Assessing Team Strategy using Spatiotemporal Data

Patrick Lucey Dean Oliver Disney Research ESPN Pittsburgh, PA, USA Bristol, CT, USA [email protected] [email protected]

Peter Carr Joe Roth Iain Matthews Disney Research Disney Research Disney Research Pittsburgh, PA, USA Pittsburgh, PA, USA Pittsburgh, PA, USA [email protected] [email protected] [email protected]

ABSTRACT Canada USA In CVPR,2009. 9(3) Shots (on Goal) 12(4) [16] W. Lu, J. Ting, K. Murphy, and J. Little. Identifying The Moneyball revolution coincided with a shift in the way 13 Fouls 11 Players in Broadcast Sports Videos using Conditional professional sporting organizations handle and utilize data in 3 Corner Kicks 8 Random Fields. In CVPR,2011. 1O↵sides 4 [17] P. Lucey, A. Bialkowski, P. Carr, E. Foote, and terms of decision making processes. Due to the demand for 38% Time of Possession 62% I. Matthews. Characterizing Multi-Agent Team better sports analytics and the improvement in sensor tech- 3 Yellow Cards 1 Behavior from Partial Team Tracings: Evidence from 0 Red Cards 0 nology, there has been a plethora of ball and player tracking the English Premier League. In AAAI,2012. 4 Saves 3 information generated within professional sports for analyt- [18] L. Madden. NFL to Follow Army’s Lead on Helmet ical purposes. However, due to the continuous nature of the (a) (b)Sensors in Attempt to Prevent Head Injury. data and the lack of associated high-level labels to describe www.forbes.com/sites/lancemadden/2012/07/16/ 8. REFERENCESFigure 1: (a) An example of standard soccernfl-to-follow-armys-lead-on-helmet-sensors-in/ statis- it - this rich set of information has had very limited use espe-[1] S. Alitics and based M. Shah. on Floor hand-labeled Fields for Tracking event in High data which-attempt-to-prevent-head-injury/ describe 16 July 2012. ” cially in the analysis of a team’s tactics and strategy. In this Densitywhat Crowdhappened. Scenes. In ECCV (b),2008. Spatiotemporal data[19] has R. Masheswaran, the po- Y. Chang, A. Henehan, and paper, we give an overview of the types of analysis currently[2] N. Allen,tential J. Templon, to describe P. McNally, the L. Birnbaum,where and and how, butS. Danesis. as it Destructing is the Rebound with Optical K. Hammond. StatsMonkey: A Data-Driven Sports performed mostly with hand-labeled event data and high- a continuous signal which is not associatedTracking with Data. a In MIT Sloan Sports Analytics Narrative Writer. In AAAI Fall Symposium Series, Conference,2012. light the problems associated with the influx of spatiotem- 2010.fixed event, using this data for analysis[20] is V. difficult. Morariu and L. Davis. Multi-Agent Event poral data. By way of example, we present an approach[3] BBC-Sports. Footballers may trial wearing microchips Recognition in Structured Scenarios. In CVPR,2011. which uses an entire season of ball tracking data from the to monitor health. [21] T. Moskowitz and L. Wertheim. Scorecasting: The www.bbc.co.uk/sport/0/football/21460038 English Premier League (2010-2011 season) to reinforce the 1. INTRODUCTION ,14Feb Hidden Influences Behind How Sports Are Played and 2013. Games Are Won. Crown Publishing Group, 2011. common held belief that teams should aim to “win home[4] M. Beetz,In N. his von 2003 Hoyningen-Huene, book Moneyball B. Kirchlechner,[14], Michael Lewis docu- mented how Oakland A’s General Manager Billy[22] NBA Beane Shot was Charts. www.nba.com/hotspots. games and draw away ones”. We do this by: i) forming a S. Gedikli, F. Siles, M. Durus, and M. Lames. [23] D. Oliver. Basketball on Paper: Rules and Tools for representation of team behavior by chunking the incoming ASPOGAMO:able to Automatedeffectively Sports use Game metrics Analysis derived from hand-craftedPerformance Analysis. Brassey’s, Incorporated, 2004. Models. International Journal of Computer Science in spatiotemporal signal into a series of quantized bins, and ii) statistics to exploit the inefficiencies in the[24] value D. Oliver. of indi- Guide to the Total Quarterback Rating. Sport,8(1),2009. espn.go.com/nfl/story/_/id/6833215/ generate an expectation model of team behavior based on[5] P. Carr,vidual Y. Sheikh, baseball and I.players. Matthews. Around Monocular the same time, Basketball on Paper [24] was published which outlined methodsexplaining-statistics-total-quarterback-rating for , a code-book of past performances. We show that home ad- Object Detection using 3D Geometric Primitives. 4 August 2011. vantage in soccer is partly due to the conservative strategy 2012.valuing player performance in basketball which[25] is Opta a far Sports. morewww.optasports.com. of the away team. We also show that our approach can flag[6] K. Goldsberry.challenging CourtVision: problem New because Visual and it is Spatial a continuous[26]team S. Pellegrini, sport. A. Ess, K. Schindler, and L. van Gool. Analytics for the NBA. In MIT Sloan Sports Analytics You’ll Never Walk Alone: Modeling Social Behavior anomalous team behavior which has many potential appli- ConferenceDue to,2012. the popularity and effectiveness of the tools that em- for Multi-Target Tracking. In CVPR,2009. [7] A. Gupta,anated P. Srinivasan,from these J. Shi, works, and L. there Davis. has been enormous interest cations. [27] M. Perse, M. Kristan, S. Kovacic, and J. Pers. A Understanding Videos, Constructing Plots: Learning a in the field of sports analytics over the last 10Trajectory-Based years with Analysis of Coordinated Team Visuallymany Grounded organizations Storyline Model (e.g. from professional Annotated teams, mediaActivity groups) in Basketball Game. Computer Vision and Categories and Subject Descriptors Videos.housing In CVPR their,2009. own analytics department. However,Image nearly Understanding all ,2008. [8] Hawk-Eye. www.hawkeyeinnovations.co.uk. H.4 [Information Systems Applications]: Miscellaneous; of the analytical works have dealt solely with[28] hand-labeled Prozone. www.prozonesports.com. [9] D. Henschen. IBM Serves New Tennis Analytics At [29] B. Siddiquie, Y. Yacoob, and L. Davis. Recognizing I.2.6 [Learning]: General event datawww.informationweek.com/software/ which describes what happened (e.g. basketball Wimbledon. Plays in American Football Videos. Technical report, business-intelligence/ - rebounds, points scored, assists, football - yardsUniversity per carry, of Maryland, 2009. ibm-serves-new-tennis-analytics-at-wimbl/ tackles, sacks, soccer - passes, shots, tackles (see[30] Figure SportsVision. 1(a))).www.sportsvision.com. Keywords 240002528, 23 June 2012. [31] STATS SportsVU. www.sportvu.com. Sports Analytics, Spatiotemporal Data, Representation [10] A. HervieuOnce andthe P. data Bouthemy. is in Understandingthis form, most sports approaches just relate videoto using parsing players in trajectories. the relevant In J. Zhang, data L. from Shao, a database,[32] Statsheet. then ap-www.statsheet.com. L. Zhang,plying and sport-based G. Jones, editors, rulesIntelligent and standard Video statistical[33] D. methods,Stracuzzi, A. Fern, K. Ali, R. Hess, J. Pinto, N. Li, T. Konik, and D. Shapiro. An Application of Transfer Eventincluding Analysis and regression Understanding and. Springeroptimization. Berlin / Heidelberg, 2010. to American Football: From Observation of Raw Permission to make digital or hard copies of all or part of this work for[11] S. IntilleAs and most A. Bobick. sporting A Framework environments for Recognizing tend to be dynamicVideo to with Control in a Simulated Environment. AI personal or classroom use is granted without fee provided that copies are Multi-Agentmultiple Action players from continuously Visual Evidence. moving In AAAI and, competingMagazine against,32(2),2011. not made or distributed for profit or commercial advantage and that copies 1999.each other, simple event statistics do not capture[34] X. Wei, the P.com- Lucey, S. Morgan, and S. Sridharan. bear this notice and the full citation on the first page. To copy otherwise, to[12] K. Kim, M. Grundmann, A. Shamir, I. Matthews, Sweet-Spot: Using Spatiotemporal Data to Discover plex aspects of the game. To gain an advantageand over Predict the Shots in Tennis. In MIT Sloan Sports republish, to post on servers or to redistribute to lists, requires prior specific J. Hodgins, and I. Essa. Motion Fields to Predict Play Evolutionrest of in Dynamicthe field, Sports sporting Scenes. organizations In CVPR,2010. have recentlyAnalytics looked Conference,2013. permission and/or a fee. [35] C. Xu, Y. Zhang, G. Zhu, Y. Rui, H. Lu, and KDD’13, August 11–14, 2013, Chicago, Illinois, USA. [13] M. Lewis.to employMoneyball: commercial The Art of Winning tracking an technologies Unfair which can lo- Gamecate. Norton, the position 2003. of the ball and players at each timeQ. Huang. instant Using Webcast Text for Semantic Event Copyright 2013 ACM 978-1-4503-2174-7/13/08 ...$15.00. Detection in Broadcast. T. Multimedia,10(7),2008. [14] R. Li and R. Chellappa. Group Motion Segmentation Using a Spatio-Temporal Driving Force Model. In [36] Zonalmarking. www.zonalmarking.net. CVPR,2010. [15] R. Li, R. Chellappa, and S. Zhou. Learning Multi-Modal Densities on Discriminative Temporal 1366Interaction Manifold for Group Activity Recognition. in professional leagues [29, 32, 9, 26]1 - to determine where alizations for the television broadcasters [9]. Partial data and how events happen. Even though there is potentially sources normally generated by human annotators such as an enormous amount of hidden team behavioral information shot-charts in basketball and ice-hockey are often used for to mine from such sources, due to the sheer volume as well analysis [23], as well as passing and shot charts in soccer as the noisy and variable length of the data, methods which [26]. Recently, ESPN developed a new quarterback rating can adequately represent team behaviors are yet to be devel- in American Football called“TotalQBR”[25] which attempts oped. The value of this data is limited as little analysis can to assign credit or blame to the quarterback depending on be conducted. What compounds the difficulty of this task a host of factors such as pass or catch quality, importance is the low-scoring and/or continuous nature of many team in the match, pass thrown under pressure or not. As these sports (e.g. soccer, hockey, basketball) which makes it very factors are quite subjective, annotators who are reliable in hard to associate segments of play with high-level behaviors labeling such variables are used. In terms of strategic anal- (e.g. tactics, strategy, style, system, formations). Without ysis, zonalmarking.net [37] attempts to describe a soccer these labels, which essentially give the game context, infer- match from a tactical and formation point of view. Whilst ring team strategy or behavior is impossible as there are no interesting, this approach is still qualitative and is based factors to condition against (see Figure 1(b)). solely on the opinion of the analyst. Due to these complexities, there has been no effective As the problem of fully automatic multi-agent tracking method of utilizing spatiotemporal data in continuous sports. from vision-based systems is still an open one, most aca- Having suitable methods which can first develop suitable demic research has centered on the tracking problem [1, 27, representations from imperfect (e.g. noisy or impartial) data, 17, 5]. The lack of fully automated tracking approaches has and then learn team behaviors in an unsupervised or semi- limited team behavioral research to works on limited size supervised manner, as well as recognize and predict future datasets. The first work which looked at using spatiotem- behaviors would greatly enhance decision making in all ar- poral data for team behavior analysis was conducted over 10 eas of the sporting landscape (e.g. coaching, broadcasting, years ago by Intille and Bobick [12]. In this seminal work, fantasy-games, video games, betting etc.). We call this the the authors used a probabilistic model to recognize a single emerging field of Sports Spatiotemporal Analytics - and we football play from hand annotated player trajectories. Since show an example of this new area of analytics by compar- then, multiple approaches have centered on recognizing foot- ing the strategies of home and away teams in the English ball plays [16, 30, 15, 34], but only on a very small number of Premier League by using ball tracking data. plays (i.e. 50-100). For soccer, Kim et al. [13] used the global motion of all players in a soccer match to predict where the play will evolve in the short-term. Beetz et al. [4] proposed 2. RELATED WORK a system which aims to track player and ball positions via The use of automatic sports analysis systems have re- a vision system for the use of automatic analysis of soccer cently graduated from the virtual to the real-world. This matches. In basketball, Perse et al. [28] used trajectories of is due in part to the popularity of live-sport, the amount player movement to recognize three type of team offensive of live-sport being broadcasted, the proliferation of mobile patterns. Morariu and Davis [21] integrated interval-based devices, the rise of second-screen viewing, the amount of temporal reasoning with probabilistic logical inference to data/statistics being generated for sports, and demand for recognize events in one-on-one basketball. Hervieu et al. [11] in-depth reporting and analysis of sport. Systems which use also used player trajectories to recognize low-level team ac- match statistics to automatically generate narratives have tivities using a hierarchical parallel semi-Markov model. In already been deployed [33, 2]. Although impressive, these addition to these works, plenty of work has centered on an- solutions just give a low-level description of match statistics alyzing broadcast footage of sports for action, activity and and notable individual performances without giving any tac- highlight detection [36, 8]2. Even though notable, the lack tical analysis about factors which contributed to the result. of tracking data to adequately train models has limited the In tennis, IBM has created Slamtracker [10] which can pro- usefulness of the above research. vide player analysis by finding patterns that characterize It is clear from the overview given above, that there ex- the best chance a player has to beat another player from an ists a major disparity in resources between industry and enormous amount of event labeled data - although no spa- academia to deal with this problem domain. Sporting orga- tiotemporal data (i.e. player or ball tracking information) nizations that receive large volumes of spatiotemporal data has been used in their analysis yet. from third-party vendors but often the people within these Spatiotemporal data has been used extensively in the vi- organizations lack the computational skills or resources to sualization of sports action. Examples include vision-based make use of it. Contrastingly, due to the proprietary nature systems which track baseball pitches for Major League Base- of commercial tracking systems, and the cost and method of ball [31], and ball and players in basketball and soccer [32, generating the tracking data, research groups who have the 29]. Hawk-Eye deploy vision-based systems which track the necessary skills can not access these large data repositories. ball in tennis and cricket, and is often used to aid in the Recently however, due to the potential payoff, some industry officiating of these matches in addition to providing visu- groups are investing in analytical people with these skill sets, or have teamed up with research groups to help facilitate a 1As nearly all professional leagues currently forbid the use of wearable sensors on players, unobtrusive data capture meth- solution. The release of STATS Sports VU data [32] to some ods such as vision-based systems or armies of human annota- research groups has enabled interesting analysis of shots and tors are used to provide player and ball tracking information. rebounding in the NBA[7, 20]. In tennis, Wei et al. [35] However, this restriction may change soon as monitoring the health and well-being of players has attracted significant in- 2These works only capture a portion of the field, making terest lately, especially for concussions in American Foot- group analysis very difficult as all active players are rarely ball [19], as well as heart issues in soccer [3]. present in the all frames.

1367 Home Away N Team W D L P GF SF GA SA W D L P GF SF GA SA 1 Man Utd 18 1 0 55 49 347 12 191 5 10 4 25 29 272 25 271 2 Chelsea 14 3 2 45 39 379 13 233 7 5 7 26 30 367 20 208 3 Man City 13 4 2 43 34 306 12 216 8 4 7 28 26 240 21 298 4 Arsenal 11 4 4 37 33 350 15 154 8 7 4 31 39 305 28 251 5 Tottenham 9 9 1 36 30 383 19 228 7 5 7 26 25 274 27 339 6 Liverpool 12 4 3 40 37 319 14 220 5 3 11 18 22 266 30 270 7 Everton 9 7 3 34 31 321 23 227 4 8 7 20 20 259 22 279 8 Fulham 8 7 4 31 30 307 23 262 3 9 7 18 19 245 20 271 9 Aston Villa 8 7 4 31 26 273 19 263 4 5 10 17 22 233 40 340 10 Sunderland 7 5 7 26 25 287 27 243 5 6 8 21 20 246 29 311 11 West Brom 8 6 5 30 30 329 30 237 4 5 10 17 26 273 41 297 12 Newcastle 6 8 5 26 41 300 27 250 5 5 9 20 15 209 30 256 13 Stoke City 10 4 5 34 31 298 18 256 3 3 13 12 15 186 30 294 14 Bolton 10 5 4 35 34 311 24 256 2 5 12 11 18 261 32 346 15 Blackburn 7 7 5 28 22 254 16 259 4 3 12 15 24 200 43 360 16 Wigan 5 8 6 23 22 290 34 227 4 7 8 19 18 221 27 284 17 Wolves 8 4 7 28 30 256 30 266 3 3 13 12 16 205 36 306 18 Birmingham 6 8 5 26 19 231 22 324 2 7 10 13 18 174 36 362 19 Blackpool 5 5 9 20 30 296 37 297 5 4 10 19 25 240 41 441 20 West Ham 5 5 9 20 24 325 31 317 2 7 10 13 19 250 39 378 SUM 179 111 90 648 617 6162 446 4926 90 111 179 381 446 4926 617 6162 AVG(per game) 0.47 0.29 0.24 1.71 1.62 16.2 1.17 13.0 0.24 0.29 0.47 1.00 1.17 13.0 1.62 16.2

Table 1: Table showing the statistics for the home and away performances for each team in the 2010 EPL season: (left columns) home matches (right columns) away columns (Key: W = wins, D = draws, L = losses, P = points (3 for a win, 1 for a draw, 0 for a loss), GF = goals for, SF = shots for, GA = goals against, SA = shots against). used ball and player tracking information to predict shots cant difference between teams at home and away (10.01% using data from the 2012 Australian Open. For soccer, re- vs 9.05% for shooting and 73.46% vs 72.99% for passing - searchers have characterized team behaviors in the English see the bottom row in Table 1). Premier League using ball-motion information across an en- However, there is a large difference between the amount of tire season using OPTA data [18]. In this paper, we extend shots (16.2 vs 13.0) and goals scored (1.62 vs 1.17) at home this method to explain that the home advantage in soccer and away. An illuminating example is the league champions is due to the conservative strategy that away teams use (or for that season, Manchester United (see top row in Table 1). more aggressive approach of the home team) which rein- At home, they were unbeaten (winning 18 and drawing 1), forces the commonly held belief that teams aim to win their but away from home they only won 5 games, drew 10 and home games and draw their away ones. lost 4. The telling statistic is that at home they scored 49 goals from 347 shots, compared to only 29 goals from 272 3. CASE STUDY: shots away from home. Comparatively, the opposition at home games only scored a paltry 12 goals from 191 shots HOME ADVANTAGE IN SOCCER while at away games they scored 25 goals from 271 shots. In soccer, there is the commonly held belief that team should 3.1 “Win at Home and Draw Away” aim to win their home games and draw their away ones. If In a recent book by Moskowitz and Wertheim [22], they you skim down Table 1, you will find that: i) all teams won highlight that the home advantage exists in all professional more home games (except for Blackpool who won the same sports (i.e. teams win more at home than away). The au- amount), ii) all teams score more goals at home (except Ar- thors hypothesized that referees play a significant role by senal and Blackburn), iii) all teams had more shots at home giving home teams favorable calls at critical moments. They compared to away, and iv) all teams gained more points at then quantitatively showed this in baseball through the use home. These event statistics tell us what has occurred, in of pitch tracking data. For soccer, hand-labeled event statis- the rest of the paper we use spatiotemporal data to help tics such as the amount of injury time, number of yellow explain where and why this occurred. Before we detail the cards and number of penalties awarded to reinforce their method, we first describe the data. hypothesis. As soccer is a very tactical game, we hypothe- size that the strategy of the home and away teams also plays a role in explaining the home advantage. 3.2 Ball Tracking and Event Data A great case study of home advantage is the 2010-2011 Due to the difficulty associated with accurately tracking English Premier League soccer season. In that season, the players and the ball, data containing this information is still home team earned an average 1.71 points out of a total 3 scarce. Most of the data collected is via an army of human points per match. This is in stark contrast with the away annotators who label all actions that occur around the ball team, which earned only 1.00 points per game: a rather - which they call ball actions. The F24 soccer data feed col- large discrepancy, especially considering that teams play ev- lected for the English Premier League (EPL) by Opta [26] ery other team both at home and away so that any talent is a good example of this. The F24 data is a time coded disparities apply to both home and away averages. In terms feed that lists all player action events within the game with of shooting and passing proficiency, there was no signifi- a player, team, event type, minute and second for each ac-

1368 FigureFigure 2: 2: An An example example of of IBM IBM Slamtracker Slamtracker (IBM (IBM Slam- Slam- FigureFigure 2: 2: An An example example of of IBM IBM Slamtracker Slamtracker (IBM (IBM Slam- Slam- TrackerTracker 2012), 2012), highlighting highlighting the the plans plans Rafael Rafael Nadal Nadal needs needs FigureFigureTrackerTracker 2: 2: 2012),An An 2012), example example highlighting highlighting of of IBM IBM the the Slamtracker plansSlamtracker plans Rafael Rafael (IBM (IBMNadal Nadal Slam- Slam- needs needs toto execute execute to to win win his his second second round round match match at at the the 2012 2012 Aus- Aus- TrackerTrackertoto execute execute 2012), 2012), to to win highlighting highlightingwin his his second second the the round round plans plans match match Rafael Rafael at at Nadal the the Nadal 2012 2012 needs needs Aus- Aus- traliantralian Open. Open. These These set set of of plans plans can can be be seen seen as as the the tactics tactics tototralian execute executetralian Open. to Open. to win win These Thesehis his second set second set of of round plans plansround can match can match be be at seen seen at the the as 2012 as 2012 the the Aus- tactics tactics Aus- Figure 2: An example of IBM Slamtracker (IBM Slam- thatthat he he should should employ. employ. Figure 2: An example of IBM Slamtracker (IBM Slam- Figuretralianthat he2: Open. should An Theseexample employ. set of of IBM plans Slamtracker can be seen (IBM as the Slam- tactics FigureTrackerFigure 2: 2012), 2: An An example highlightingexample of of IBM IBM the Slamtrackerplans Slamtracker Rafael (IBM Nadal(IBM Slam- Slam-needs tralianthatFigure Open. he should 2: These An employ. exampleset of plans of IBM can be Slamtracker seen as the (IBM tactics Slam- Tracker 2012), highlighting the plans Rafael Nadal needs Trackerthat he should 2012), employ. highlighting the plans Rafael Nadal needs FigureFigureTrackertoTracker execute 2: 2: 2012),An An 2012), to example winexample highlighting highlighting hisFigure secondof of IBM IBM 3: the round the Figure Slamtracker plansSlamtracker plans match 2. Rafael Rafael at (IBM the (IBMNadal Nadal 2012 Slam- Slam- needs needs Aus- thatFigure heTracker should 2: Anemploy.2012), example highlighting of IBM the Slamtracker plans Rafael (IBM Nadal Slam- needs to execute to winFigure his second 3: Figure round 2.match at the 2012 Aus- to execute to win his second round match at the 2012 Aus- TrackerTrackertotralianto execute execute 2012), 2012),Open. to to win highlighting These highlightingwin hisFigure his set second second of the3: 3: plans the round Figure Figure round plans plans can match match Rafael2.2. be Rafael seen at at Nadal the the Nadalas 2012 2012the needs tacticsneeds Aus- Aus- theirTrackerto overall execute 2012), goal to win highlighting (i.e., his did second they the roundwin plans the match match Rafael at or the Nadal not?), 2012 needs weAus- tralian Open. TheseFigure set of 3: plans Figure can 2. be seen as the tactics traliantheir Open. overall These goal set (i.e., of did plans they can win be the seen match as the or not?), tactics we tototralian executethat executetralian he Open. should to Open. to win win These employ. ThesehisFigure his second set second set of 3: of round plans Figure plansround can match can match 2. be be at seen seen at the the as 2012 as 2012 the the Aus- tactics tactics Aus- Figuretheircantotheir 2:tralian execute find overallAn overall examplethe Open. to goal plans goal win These (i.e., of (i.e.,his or IBMa did second set did sequence they of Slamtracker they plans round win win of thecan matchthe plans match be(IBM match seenwhich at the orSlam- or as not?), not?),2012 correlate the tactics weAus- we Figuretralianthat2: Open. he An should Theseexample employ. set of of IBM plans Slamtracker can be seen (IBM as the Slam- tactics that hecan should find the employ. plans or a sequence of plans which correlate thetralianthat periodthatFigure he Open. he should should of 2: These the Anemploy. employ. observations). exampleset of plans ofIBM can Banjeer be Slamtracker seen as the (IBM tactics Slam- Trackertheirtheircantralian overall2012), findoverall Open. the highlightinggoal goal plans These(i.e., (i.e., or didaset did the sequence they of they plans plans win win Rafael theof can the plans match be match Nadal seen which or or not?), asneeds not?), correlate the we tactics we theTrackerthat period he should 2012), of the employ. highlighting observations). the Banjeerplans Rafael Nadal needs Figure 3: Figure 2. withcanthat a find player he theshould winning plans employ. or and a sequence losing. These of plans sequence which correlate of plans thethat periodFigure heTracker should of 2: the Anemploy.2012), observations). example highlighting of IBM Banjeer the Slamtracker plans Rafael (IBM Nadal Slam- needs Figure 3: Figure 2. to executecanwiththatwith find a to he player awin the should player plans his winning employ.second winning or a and roundsequence and losing. losing. match of These These atplans the sequencesequence 2012which Aus- correlate of of plans plans thetoMulti-agent execute period of to the win plan observations). his secondrecognition round Banjeer explores match at thean explanation2012 Aus- of Figure 3: 3: Figure Figure 2. 2. canorwith findtactics thea playerwill plans vary winning or according a sequence and losing. to of the plans These opposition’s which sequence correlate learnt of plans suc- theMulti-agenttheirMulti-agentTracker periodto overall execute of 2012), the goal plan to observations). win highlightingFigure (i.e., recognition his did second 3: they Figure theexploresround winBanjeer explores plans 2. the match match Rafael an an at explanation explanation or the Nadal not?), 2012 needs weAus- of of tralianwithor Open.tacticsor atactics player Thesewillwill winning vary set vary of according accordingplans and losing. can to to bethe Thesethe seen opposition’s opposition’s as sequence the tactics learnt learnt of plans suc- suc- thethetralian periodMulti-agent observedtheir Open. of overall the These teamobservations). plan goal recognitionset trace, (i.e.,Figure of did plans i.e., 3:they Banjeer Figurecanexplores activity win be the seen2. sequencesmatch an as explanation the or not?), tactics of we a of set FigureFigure 3: Figure 3: Figure 2. 2. withcessfulor atactics player andwill winning unsuccessful vary andaccording losing. plans. to These the opposition’s sequence of learnt plans suc- FigurethethetheirMulti-agentcanto observedtheir 2:observedtralian execute find overallAn overall examplethe Open. to teamgoal team plans plan goal win These (i.e., oftrace, (i.e.,his orrecognitionFigure IBMa did second set did sequence i.e., i.e., they of Slamtracker they 3: plans round activity activity win Figure win explores of thecan matchthe plans match sequences be2.(IBMsequences match seenanwhich at the orexplanationSlam- or as not?), not?),2012 correlate the of of tactics a weAus- wea set set of thattheiror hecessfultactics shouldcessful overall andwill employ. and goal unsuccessful vary unsuccessful (i.e., according did plans. they plans. to win the the opposition’s match or not?), learnt we suc- Multi-agenttheofthattheircan agents, observed hecan findoverall should find the by plan thegoal team employ.plans identifying plans recognition(i.e., trace,or ora did sequence a i.e., thetheysequence explores dynamic activity win of theof plans plans match sequencesan team which explanation which or structures not?), correlate correlate of a we of set and the period of the observations). Banjeer or tacticscessfulFortheir teamwill overall and vary sports, unsuccessful goal according such (i.e., as plans.did to American thethey opposition’s win Football the match learnt and or Basket- not?), suc- we Trackerofthetheirof agents,with agents,tralian observedcan overall2012),that a find player he Open. by by highlightinggoal theshould identifyingidentifying team winning plans These(i.e., employ. trace, ordid set andthe a they thesequence of i.e.,plans losing. plans windynamic dynamic activity Rafael the canThese of match plans be team team Nadal sequence sequences seen which or structures structures not?), asneeds the correlate of weplans oftactics and aand set thethe period period of of the the observations). observations). Banjeer Banjeer cancessfulFor findFor andteam the team unsuccessful plans sports, sports, or such a such sequence plans. as as American American of plans Football Football which and and correlate Basket- Basket- tothe executeofteamcan observedwith agents,thatwith find behaviors a to he player awin the should by player team plans hisidentifying winning of employ.second winning trace, agentsor a and roundsequence i.e., and the based losing. losing.activity dynamic match on of These These atplansa sequences library the team sequencesequence 2012which structures of Aus- teamcorrelate of of plans a plans plans. setand theMulti-agent period of the plan observations). recognition Banjeer explores an explanation of cessfulball,theircanFor and a overall similar find team unsuccessful the sports, goal approach plans (i.e., such plans.or didcan a as sequence they American be taken win of the asFootball plans matchthematch which and or not?), Basket- iscorrelate seg- we theteamofcanteam periodor agents,with findtactics behaviors behaviors theaof player by thewill plansFigure identifying observations). ofvary of winning or agents according a 3: sequence Figure and based based the losing. Banjeerto dynamic 2.of on onthe plans Thesea aopposition’s library library whichteam sequence of of structures correlate team learnt team of plansplans. suc- plans. and theMulti-agentMulti-agent period of the plan observations).Figure recognition 3: Figure explores Banjeerexplores 2. an an explanation explanation of of withForball, a player team a similar sports, winning approach such and as losing. canAmerican be These taken Football sequence as the matchand of Basket- plans is seg- tralianofteamIt agents,withtheor has Open.tacticsor behaviorsperiod a importanttactics player by Thesewill identifying ofwill winning the vary of set applicationsvary agentsobservations). of according accordingplans and the based losing. dynamiccan to in to beonthe Banjeer analyzingThese the seen aopposition’s libraryopposition’s team as sequence the structures dataof tactics learntteam learnt offrom plans plans.suc- andsuc- auto- thethe periodMulti-agent observed of the teamobservations).plan recognition trace,Figure i.e., 3: Banjeer Figureexplores activity 2. sequences an explanation of a of set ball,Formentedcanball,with team a find asimilar into asimilar sports, theplayer distinct plansapproach approach suchwinning plan or as a can segments.American sequenceand can be losing.be taken taken However, Footballof These as plans as the the sequence and whichmatch due match Basket- to correlateisthe is of seg- seg- dif- plans withItMulti-agentcessful hasor atacticsplayer important andwill winningplan unsuccessfulvary applications recognition andaccording losing. plans. in toexplores These analyzing the opposition’s sequence an data explanation of fromlearnt plans auto-suc- of thetheMulti-agent observed observed team team plan trace, recognition i.e., i.e., activity activity explores sequences sequences an explanation of of a a set set of or tacticsmentedwill into vary distinct according plan segments. to the opposition’s However, due learnt to the suc- dif- thatteamItteamItmatedtheiror hethe has has behaviorstactics shouldcessful period behaviorsoverall important important monitoring,will employ. and ofgoal of vary theunsuccessful of agents applications (i.e., observations). agentsaccording surveillance did based they based plans. to inin winon the analyzingand analyzing on a Banjeer the opposition’s library a intelligence match library data of data or team of not?), learnt from from teamanalysis plans. we suc- auto- auto- plans. in Multi-agenttheof agents, observed by plan team identifying recognition trace,Figure i.e., the 3: explores dynamic activity Figure sequencesan 2. team explanation structures of a of set and theirball,ball,mented overallwith a a similar similar a goal into playerdistinct (i.e., approach approach winning did plan they can canand segments. win be belosing. taken the taken match However, Theseas as the theor sequence match not?), duematch to is we the is seg-of seg- dif- plans ormatedcessfultacticscessfulMulti-agentFortheir monitoring,team andwill overall and unsuccessful vary sports, unsuccessful plan goal according surveillance such recognition(i.e., plans. as plans.did to American thethey and opposition’s explores win intelligence Football the match an learnt explanationandanalysis or Basket- not?), suc- wein of theof agents, observed byidentifying team trace, the i.e., dynamic activity team sequences structures of and a set ficultymentedor tactics with into trackingwill distinct vary plan all according the segments. players to the However, and opposition’s ball duein a to confined learnt the dif- suc- ItthematedIt hascancessful observedhas importantMulti-agent find monitoring,importantFor and the team unsuccessful team plans applications sports, plan applications trace,or surveillance a recognitionsuch sequence plans. i.e., as inAmerican activity analyzingin and of analyzing explores plans intelligence sequences Football which data an data from and correlate explanation analysis offrom Basket- auto- a set auto- in oftheofofteam observed agents, agents, behaviors by team identifying of trace, agents i.e., the the based activitydynamic dynamic on a sequences library team team structures structures of team of a plans. setand and cancessfulmented findficultyorficulty thetactics and intowith plans withunsuccessfulwill distinct tracking or tracking vary a sequence plan according all plans. all segments. the the playersof players to plans the However, and opposition’swhich and ball ball correlatedue in in a to a confinedlearnt confinedthe dif- suc- matedgeneral.cessfultheball,theirForcan observedFor monitoring, and ateam overallIt similar find team is unsuccessful also sports, the sports, goal team approach a plans surveillancedifficult (i.e.,such suchtrace, plans.or didascan a as task sequenceAmerican i.e., they American be and since taken activity win intelligence of theFootballan asFootball plans observed matchthe sequencesmatch which and and or analysis Basket-not?), team Basket- iscorrelate of seg- tracea we in set theofteam period agents, behaviors of by theFigure identifying observations). of agents 3: Figure based the Banjeer dynamic 2. on a library team of structures team plans. and mentedspace,ficultycessful into work with distinct and in tracking unsuccessful this plan space segments.all hasthe plans. playersbeen However, limited and ball dueto small in to a the confined datasets dif- the periodofgeneral.matedgeneral.with agents,theFor ofball, a observed monitoring, theplayer team It It by a is is observations). similar alsoidentifying alsosports, winning team a approachdifficult suchsurveillance and trace, theas Banjeer losing. task canAmerican dynamic i.e., besince since Theseand activitytaken an anFootballintelligence team sequence observedas observed the sequences structures matchand of team team Basket- analysis plans is trace seg-of and trace a in setofteamteamIt agents,the has behaviors behaviorsperiod important by identifying of the of applications agentsobservations). agents the based based dynamic in on on Banjeer analyzing a a library team library structures dataof of team team from plans. plans.and auto- withficulty aspace,Forcessful playerspace, team with work winning and work sports, tracking in unsuccessful thisin and this such space all losing.space as the has American plans. has players These been been limited sequence and Footballlimited ball to to small ofinand small plansa Basket- confined datasets datasets matedgeneral.isofball, oftenFormentedcanball, monitoring, agents,with team a find composed It asimilar into is asimilar sports, theplayer also by distinct plansidentifyingapproach asurveillance approach suchdifficultwinning of manyplan or as a can segments.American sequenceand taskcan possible the be and losing.be since dynamic taken taken intelligenceHowever,Football of team an These as plans as observed theteam plansthe sequence and whichmatch due match structures analysis in Basket- toteam the correlateisthe is ofseg- library, seg-trace dif- plans in and teamItMulti-agent has behaviors important plan of applications recognition agents based in explores analyzing on a library an data explanation of from team auto- plans. of ficultyandspace, onlywithFor work focustrackingteam in sports, on this all single space the such playersplan has as beenAmerican recognition and limited ball Football toin (Intille small a confined and and datasets Basket- Bo- Multi-agenttheirteamisgeneral.isorball, often overall oftentactics behaviorsmented a composed composedsimilar It plangoalwill is into also (i.e., varyrecognition of approach distinct agents a of didaccording difficult many they plan based canexplorespossible winpossible segments. task to be the on taken sinceopposition’s matcha team team an library However, as an explanation plans theor plansobserved not?), of match due inlearnt teamin the wetheto isteamthe library,suc-of plans. library, seg- dif- traceteamItItmatedthe has has behaviorsMulti-agent period important important monitoring, of of the agentsplan applications observations). surveillance recognition based in in on analyzingand analyzing aexplores Banjeer library intelligence dataan of data teamexplanation from from analysis plans. auto- auto- inof or tacticsball,space,and awillsimilar work only vary in focus approach according this on space single can to has the be plan been opposition’s taken recognition limited as the to learnt match (Intillesmall suc- is datasets and seg- Bo- general.isandball,ofteam oftenmentedficultywithmented team-matesagents, aor It similar behaviorscomposed istactics a withinto playeralso into by distinct approachtracking awill distinct difficultidentifying winning may of of vary manyagents plan plandynamically all canaccording taskand segments. the possiblesegments. be based the losing. playerssince taken dynamicto on However,teamthean changeTheseHowever, as and a observed opposition’s the library plans ball sequenceteam match in due duein in ofthe team ato structuresthe tois team confined learnt the observing theseg- of library, trace dif- plans dif- plans. suc- andtheItmated observedhas important monitoring, team applications trace, surveillance i.e., activity in and analyzing intelligence sequences data analysis offrom a set auto- in space,andbickandForball,only work 2001; only team a focus in similar focus?; this sports,?). onspace Conversely,on approach single single such has plan as beenplan canAmerican for recognition berecognition limited team taken sports Football to as small (Intille (Intille the such match anddatasets as and and soccer, Basket- is Bo- Bo- seg-thecanIt observedisand hascessful find often team-matesimportantficulty the andcomposed team plans withunsuccessful trace, or applications tracking may a of sequence i.e., many dynamically plans. all activity the possible in of players analyzingplans sequences change team which and ballplans datain correlate the of in from in a a observing confinedthe set auto- library,Itmated hasmatedgeneral.the importantMulti-agent observed monitoring, It is also applications team plan a surveillancedifficult trace, recognition task in i.e., analyzing and andsince activity explores intelligence intelligence an observed sequencesdata an from explanation analysis analysis team ofauto- tracea in setin of cessfulmented andbick intounsuccessful 2001; distinct?; ?). plan Conversely,plans. segments. for However, team sports due such to the as soccer, dif- is oftenandandprocess.mentedmentedteamItficultyspace,orteam-matesficulty hasteam-mates composedcessfultactics intobehaviors importantinto withwork with distinct andwilldistinct tracking in tracking of unsuccessfulmay thisvary of many planapplications plan space agents dynamicallyaccording dynamicallyall segments.all possiblesegments. the hasthe plans.based players playersbeen to in However, team the analyzingHowever,change onlimitedchange and opposition’sand a plans libraryball balldueto in in due insmall data the in tothe thea toofa the observingconfinedlearntconfined fromobserving datasetsthe library, team dif- dif- suc- auto- plans.the periodofmatedgeneral. agents, of monitoring, the It by is observations). identifyingalso a difficult surveillance the Banjeer task dynamic since and an intelligence team observed structures team analysis trace and in andandbickanball,bick only onlyabundant 2001; a2001; focus similarfocus?; of? on;). on spatio-temporal). approach Conversely, single Conversely,single plan plan can forrecognition for recognitionbe tracing team team taken sports sports data as (Intille (Intille the suchis such available match and as as and soccer, soccer, Bo- is Bo- but seg-of agents,withmatedprocess.ficulty aFor player bymonitoring,space, team withidentifying winning work sports, tracking in surveillanceand this such the all losing.space as dynamic the American has players These and been team intelligence sequence and Footballlimited structures ball to ofinand small plansa analysis Basket- confined and datasets in matedgeneral.general.istheof often monitoring, agents, observed composedIt is also by identifying team asurveillance difficult of many trace, task task possible the i.e., and since sincedynamic activity intelligence team an an observed observed team plans sequences structuresanalysis in team teamthe library, trace of trace in and a set For teamanmented abundant sports, into? such? of distinct spatio-temporal as American plan segments. Football tracing However, data and is Basket- availabledue to the but dif- andprocess.andprocess.ficultyItmated team-matesspace,Thereandcessfulspace, hasteam-mates onlywithFor workimportantmonitoring, have work and focustrackingteam in mayunsuccessful been in this sports,may on this applications dynamically allspace singlemany surveillance space dynamically the such has playersplans.plan techniqueshas as been beenAmerican recognition in change and and limited analyzinglimited change intelligence ball designed Football in to toin (Intille the smallin small a data confinedthe observing andto datasets andanalysis datasets observing fromautomat- Basket- Bo- auto- inMulti-agentteamgeneral.is often behaviors composed It plan is also recognitionof agents a of difficult many based explorespossible task on since a team an library an explanation plans observed of in team the team library,of plans. trace bickficultybickanmented 2001; abundant 2001; with into;? tracking; ). of? distinct). Conversely, spatio-temporal Conversely, all plan the segments. for players for team tracing team and sports However, sports data ball suchisin such available adue as confined as soccer, to soccer, the but dif-teamor behaviorstacticsball,space,Thereand awillsimilar work only have of vary agents in focus approachbeen according this on space basedmany single can to hastechniques on the be plan been a opposition’s taken library recognition limited as designed ofthe to learntteam match (Intillesmall to plans.suc- is datasets automat- and seg- Bo- general.isisandof oftenteam often team-matesagents, It composedbehaviors composed is also by a difficultidentifying may of of manyagents many dynamically task possible possible based the since dynamic on teaman teamchange a observed library plans plans team in in ofthe in team structuresthe theteam observing library, library,trace plans. and asan itficulty abundant is low-scoring? with? of tracking spatio-temporal and allcontinuous the players tracing it is and data extremely ball is available in a difficult confined but process.general.process.space,matedThereandbickandForball,only workIt 2001; only ismonitoring, have team aalso focus in similar focus?; thisbeen sports,? a). ondifficultspace Conversely,on approach singlemany surveillancesingle such has task plantechniques as beenplan canAmerican for since recognition berecognition limited team and taken an sports designed observed intelligence Football to as small (Intille (Intille the such match anddatasets teamto as and and automat-soccer, analysisBasket- is traceBo- Bo- seg-the inIt observedisand has often team-matesimportant composed team trace, applications may of i.e., many dynamically activity possible in analyzing sequences change team plans datain the of from in a observing the set auto- library, ball,space,an aasficulty similarabundant itas is work it low-scoring is with approach low-scoring in of this tracking spatio-temporal space can and and allbe has continuous thecontinuous taken been players tracing aslimited itthe it is and data ismatch extremely to extremely ball small is available is in seg- datasets a difficult confineddifficult but cessfulicallymentedandgeneral.There andbick only recognize intounsuccessful have 2001;It focus distinctis alsobeen?; on? team). a plansinglemany Conversely,plans.difficult plans segments. plan techniques task given recognition for since However, team an designed an sports observed observed (Intille due such to to the andas teamautomat- team soccer, dif- Bo- trace trace is oftenandandprocess.teamItteam-mates hasteam-mates composed behaviors important of may of many applications agents dynamically dynamically possible based in team analyzingchange onchange a plans library in in in data the the the of observing fromobserving library, team auto- plans. anbecause abundantasspace, it is the low-scoring ofwork game spatio-temporal in thisis not and space segmented continuous has tracing been into datait limited is “discretized” extremely is available to small difficult butdatasets plays It hasisically oftenandically importantgeneral.bickThereanball,bick only abundant composedmented recognize 2001;recognize a2001; focushave similarIt applications is? into;? ofalso? beenon;).? ofdistinctteamspatio-temporal team). approach Conversely, single manya Conversely, manydifficult plansplan inplan possible can analyzingtechniques segments. given forrecognitiongiventask for be tracing team team takensinceteam an an sports However,data sportsobserved observeddatadesigned plans anas (Intille theobserved from suchis such in available match due and the asteam auto- as teamto tosoccer, library,soccer, Bo- automat-team isthe trace but seg-trace dif- traceof agents,matedandprocess. team-mates bymonitoring, identifying may surveillance the dynamically dynamic and team intelligence change structures in the analysis and observing in mentedandas it into onlybecause is low-scoring distinct focus the on plangame single andsegments. is not continuous plan segmented recognitionHowever, it into is due extremely (Intille“discretized” to the and dif- difficult Bo- plays ForThereicallyasficultybickis input. team oftenan 2001; recognize have with abundant sports, composedAvrahami-Zilberbrand beentracking; ). such of team Conversely, many spatio-temporal of as all many plansAmerican the techniques players possible forgiven team Football andtracing and an designed team sports Kaminkaa observed ball data and plans insuch is Basket- ato availableconfined in asteam automat-presented the soccer, library, trace but a andprocess.process.Itmated team-matesThere has importantmonitoring, have may been applications dynamically manysurveillance techniques in change and analyzing intelligence designed in the data observing to analysis fromautomat- auto- in asbecause(i.e. itspace,because isand plans).low-scoring workonlythe the game Consequently, focus gamein this and is on is not space notcontinuous single segmented segmented has such plan been analysis it recognition into is limitedinto extremely “discretized” “discretized” is to still (Intille small difficult conducted datasets and plays plays Bo-matedandicallybickas monitoring,an team-matesas input.mentedan 2001; abundant itficulty recognize abundant is Avrahami-Zilberbrandlow-scoring? into;? with? surveillance). of? distinctmay of Conversely, spatio-temporal tracking team spatio-temporal dynamically and plan plans allcontinuous and segments. for the given teamintelligence tracingplayers and tracing change sports it However,Kaminkaaan is data and data observed extremely such in analysisisball is the available dueavailable as inpresentedobserving soccer, ato difficultteam confined in the but but dif- trace ateamgeneral. behaviorsprocess.There It is have of also agents been a difficult basedmany tasktechniques on asince library an designed observed of team to plans.team automat- trace ficultybickbecause with(i.e. 2001; tracking theplans).?; game?). Consequently, Conversely, all is the not players segmented for such team and analysis into ball sports “discretized” in sucha is confined still as conducted soccer, plays ball,icallyasasDynamicspace,anisand a input. input. oftensimilarrecognizeabundantas team-mates work it Avrahami-Zilberbrand iscomposed Hierarchical approach low-scoring in of thisteam spatio-temporal may spacecan of plans andmany Groupdynamicallybe hascontinuous giventaken been possible Model tracing and and aslimited an the Kaminkaachange observedKaminkaa it team(DHGM), data ismatch to extremely small is plans in available is teamthe presented seg- datasetspresented which in difficultobserving the trace but library, indi- a a process.icallymatedgeneral.ThereThere recognize monitoring,have It is alsobeen team a many difficult surveillance plans techniques techniques task given since and an designed designedan intelligence observed observed to to teamautomat- team automat- analysis trace trace in because(i.e.byand(i.e. humansbick plans). only the plans). 2001; game focus whichConsequently, Consequently,?; is? on). not means Conversely, single segmented that such plan such it for isanalysis recognition into analysis subjective, team “discretized” sports is is still still(Intille unrepeatable, such conducted conducted plays as and soccer, Bo-general.process.asanDynamicas Itbecauseinput. abundantficultyas is itspace,also itis is low-scoring Avrahami-Zilberbrand Hierarchical with the low-scoringa ofwork difficult game spatio-temporal tracking in thisis and task not and Group spaceall continuous segmentedsince continuous the has tracingModel players an been and observed itinto datait (DHGM), limited is andKaminkaa is extremely“discretized” extremely is ball available team to in small which tracea presented difficult difficult confined butdatasets plays indi-It hasa is oftenically importantThere composed recognize have applications been of team many many plans in possible analyzingtechniques given team an data observeddesigned plans from in the team auto- to library, automat- trace space,an abundantworkby humans in this of spatio-temporalspace which has means been that limited tracing it is subjective, to data small is available datasets unrepeatable, but mentedasDynamicDynamiccated input.andasandprocess. it into onlybecause is team-matesthe Avrahami-Zilberbrand low-scoring distinct Hierarchicalfocus connection the on plangame single may andsegments. is betweenGroup Group not dynamicallycontinuous plan segmented ModelrecognitionHowever, Model and agents, Kaminkaa it (DHGM),into is(DHGM), change due extremelyto (Intille“discretized” trackto the in presented which and whichdif- the difficult Bo- observingdynam- plays indi- indi- a icallyThereicallyasgeneral.is input. often recognize recognize have composedAvrahami-Zilberbrand It beenis also team many a of difficult many plans plans techniques possible given giventask and since an an designed team Kaminkaa observed observed an plans observed to in team automat-presented team the team library, trace trace trace a (i.e.(i.e.byandbickby plans). humans plans).an oftenhumans 2001; abundant Consequently, unreliable Consequently,which? which; ? of). meansConversely, spatio-temporal means as there such that suchthat itareanalysis for it is analysis is teamnosubjective, tracing subjective, quantitative is sports is stilldata still unrepeatable, such unrepeatable,conductedis conducted available means as soccer, of butis oftenascatedThere composedbecause(i.e. itspace,because isand the plans).low-scoringhave workonlythe connection the of been game Consequently,focus many gamein thismany and is on is possible not spacebetween notcontinuous single segmentedtechniques segmented has such planteam agents, been analysis it recognition plans into is limitedinto designed extremely to “discretized” “discretized” in istrack the to still (Intille small tolibrary, the difficult conducted automat- dynam- datasets and plays plays Bo-matedandicallyas monitoring, team-mates input. recognize Avrahami-Zilberbrand surveillance may team dynamically plans and given intelligence and change Kaminkaaan observed in analysis the presented observing team in trace a and onlyand focus often on unreliable single plan as recognition there are no (Intille quantitative and Bo- means of ficultyDynamiccatedDynamiccatedicallybickbecauseprocess. with(i.e. 2001; the changed Hierarchical tracking theplans). connectionHierarchical?; game?). Consequently,structures Conversely, all is the Groupnot between between Group players segmented of forModel such groupsagents,Model team andagents, analysis into (DHGM),ball sports of (DHGM),to “discretized”toin agents, track sucha trackis confined still which as the the conducted Sukthankar soccer,which dynam- playsdynam- indi- indi-icallyasasDynamicisand input. input. oftenrecognize team-mates Avrahami-Zilberbrand composed Hierarchical team may of plans many Groupdynamically given possible Model and and an Kaminkaachange observedKaminkaa team(DHGM), plans in teamthe presented presented which in observing the trace library, indi- a a byasbyandhumans itan humans is oftenabundant low-scoring which unreliable which of means spatio-temporal means and as continuous that there that it it isare is subjective, tracing no subjective, it is quantitative extremely data unrepeatable, unrepeatable, is available difficultmeans of butand team-matesbecause(i.e.byand(i.e.There humansbick plans). only the plans). 2001; mayhave game focus whichConsequently, Consequently,? dynamically been; is? on). not means Conversely, single many segmented that such plantechniques such change it for isanalysis recognition into analysis subjective, team in “discretized” designedthe sports is is stillobserving still(Intille unrepeatable, such conducted conducted to plays as and automat- soccer, Bo-general.process.asDynamic It input. is also Avrahami-Zilberbrand Hierarchical a difficult task Group since Model an and observed (DHGM), Kaminkaa team which trace presented indi- a bick 2001;verifyingandas? often it; ? is). such low-scoring Conversely, unreliable analysis as andfor ( there? team; continuous?). are sports This no isquantitative suchit despiteis extremely as soccer, the means use difficult of space,icallyicallycatedicallyan(i.e. abundantworkThere recognizeby plans). changedthe changed humans in connectionhave this of Consequently, spatio-temporalspace teamwhich structures structures been has plans means manybetween been of of such given thattechniques groups groupslimited tracing itagents, analysis is an subjective,to of dataof observed small agents, toagents, isdesigned is stilltrack available datasets unrepeatable, conductedSukthankarteam Sukthankar the to but dynam-traceautomat-asDynamicDynamiccated input.andprocess. team-matesthe Avrahami-Zilberbrand Hierarchical connection may betweenGroup dynamically Model Model and agents, Kaminkaa (DHGM), (DHGM),change to track in presented which which the observingdynam- indi- indi- a becauseandverifyingasverifying often it is the low-scoring unreliable such game such analysis is analysis not as and segmented there (? continuous; (??; are).?). This no intoThis quantitative isit “discretized” is is despite despite extremely themeans the plays use difficult use of ofprocess.catedicallyand(i.e.icallybyandbick thebySycara plans). humansan changed oftenhumans connection2001; recognize abundant Consequently, proposed unreliable which? which; structures? of). team meansConversely, betweenspatio-temporal means another as plans there suchthat of that groupsagents, recognizingitareanalysis givenfor it is is teamnosubjective, tracing subjective, of anto quantitative is sports agents, track observed stilldata algorithm unrepeatable, such unrepeatable,conductedis the Sukthankar available means asdynam- team soccer, that of trace but en-is oftenDynamiccatedThere composed the have Hierarchical connection of been many many possible between Group techniques team Modelagents, plans designed (DHGM), to in track the tolibrary, the whichautomat- dynam- indi- an abundantandnewverifying oftenbecause statistics of unreliable spatio-temporal such the including game analysis as is there notspatial (tracing? segmentedare; ?). location no This data quantitative isinto is of available despite events “discretized” meansbeing the but use of gen- plays of andasandicallyandas input.by only it Sycara humansSycaraand is focus changed Avrahami-Zilberbrandlow-scoring often proposed proposed on which unreliable single structures means and anotherplan continuous as that recognition there of recognizing recognizingit groups and is are subjective, it noKaminkaa is (Intille ofquantitative extremely algorithm agents, algorithm and unrepeatable, presented difficultBo- meansSukthankar that that en- of en- a Dynamiccatedcatedicallyprocess.There the changed Hierarchical connection have structuresbeen Group betweenmany between of Modeltechniques groups agents, agents, (DHGM), of to todesigned agents, track track which the the Sukthankarto dynam- automat- dynam- indi- (i.e.verifyingnew plans). statistics such Consequently, analysis including (? suchspatial; ?). analysis This location is isdespite of still events conducted the being use gen- of ThereicallyCharacterizeandcodedbyicallyasandverifying havehumansan changedSycaraand input.as oftenabundant the often recognizeitbeen dynamicis whichAvrahami-Zilberbrand proposed unreliableteamsuch low-scoring structuresunreliable many of analysismeans tacticsspatio-temporal team team techniques another as as thatand of there membership plansas ( there? groups; continuousit a recognizing? isare). function are given subjective,designed This tracing no and no of quantitative isquantitativeto anagents, itKaminkaa of despitedatais prune algorithm observed to extremelyunrepeatable, the isautomat- Sukthankar theplaybook available the means means presented size that use team difficult of ofen- but the traceand a ically team-matescatedically recognize the changed connection may team structuresdynamically plans between of given groupschange agents, an of in observed agents, the to track observing Sukthankarteam the dynam-trace verifyingneweratedbecausenew(i.e. statistics statistics by such plans). the humans analysisincluding game including Consequently, ( is?) not (? andspatial; spatial segmented?). player This suchlocation location tracking is analysis into despite of of events “discretized” events information is the still being being use conductedgen- of gen- plays (?; bickDynamiccodedbecause 2001;verifying the? Hierarchical the; ? dynamic). game Conversely, such is teamanalysis not Group segmented formembership( team? Model; ?). sports intoThis (DHGM), to“discretized” such is prune despite asthe soccer, which the size plays use of indi- the of catedicallyicallyandically theThere Sycara changed changed connection recognize have proposed structures been team between another many plans of of techniquesgroupsagents, groups recognizing given of anto of agents, track observedagents, designed algorithm the Sukthankar Sukthankar dynam-team to that automat- trace en- as it is low-scoringerated by humans and continuous (?) and player it is extremely tracking information difficult (?; • andcodedandcodedplanandand SycaraasDynamicverifyingnewasverifying Sycaraoften input. library.often theit statistics is proposed dynamicunreliable low-scoring unreliable Avrahami-Zilberbrandsuch Hierarchicalproposed such Banerjee including analysis analysis teamanother as as another and there et membershipthere membershipGroupspatial (? al.continuous(; recognizing are;? areproposed). recognizing). location Modelno This Thisno quantitative and quantitative to isitto is (DHGM), ofprune is Kaminkaadespitealgorithmto prune despite eventsextremely formalizealgorithm the themeans thebeing the means size which size that presenteduse difficultuse of MAPRgen-that of en- of ofthe indi- theprocess. en- aasand input. Sycara Avrahami-Zilberbrand proposed another recognizing and Kaminkaa algorithm presented that en- a bynewerated(i.e. humans statistics plans). by which humans including Consequently, means (?) andspatial that player it such location is subjective, analysistracking of events isinformation unrepeatable, still being conducted gen- (?; icallyansize. abundant recognizebecause of team spatio-temporal the gameplans is given not tracing? segmented an? observed data isinto available “discretized” team buttrace plays icallyandicallyicallySycara changed changed recognize proposed structures structures team5 another of4 plans of groups10 recognizing groups given8 of15 anagents,of12 agents, observed algorithm5 Sukthankar4 Sukthankar10 team that8 en- trace15 12 becausenew?)erated statistics beingtheby gamehumans by generated. including humans is notwhich segmented ( spatial? means) and location playerthat into it “discretized” is tracking of subjective, events information being plays unrepeatable, gen- (?; catedplancodedplan(i.e.verifyingnewbecause the library.new plans). library. statisticsthe connection statistics dynamic such the Consequently, Banerjee Banerjee includinggame analysis including team between is notet spatial ( membership? al. suchspatial segmented; ? proposed proposed). agents, location analysis This location isinto to toof totois despite of trackprune events “discretized”formalize still formalize events conducted the thebeing being dynam- sizeuse MAPR MAPRgen- gen-plays of of theThereCharacterizeandcodedas have Sycara input. the been dynamic Avrahami-Zilberbrandproposed team many5 tactics4 team techniques another10 membership as58 a recognizing415 function designed10 and12 to8 Kaminkaa of prune algorithm15 to the× automat-12 theplaybook presented size that× of en- the a × and often?) being unreliable generated. as there are no quantitative means of as input.ascodedplan itwithverifyingDynamiccated iserated Avrahami-Zilberbrand thenew low-scoring library. a(i.e. dynamic thenew statistics by such plans). Hierarchicalconnection humans Banerjee model, analysis and including team Consequently, continuous (? revealing membership) et ( between? and;Groupal. spatial? and). player proposed This itsuch Kaminkaa locationtheModelis agents, trackingextremely isto analysis distinction prunedespite to of (DHGM), to formalize events presented informationtrack is the difficult the still size betweenbeing usethe conducted which MAPR of a ofdynam- gen-the (?; the indi-Dynamiccodedandcoded Sycara the the Hierarchical dynamic dynamic proposed× 5 team Group4 another× membership×10 Model8 recognizing× ×15 (DHGM), to to121× prune prune algorithm the the which size size of thatindi- of the the en- eratederated?)by being humansby by humans generated. humans which ( ()? means) and and player player that it tracking is tracking subjective, information information unrepeatable, ( ; (?; withby humanserated a new by which model, humans means revealing (?) that and it player is the subjective, distinction tracking unrepeatable, information between (the?; • andcodedplan SycaraasDynamic input. library. the proposed dynamic Avrahami-ZilberbrandHierarchical Banerjee1 teamanother× et membership Group al. recognizing×× proposed Model and×× to (DHGM), prune Kaminkaaalgorithmto formalize× the size which that presented ofMAPR en- the indi- a H(X, Y )= p(x, y) log p(x, y) (4) p(x, y) (i.e. plans).?In)and being this Consequently, often paper, generated. unreliable we? propose such as analysisthere a method are is no still to quantitative overcome conducted means this? is- of icallyShowwithplannewnewerated(i.e.) statisticschangedthat beinglibrary.a statistics new plans). by teams generated. humans model,including Banerjee structuresincluding Consequently, have (? revealing unique) spatial andspatialet of al. player groups locationplaying such proposed location the analysistrackingdistinction of of styles, of agents,events to events isinformation formalize despite still being Sukthankar beingbetween conducted gen- hav- gen- MAPR (?; theicallysize.plan recognize library.1 team Banerjee plans1 et given5 al.4 proposed an10 observed8 2 to15 formalize team12 trace MAPR H(X, Y )= H(X,− Y )=p(x, y) logI(Xp(;x,Yp()= yx,) y) log(4)p(x, y) p(x,(4) y) log (5) verifying) beingIn generated. this such paper, analysis we propose (?; ?). This a method is despite to overcome the use this of is-Dynamicbecauseplanwithhardnesscatedically library.? Hierarchicalerated theaby new the changed gamehumans of by Banerjee connection model, singlehumans is notwhich Group structures segmented andrevealinget (? means) al. Modelbetween and multi-agent proposed of playerthat intothe groups (DHGM), it agents, “discretized” isdistinction tracking subjective,to plan of formalize agents,to which informationrecognition, track between plays unrepeatable, indi- Sukthankar MAPRthe dynam-( the?; and catedplancodedDynamic the library. the connection dynamic Hierarchical Banerjee1 2 team between et membershipGroup× al. proposed agents, Model× to to to (DHGM), trackprune× formalize the the dynam-which size MAPR of indi- the H(X, Y )=− x X−yp(Yx, y) log p(x, y) (4) ?) beingandIn this often generated. paper, unreliable we propose as there a method are no toquantitative overcome means this is- of • hardnessanderated often?Characterize) being by of unreliable humans generated.single ( teamand? as) and there multi-agent tactics player are no as tracking quantitative a plan functionrecognition, information means of the of ( andplaybook?; as input.codedplanwithcated Avrahami-Zilberbrand the library. a dynamic thenew2 connection Banerjee model, team1 revealing2 membership et between al. and proposed Kaminkaa the agents, to distinction prune to to formalize presented track the size between the MAPR of a dynam- the the H(X,x −X Y )=y Y￿∈ ￿∈x X y pY(x, y) log p−(x, y) (4) p(x)p(y) by? humanssueverifyingIn by which this representing paper, means such we analysis that team propose it behavior is (? subjective,; a? method). Thisas a tospatio-temporal unrepeatable, is overcome despite the this use oc- is- of andingwitherated? Sycara partial)by being a humansby newproposedhumans team generated. model, which tracings. (? another) means andrevealing We player that recognizing show itthe tracking issubjective, this distinction algorithm informationby framing unrepeatable, between that (? the; en- the with a new model,2 revealing the distinction3 between the ￿∈ x￿∈X y Y ∈ ∈ x X y Y sue by representing team behavior as a spatio-temporal oc- (i.e.withhardnesshardnesssolveically plans).and a•?In) newand Sycara MAPRbeing this changed Consequently, of often model, paper, generated. single proposed problems unreliable structures we revealing and propose such another multi-agent multi-agent asin analysisthere the of the a recognizing method groups model are distinction is no plan plan still to quantitativeusing of overcomerecognition, conducted agents,recognition,algorithm between a first-cut means Sukthankarthis that the andis- and ap-ofen- icallyShowwithplancated changedthat library.a new the teams3 connection model, Banerjee structures have3 revealingunique et between of al. groups playing proposed the agents,distinction of styles, agents, to to formalize despite track Sukthankar between the hav- MAPR dynam- the ￿∈ −￿∈ ￿ ￿ ￿∈ ￿∈ newsueverifyingIn statistics by this representing paper, such including we analysis propose team spatial behavior (? a; location? method). This as ofa to isspatio-temporal events overcome despite being the this gen- use oc- is- ofcated theverifying?) being connectionsize.In generated. this such paper, analysis between we propose (? agents,; ?). This a method to is track despite tothe overcome the dynam- use this of is-Dynamicplanwithhardnessically library. Hierarchical a new changed of Banerjee model, single3 Group structures2 andrevealinget3 al. Model multi-agent proposed ofthe groups (DHGM), distinction to plan of4 formalize agents, which recognition, between indi- Sukthankar MAPR the and x X y Y and oftenIncupancysue thisnew unreliable by paper, statistics representing map webased as including propose there on team ball are a spatialbehavior nomethod movement quantitative location as to a overcome over spatio-temporal of means events fixed this windows beingof is- oc- gen- bycodedproblemsolvehardness? humanssolve) beingsueandIn theverifying MAPRIn byMAPR thisas often whichdynamic generated. this representing ofa paper,team paper, singleunreliable problemsmeans problems such identificationteamwe we analysis and thatpropose team propose membership as in it multi-agent there behavior is the the ( a? subjective,; a method modeltask? aremodel method). Thisas no. to a prunetoquantitativeplanusing using tospatio-temporal unrepeatable, is overcome overcome despite recognition, thea a first-cut first-cut size the means this this of use oc-is- is-the ap- ap-of of and• andingwithhardness Sycara partial aCharacterize newproposed teamof4 singlemodel, tracings.4 another teamand revealing multi-agent We tactics recognizing show the as this distinctiona plan function algorithm by recognition, framing of between that the the anden-playbook the ￿∈ ￿∈ p(x, y)=p(x,H y()X)+H(Y ) H(X, Y ) eratedcupancy by humans map based (?) and on playerball movement tracking over information fixed windows (?; icallyhardnesssolveproach, changednewandcodedsue statistics SycaraMAPR of theprovided by structures single representing dynamic includingproposed problems and that of team multi-agentspatiala team groupsanother in fully membership the behavior location observed modelof recognizing agents, plan as of usingto a eventsteam spatio-temporal recognition, prune Sukthankar algorithm a being trace first-cut the size gen- and and thatoc-of ap- a thecated li- en-withhardnesshardness thesolveicallyand a• connection new Sycara MAPR changed of model, single proposed4 problems between structures revealing3 and4 another multi-agent multi-agent inagents, the of the recognizing groups model todistinction track plan plan5 usingof therecognition, agents,recognition,algorithm between dynam-a first-cut Sukthankar that the and andI ap-en-(X; Y )=I(X; Y )= p(x, y) logp(x, y) logp(x, y)(5)p(x, y(5)) − suecupancynew bystatistics representing map based including team on ball behaviorspatial movement location as a overspatio-temporal of events fixed windows being oc- gen- proach,sueverifyingIn by this provided representingCharacterize paper, such we that analysis propose team a team fully behavior (? a; tactics observed? method). This as as a to isaspatio-temporal team overcomefunction despite trace the of and this usethe oc-a is- li- playbookof size.5 5 I(X; Y )=I(X; Y )=− p(x, y) logp(x, yp)(x log)p(y)(5) (5) verifyingsueof bycupancy time,erated such representing which analysis bymap humans we based team call (?; on (play-segments?? behavior).) ball and This movement player is as despite a.spatio-temporal tracking By over analyzing the fixed information use windows of a matchoc- (?; andplanproach,solve oftenIncupancy library.sue thisnew MAPR unreliable byprovided paper, statistics representing map BanerjeeCharacterize problems webased as that including propose there et on ateam al.fully ball are in team a spatialbehavior proposed the nomethod movement observed quantitativetactics model location as to to a overcome overasteam spatio-temporalusing formalize of a means events fixed function trace a this first-cut windows beingof MAPRand is- of oc- gen- a the li- ap- playbookcodedproblemsolvehardnesssolve the MAPR MAPR as dynamic ofa team single5 problems problems identificationteam4 and5 membership in multi-agent the the modeltask model. to pruneplanusing6 using recognition, thea a first-cut first-cut size of theap- ap- and − x X y Y p(x)p(y) p(x, y) of time, which we call play-segments. By analyzing a matchandsolve Sycaraproach,braryeratedcoded MAPR proposed ofShow by provided the full humans problemsdynamic that team another teams that (plans) team anda in recognizing havefully were the player membership unique observedmodel given tracking algorithm as playingusing inputteam to information prune a trace( styles, first-cut that). the and en- despitesize ( ap-a; li- ofically the hav-hardnesssolveproach, changedand SycaraMAPR of provided structures single proposed problems6 and that of multi-agent a groupsanother in fully the observed modelof recognizing agents, plan using team recognition, Sukthankar algorithm a trace first-cut and and that ap- a li- en- x −X y Y ∈ −∈ p(x)p(y) p(x)p(y) ?cupancy)of beingerated time, generated. mapwhich by humans based we call on ( play-segments) ball and movement player tracking. By over analyzing fixed information windows a match ( ; Automaticallysueplancupancynewcupancy by library.•statistics representing• map map analyzeBanerjee based including based? team on an on ball et unseenbehavior spatial ball al. movement movement proposed location match as a overspatio-temporal ofoverin to events termsformalize fixed fixed? windows being of windows tac- MAPR? oc- gen- coded the6 dynamic team membership to prune the size of the I(X; Y￿)=∈ x￿∈X￿y Y￿x X y pY(x, y) log (5) newcupancy statisticsasof as? time,) a being mapplaybook including which based generated. weof spatialon call play-segments,? ballplay-segments location movement of weevents over. By can analyzingfixed being characterize windows gen- a match the? verifyingwithproach,suebraryof• a bycupancy time,erated newof such representing provided fullsize. which model, analysis bymap teamsize. humans we based plansthat team revealing call (?; on a(play-segments? were? behavior). fully) ball and This given movementthe observedplayer is as distinction despiteas a.spatio-temporal tracking Byinput over teamanalyzing the (? fixed). information betweenuse trace windows of a matchoc- and the ( a?; li- planproach,solveproach, library. MAPR provided providedCharacterize BanerjeeCharacterize6 problems that5 et6 a team al.fully in team proposed the tactics observed observedtactics model as to7 ateam asteam using formalizefunction a tracefunction trace a first-cut of and MAPRand the of a a li- playbook the li- ap- playbook ￿∈ −￿∈ ￿∈ ￿∈ p(x)p(y) coded• proach,brary theplan) being dynamicof ofprovideding library. time, full generated. partial teamwhich team Banerjeethat teamplans wemembership a call fully tracings.wereplay-segments et observed al. given toproposed We prune as team. show input By the toanalyzing trace size( this formalize?). of byandthe a matchframing a MAPRli-andsolve theSycaraproach,brarycodedplan MAPR proposed ofShow library.• provided the full7 problemsdynamic that team Banerjee another7 teams that plans team a in recognizing havefully wereet the membership al.unique observedmodel given proposed algorithmas playingusing inputteam to to prune formalizea trace( styles,? first-cut that). the and en- despitesize MAPR ap-a li- of the hav- = H(X=)+HH(X(Y)+x) XHyH(Y(X,) YSummaryH) (X, Y ) and Future Work ofIn time,as this as whichplaybook paper, a playbook we we call proposeofplay-segments play-segments, a method. Byweto overcome cananalyzing characterize this a match is- the tics.brary?cupancywithoferated time, of a fullnewmap which by team humans basedmodel, we plans call on (? revealingplay-segments) wereball and movement given player the as tracking. distinction By input over analyzing (fixed? information). between windows a match (? the; Automaticallysize.• analyze7 6 7 an unseen match10 in terms of tac- = H(X=)+H￿H∈(−(YX￿)∈)+HH−((X,Y ) Y ) H(X, Y ) eratedofas time,behavior? byas) as being as humans a whichIn a thisplaybook of generated. eachwe paper, (?) callof and teamof play-segments,weplay-segments play-segments, playerwhich propose tracking allows a method. Bywe we for information analyzingcan can tactical to characterize characterize overcome analysis a( match?; this the the of is- newhardnessbrarycupancy statisticsasof as of? time,) aof being mapfullplaybook including single which team based generated. we andplansof spatialon call play-segments, ball multi-agentplay-segments were location movement given of weevents plan as over. By caninput analyzingfixedrecognition, being characterize (? windows). gen- a match theand • withbraryproach,braryplan• a of newof library. fullprovided full10 model, team teamsize. Banerjee10 plans that revealing a were fully et al. given given the observed proposed distinction as as input input team to ( (? formalize?). between). trace and theMAPR a li- − − behavior of each team which allows for tactical analysis ofplanbrary library.ofwithhardnessIn of time,as this fullproblem Banerjeea as which newpaper,Show ateam ofplaybook single model, weas plans thatwe et a call propose al.team teamsof wereandplay-segments play-segments,proposed revealing identification multi-agent have given a method unique toas the. inputformalize Byweto distinction plan playing overcometask cananalyzing (?. characterize). recognition, MAPR styles, this abetween match is- despite the andcoded the5 proach, hav-tics.4brary thewith10 dynamic ofprovideding a8 full new partial15 team team10 model, that12 plansteam membership7 a10 fully revealing were tracings. observed given to the We prune as distinction inputteam show15 the (trace size? this). ofbetween byandReferences the framing a li- the the = HTo(X Do)+H(Y ) H(X, Y ) ?) beingsueasbehavior as by generated.In a representingplaybook this of paper, eachof team we team play-segments, propose which behavior allows a method as we a for spatio-temporal can tactical to characterize overcome analysis this oc- the of is- eratedofas time,behavior?English byas) as being as humans a• whichInplaybook a thisplaybook of generated.Show eachwePremier paper, (?) callof and team thatof play-segments,weplay-segments play-segments, playerteamswhich propose League tracking allowshave a method. Bywe Ballunique we for information analyzingcan can tactical to Action characterize playingcharacterize overcome analysis a( match? Data; styles, this the the of is- despite×hardnessbrary hav-× of of full15 single team× 15 andplans multi-agent were given plan as input recognition, (?). and − asa asbehavior team’s asueplaybook by performance. representingof eachof play-segments, team We team which show behavior allows the we efficacy can foras a characterizetactical spatio-temporal and usability analysis the of oc-withsolveShow asueas newhardnessEnglish as MAPRby thebehavior a model, representingplaybook utilizationing• of problems ofPremier partial revealingsingle eachof team ofplay-segments, team team and thisin behaviorLeague thewhich the tracings. multi-agent approach distinction model allows as Ball we a We spatio-temporalonusing for can planAction show an tacticalbetween characterize aiPad, recognition, first-cutthis analysis Data which theby oc- theframing ap- of1 plan andbrarylibrary. thewithhardness of fullproblem Banerjeea newShow team ofShow15 single model,as plans that et a10 that al.team teams15 wereand proposedrevealing teams identification multi-agent have given haveAllen, unique toas the unique inputformalize N.; distinction20 plan playing task Templon, (? playing.). recognition, MAPR styles, J.; between styles, McNally, despite and despite the P.; hav- Birnbaum, hav- L.; and cupancybehaviora team’sa team’s map ofperformance. eachperformance. based team on ballwhichWe We show movement show allows the the efficacy for efficacy over tactical fixed and and usability analysis windows usability of of • ?) beingassolvebehaviora asEnglishEnglishbehavior team’s agenerated.InsueAutomaticallyplaybook MAPR this by performance.of representingof paper,ing eachPremier each problemsof partial team play-segments, we team analyze propose Wewhich team Leagueteam Leaguewhich in show behaviorthe allows antracings. a allows themethod modelwe unseen Ball Ball efficacy for can foras tactical Wetoa Actionusingcharacterize tactical Action match spatio-temporal overcome and show a analysisusability analysis in first-cut Data Datathis terms the this byof of2oc- is- ap- of framing5 5 tac-solveShow54545451045English104 theMAPR the5104584105•10554810 utilization845152010448101548 problems151084 Premier101281015102081215101281581581512 of8151281512 thisin1515121512 League the12 approach121212 model Ball onusing an Action aiPad, first-cut which Data ap- SummarySummary and Future and Future Work Work Inbehavior thisoursuea team’scupancy paper,method by of representing each performance. we onmap propose team the based 2010-2011 which team a Weon method ball behaviorallowsshow English movement the to for overcomeefficacy as tactical Premier a spatio-temporal over and analysis this fixedLeague usability is- windows of soc- of oc- proach,showscupancya it’s providedteam’s viabilitymapproblem performance. based that to as on bea a fullyballteam used We movement observedidentificationshow for amateur the efficacy over team level task fixedtrace and. use. windows usability and a ofli-with×× a×× newhardnesssolve×English×English××× model,× MAPR×××ing×•× of××20 PremierPremier partial× revealingsingle× problems×××××× team20 and×× League× the× in× tracings. multi-agent× the distinctionHammond, model Ball Ball We25planAction showusing ActionK. between 2010. recognition, this aReferences first-cut DataStatsMonkey: DataReferencesReferences thebyReferencesReferences framingReferencesReferences ap-References and AReferencesReferences Data-DriventheReferencesSummarySummary Sports and Future and Work Future Work our method on the 2010-2011 English Premier League soc-hardnessInbehaviorbehaviorsolveproach,a thisEnglishsue• team’sof paper, single by MAPR ofprovidedof performance.representing each weeach Premierand propose teamProblemproblems team multi-agent that which team whichWe aLeague methodfully show in behaviorallows Formulation allows the observedplan the to modelfor efficacy overcomeBall for as recognition, tactical a tactical spatio-temporal team using Action and analysis this traceusability analysis a is-and first-cut Data and of of1 oc-3 a1 li-1 ap-•111 1 111English11 1Automatically25 ing Premier25 partial20 analyze Leagueteam antracings. unseen Ball We Action match show in Data this termsTo Do by ofTo framing Dotac- the sueof byaour team’s time,cupancy representing method which performance. map on we team the based call 2010-2011 behaviorplay-segments We on ballshow as movement English the a spatio-temporal. efficacy By Premier analyzing over and fixed League usability a oc- match windows soc- of Englishoura team’scupancytics. method Premier performance.problem onmap the based 2010-2011 as League Weon a team ball show English movement identification Ballthe efficacy Premier Action over and task fixedLeague usability Data. windows soc- of hardnessproach,showssolve of it’s provided single MAPR viabilityproblem25 and that problems to asmulti-agent bea a25 fullyteam used in observedidentification for theAllen, planNarrative amateurAllen,Allen, modelAllen, N.;Allen, recognition,30 teamN.;Allen, N.; Templon,Writer.Allen,Allen, levelN.;usingTemplon, task Templon,N.; Allen, traceN.;Allen, Templon,Allen, In.Templon,Allen, Templon,N.; use.N.; J.;Templon,AAAI a N.; Templon,and J.;McNally,Templon,N.;and J.;first-cut N.; N.; FallMcNally, Templon,J.;McNally, Templon,J.; a Templon,J.; Templon,To McNally,J.; Symposiumli- McNally, McNally, DoJ.;P.; McNally, ap- McNally,P.;Birnbaum, J.; J.; J.; P.;Birnbaum, McNally, P.;McNally, Series P.;McNally, P.;Birnbaum, Birnbaum, Birnbaum,P.; L.;.Summary Birnbaum, P.; L.;and P.; P.; Birnbaum,L.; andBirnbaum, L.; Birnbaum, L.; and and L.; and and and L.; L.; Future L.; L.; and and and and Work a team’scerourof data. method performance.time, which on the we We 2010-2011 call showplay-segments the English efficacy. Premier By and analyzing usability League a of soc-match braryofa team’s time, ofour full whichmethod performance. team weProblem on plans call theplay-segments 2010-2011 wereWe show Formulation given the English as. efficacy By input analyzing Premier (and?). usability League a match soc- of242 2222 2English2proach,22English•22 2 provided30 Premierproblem Premier30Problem25 that as League a a League fullyteam Formulation observed identification Ball Ball Action team Action task trace Data. Data and a li- To Do asour as methodcer a playbook data. on theof 2010-2011play-segments, English we can Premier characterize League the soc- solveThesue MAPRa rest byproach, team’sbraryourcercupancy representingour of methodofdata. theofmethod performance.problemstime,Automatically providedfull paper map whichon team onProblem team the based describesthe in we 2010-2011plans that Webehavior 2010-2011 the call on showaanalyzewere ballmodelplay-segments fully Formulation the asthe movement English given English method aobserved anefficacyusing spatio-temporal unseen as Premier. Premier inputa Byof over and first-cut teamdoing analyzing match usability fixed( League? League). trace oc- this,in1windows ap- a ofsoc-terms soc-matchand in a of5 li-brary tac- oftics. full9 team30Problem plans13 30 were Formulation given17Hammond,Avrahami-Zilberbrand,Hammond,Hammond, asHammond,Hammond, inputHammond, K.Hammond,Hammond, K. K.2010.(?Hammond,Hammond,). 2010.K.Hammond, 2010.Hammond, K. K. K.StatsMonkey:2010. 2010. 2010. K. K.D.;StatsMonkey: 2010.StatsMonkey: 2010.K. 2010.StatsMonkey:Banerjee, K. StatsMonkey: K. StatsMonkey:K. StatsMonkey:2010. 2010. 2010. 2010.StatsMonkey: A Data-DrivenTo B.;StatsMonkey:A StatsMonkey: DoStatsMonkey:Data-Driven AKraemer, A Data-Driven A AData-Driven Data-Driven A Sports Data-Driven L.; A SportsA AData-Drivenand Data-Driven Sports Data-DrivenSports Sports Sports Sports Sports Sports Sports cupancyourcerof method data.map time, based whichon the on we 2010-2011 ball call movementplay-segments English over Premier. fixed By analyzing windowsLeague soc- a match t0,t1cer,t2 data.,t3,t4,t5,t6,t7,t8,t9 353solve3333 3 MAPR3proach,brary3333 3 of problemsAutomatically providedfull teamProblem in30 plansthat the aanalyze were model fully Formulation given observed anusing unseen as inputa first-cut team match (AAAI). traceAAAIAAAIin ap-FallAAAIAAAI termsAAAIFall Fall andSymposiumAAAI FallAAAI SymposiumAAAI SymposiumFall Fall a Fall ofSymposiumAAAI li-SymposiumAAAI FallAAAI SymposiumAAAI tac-Symposium Series Fall Symposium Fall Series Fall Symposium Series Symposium Series Symposium Series Series Series Series Series Series ceras data. as a playbook of play-segments, we can characterize theproach,cupancyourtas0ourbrary providedcer,t asof method1 method,t data.map atime,Show2•playbook of,t based3full whichon,t that ontheAutomatically4 the,t team the onProblema5of utilization we,t 2010-2011 fully 2010-2011ballplay-segments,6 plans,t call7 movement,t observedplay-segments8 were,t ofanalyze EnglishFormulation9 English this given overweteam approach Premier an can Premier. as fixed By trace unseen inputcharacterize analyzing windowsLeague and Leagueon (? match). an a soc- li-a iPad, the soc-match in terms whichThet ofrest0,t tac-1 of,t2 the,t3 paper,t4,t5 describes,t6,t7,t8Table,t the9Narrative 2:methodLyle,NarrativeNarrative TableNarrative J.Narrative 2010.Writer.Narrative showing ofWriter.Narrative Writer.Narrative Multi-Agentdoing Writer.Narrative InNarrative Writer.NarrativeNarrative? Writer. theIn In Writer. Writer. this,In identification In In Writer. Writer.In PlanWriter. Writer. inIn In Recognition: In In In In rate by Formalization breaking...... behaviorcer data. of each team which allows for tactical analysis of additiont0,t1cer,t toas2 data.experiments,t astics.3 a,tplaybook4Problem,t5,t6,t whichof7,t play-segments, Formulation8,t show9 the suitability we can characterize of these46 the4proach,4444Table4t40brary provided4,t4 2:441Table,t Table4Show2• of,t 2:3 full showing,t that Table theAutomatically4,t teamProblem a5 utilization showing,t fullythe6 plans,t identification7,t observed the8 were,t identificationof analyzeFormulation9 this given rate team approach anby as rate breakingtrace unseen input by and breakingon ( match). an a li- iPad, in terms which of tac- 5 4 10 8 15 12 of time,ceras data. which as a playbook we call play-segmentsof play-segments,. By analyzing we can characterize a match the tbehavior0cer,tWe•1 data.,t refer2,t of3•,t each to4,tteam5 team,t6,t behaviors which7,t8,t allows9 , , for as short,tactical observable analysis of seg-additionat team0,tTable1 over,tto2 2: experiments,t aTableTabletics.3 match.,t4Problem,t showing 2:5 TableGiven,t6,t which the showing7 that,t identificationFormulation8the,t ashow team’s9 plans theAvrahami-Zilberbrand,andAvrahami-Zilberbrand, identificationAvrahami-Zilberbrand, the into Algorithms. possessionAvrahami-Zilberbrand, rateAvrahami-Zilberbrand, suitability differentAvrahami-Zilberbrand, byAvrahami-Zilberbrand,Avrahami-Zilberbrand, breakingAvrahami-Zilberbrand,Avrahami-Zilberbrand, rate InAvrahami-Zilberbrand, stringtimeAvrahami-Zilberbrand,AAAI? byof D.; series breakingD.; theseis. D.;Banerjee, D.;Banerjee, Banerjee, andD.; D.; D.; Banerjee, Banerjee, the Banerjee,D.; B.; Banerjee, field Banerjee,B.;D.;Kraemer, D.; D.; B.; Kraemer,Banerjee, intoB.; Banerjee, B.;Banerjee, B.;Kraemer, Kraemer, differ- L.;B.; Kraemer, andL.;Kraemer, B.; B.; L.;B.; and L.;Kraemer, Kraemer, L.;and Kraemer, and L.;and and and L.; L.; L.; L.; and and and and behavior of each team which allows for tactical analysis ofbraryof oftime,cerEnglish fullWeas data. whichshows as teamrefer a playbook we Premierplanstotics. it’s callteam viability wereplay-segmentsof behaviors play-segments, Leaguegiven to be as, . used Byinput, Ball as analyzingwefor short, (? Action). can amateur observable characterize a match Data level seg- use.5 the75 5555the5t05a plans,t5We team5•155the,t5 into refer2over plans,t3 different•,t ato into4 match.,tteam5 different,t time6 Given,t behaviors series7,t time that8,t and series9 a team’s the, and field, as possession the into short, field differ- into observable string differ- is seg- 5 ×4 10 8× 15 12× a team’s performance. WeRelated show the Work efficacy and usability of taskst0 we,tWe1 mentioned.,tbehavior refer2,t3,t to4 ofteam,t each5We,t6 behaviors first,t team7,t describe8 which,t9, B allows, our as short, dataset. for tactical observable analysis seg- ofbraryof oft lengthEnglish,tthe fullWe plans,tshows teamreferTTable1,t, into the,t Premierplansto 2:tics. it’s differentresultingteam,t Table viability were,t showing behaviors time,t number Leaguegiven,t series to the,t be ofas, andLyle,identificationusedBanerjee, play-segments inputLyle,,Lyle, Ball the as J.Lyle, 2010.Lyle, fieldJ.for short, J. (Lyle, 2010.? 2010.Action B.; J.). amateurLyle, J.intoLyle,Multi-Agent 2010.rate Kraemer, J.2010. observableLyle, Multi-Agent Multi-AgentLyle, 2010. differ-J.Lyle, J.forbyLyle, Multi-Agent2010. 2010. Multi-AgentJ. Multi-Agentbreaking eachDataJ. Multi-AgentlevelJ. 2010. J.L.; 2010. Plan 2010.Multi-Agent 2010.Multi-Agent and Plan Plan seg-Multi-Agent Recognition:use. Multi-Agent Lyle,Multi-AgentPlan Multi-Agent PlanRecognition: Recognition: Plan Plan Recognition: J. Recognition: Plan Recognition: 2010.Recognition: PlanFormalization Recognition: Plan Plan Formalization Multi-Agent Recognition: Recognition: Formalization Recognition: Formalization1 Formalization Formalization×5 Formalization Formalization5 Formalization Formalization4 410× 108 815× 1512 12 as as abehaviorplaybook ofof each play-segments, teamRelated which we allows Work can characterize for tactical analysisthe of t0,tmentsa1 team’sWe,t2English,t refer of3 performance.,tShow coordinated4 to,t5team,t Premier the6,t behaviorsutilization7 We,tRelated movement8,t show League9 , the ofB Work, thisefficacy as and Ball short, approach action and Action observable usability executed on anData of iPad, seg- by atasks whicht0ent,t0 we1quantizationWe,t1a2 mentioned.,t team refera23 team,tthe3 over4to areas.,t plans over45team a,t match.5 We The6 a into,t match.6 behaviors7 resultsfirst different,t Given78,t Given describeent8 are9 that time quantization9 the, that aB traceteam’s series, a our as team’s of short, and dataset. possession the areas. possessionthe team observable field The string results into string differ- is are seg- is the trace of the team 5 4 10 8 15 12 a team’s performance. We show the efficacy and usability of as as abehaviorplaybook ofof each play-segments, teamRelated which we allowsB Work can characterize for tactical analysisthe 106 of6 6666 6ments6of6We6 length66entEnglish6 refer ofthe quantizationTaShow coordinated plans1 team, to theteam intoPremier resultingover the areas. different behaviorsutilizationa match. movement The number Leaguetime( resultsT Given1,and ofseriesTB)+1and, areplay-segmentsand that Algorithms.this as andand Algorithms.theBall Algorithms.and short,a approach team’sand traceactionAlgorithms. the Algorithms.andand Algorithms. Action field observableInofand Algorithms.and Algorithms. possessionandAAAI fortheandIn executedIn into Algorithms. Algorithms.onAAAIAAAIAlgorithms. eachteam In Algorithms.. In Indiffer-AAAI anData InAAAI..AAAIAAAI In IniPad,string seg-. AAAI. byAAAI. In. In InAAAI Ina is whichAAAI..AAAIAAAI. . .. 1 × × × × × × Related Work mentsWea refer of team’s coordinated to performance.team behaviorsRelated movement We show, WorkB and, the as actionefficacy short, executed observable and usability2 by seg- ofa 6 ent quantization10 areas.14 The resultsrecognition.− areBPlan the Recognition: trace of the teamFormalization and Algorithms. In 2AAAI. × × × our method on the 2010-2011Related English Work Premier League soc- mentsEnglish of• coordinatedShow Premier the movementRelated utilization League Work and of thisBall action approach Action executed on Data anby iPad, a possessionrecognition.mentsments whichWeof ofreferlength of• is lengthcoordinatedent coordinated therefore: toT quantization1,teamT the1, the resultingN behaviors resulting movement areas.= number( The18T number1T,BT resultsand)+1 of and, as. play-segments of If action are actionplay-segmentsshort, the the possession trace executed observable executed forof the for each teamby each by seg- a a 1 1 behaviorWitha team’s of the each obvious performance. team whichapplications We allows show to for sport the tactical efficacy and military analysis and usability domains, of of behaviorWementsteamourWithaThe refer method team’s (e.g. of of therest each tocoordinatedshows passobvious performance. onteam of team thethe from it’s behaviors 2010-2011which applications paper viability agent movement Weallows describes A, show Englishto to to, forbe agent sportas andthe used tactical short, the Premier efficacy andaction B) for method military. analysisobservable Aamateur Leagueand executedplan ofusability domains, ofcandoing level soc- seg- bybe157 use. of7 de- athis,7777 in7Wementsteam7possessionrecognition.77The7 refer7recognition.English (e.g.7 ofent rest tocoordinatedofshows quantization pass is lengthteam ofShow therefore:the fromPremier it’sT behaviors1 paperthe, viabilityareas. theagent movementN utilization resulting= describes The LeagueA, to toresultsBanerjee,−, be number agent asBanerjee,Banerjee, and usedof short, areBanerjee, the.Banerjee, thisBall actionIfB.; B)Banerjee, the of formethod the B.;Banerjee, Kraemer,B.;Banerjee, play-segments. traceapproach observable Aamateur Kraemer,possessionActionBanerjee,B.; Kraemer,Banerjee, B.;executed B.;Banerjee,planBanerjee, ofB.;Kraemer, Kraemer, Kraemer,of theB.;L.; B.;Kraemer,canL.;doingand levelKraemer,L.; teamB.;Kraemer, on B.; Data B.; seg-B.;andL.;and forbybe Kraemer,Lyle, L.; anKraemer, L.; Kraemer, use.Kraemer,andL.; eachLyle,de- Lyle,aandthis, iPad,and J.L.; and Lyle, Lyle,2010. J.and Lyle, L.; Lyle,in L.; 2010. L.;J.which andLyle,J. andMulti-Agent 2010.J. J.and 2010. Lyle, 2010.Multi-Agent Lyle,J. Lyle,2 Multi-Agent2010. Multi-Agent J. J. Multi-Agent J.2010. 2010. Multi-Agent2010. 1 Multi-Agent Multi-Agent Multi-Agent Multi-Agent Withour method the obvious on the applicationsRelated 2010-2011 Work to English sport and Premier military League domains, soc- team (e.g.our method• passProblem from on the agentRelated 2010-2011 Formulation A to Work agentB English B) . Premier A plan Leaguecan be de-soc- string is smallerrecognition. inProblem duration than FormulationT ,( weTBeetz,TB1 (T discardT1)+1 M.;T )+1 von it. To Hoyningen-Huene, repre- N.; Kirchlechner,3 B.; Related Work EnglishteammentsWith (e.g. of Premier thecoordinated pass obvious fromProblem LeagueRelatedapplications agent movement A BallWorkFormulationtoB toagent sport andAction B) and action . Amilitary Dataplan executedcan domains, be1020 by10de-1010 a10English1010mentsteam1010101010possession10 (e.g. ofpossessionrecognition. Premier coordinated• pass is from therefore: is League therefore: agent movementN AN= toBallB= agent− andAction−( B)T1 action. .T If A)+1. the Ifplan Data the possession executedcan possession be de- by a 2 2 a team’scerWithresearchour data.With performance. themethod the obvious interest obvious on theWe applications into applications2010-2011 show MAPR the to efficacyhas to Englishsport sport grown and and and Premier quite military usability military substantially League domains, domains, of soc- aments team’steamfinedcerWithresearchouraddition data.Withof (e.g.performance. as themethodcoordinated anthe passobvious interest orderedobvious toshowson from experiments theWe applications into applicationsit’smovement agent sequence2010-2011 show MAPR viability A the whichto to efficacyhas ofto agent andEnglish sport to sport team grown be show action andB) and usedand behaviorsPremier . quite military Ausability themilitary forexecutedplan suitability substantially amateur League domains,can domains, ofdescribing be by soc- levelde- ofa these use.mentsteamteamfinedstringaddition of (e.g. (e.g. as is coordinated smalleranpossession pass ordered toshows from experiments inProblem duration isit’smovement agent agentsequence therefore: viability than A A whichto FormulationtoTPlan ofN agent, agentandGedikli,to wePlanPlan team= Recognition:T be discardshowPlanaction Recognition:Plan B)Recognition:T B) used S.;Plan−behaviors Recognition: .PlanRecognition: .Siles,Plan A the A it.Recognition: forexecutedPlanplanPlanFormalizationRecognition:plan To Recognition:.Plan suitabilityPlan F.; If amateurFormalization Formalizationrepre-Recognition:Recognition:Durus, thecan Recognition:can Recognition:Formalization describing Formalization Formalization possession Formalizationbe by be M.; Formalization Formalization andlevelde- ofade- and Formalizationand and Algorithms. Formalization these Formalization Formalization Lames, use.and Algorithms. and and and Algorithms. Algorithms. and Algorithms.M. Algorithms. In Algorithms.and 2009. andAAAI3 Inand Algorithms.AAAI Algorithms.In. Algorithms.AS-InAAAI InAAAI.. AAAIAAAI In2. .AAAI. In. In InAAAI In..AAAIAAAIAAAI. . .. cer data. Partialfinedcer as TeamThe an data. restordered Tracing of the sequence paper from describes of team Ball behaviorsthe Action method Datadescribing of doing15251515 this,151515Partialtsent15,t15 in1515 each,t1515string15,tstringTeam play-segmentThe,t is smallerrest is,t smaller Tracing,t of inthe,tp duration in,t= paper duration,t fromp0,...,p thandescribes thanT Ball,T weT ,1 wethe discard Action,T the discard method quantized it. To it. Dataof Torepre-doing repre- this, in 43 3 Withresearch the obvious interest applications into MAPR to hassport grown and military quite substantially domains, t0finedteam,tWith1research,tresearch2 (e.g.as the,t an3,t obvious interestpass4 ordered,t interest5 from,t applications6Problem into,t sequence into7 agent,t MAPR8 MAPR,t9 A to has toFormulation of hassport agent team grown grown and B)behaviors quite military quite . A substantiallyplan substantially domains, describingcan be de- 0finedteamfinedsent1 2 (e.g.aseach as an3 an play-segment pass4 ordered ordered5 from6Problem sequence7 agentp =8 9 Ap ,...,p oftoFormulation ofBeetz, agentBeetz,team teamBeetz,−Beetz, M.;Beetz, behaviorsTM.;Beetz, B)behaviors M.; von,Beetz, theBeetz, M.; . von M.;von A Hoyningen-Huene,Beetz, quantizedM.;Beetz,Beetz, vonplanBeetz, Hoyningen-Huene,von Hoyningen-Huene,M.;von M.; von describing Hoyningen-Huene, M.;describing Hoyningen-Huene,vonM.;Hoyningen-Huene, voncan M.; Hoyningen-Huene,M.; von Hoyningen-Huene,von Hoyningen-Huene, vonbe von Hoyningen-Huene, Hoyningen-Huene,de- N.;Hoyningen-Huene, Hoyningen-Huene, N.; Kirchlechner, N.; Kirchlechner, N.; N.; N.; Kirchlechner, Kirchlechner, N.; Kirchlechner,4 Kirchlechner,N.; B.; N.; N.; B.;Kirchlechner, Kirchlechner, Kirchlechner, B.;3 B.; B.; B.; B.; B.; B.; B.; B.; ourWith methodresearchrecently.cerresearch the data. onobvious interest The the interest2010-2011 significance applications into into MAPR MAPR English of to considering has sport has Premier grown grown and military group quite League quite substantially actions substantially domains, soc- in or- ourteamfinedaWith methodt reciperecently.0 (e.g.certasks,tresearch the as1 data.,t pass onobviousan2 used we,t The the ordered3 interest mentioned.from,t by2010-2011 significance4 applications,t a5 agent,t intoteam sequence6,t MAPR7 A We English,t to of to8 to achieve,t considering first agentof sport9 hasteam Premier describe grown and B) a behaviors goal. military groupA League quiteplan our (? actions). substantially domains,dataset.cansoc- A describing team be in de- or-20 per-20202020team20balleach20fineda2020t recipe position20 time0 (e.g.20tasks20,teach20 as1 step,t pass an time2 used westring,t atis ordered3 then each stepmentioned.from,t by4 is,t used is smallertime a5 agentthen,t teamsequence to6 step,t usedpopulate in7 A{ We,t to isdurationeach to to8 achievethen,t0 populate first agentof each9 timePOGAMO: team used describe entry.than stepT B) each a to1 behaviors} isgoal ., Apopulate entry.then we Automatedplan our ( discardused?). dataset.can A todescribingeach Sportspopulate teamit. be To de- repre-Game per- each Analysis entry. Models.5 In- t ,ta recipe,trecently.,taddition usedProblemThe,t The by rest,t to significance a experiments teamof,t Formulation the to paper achieve of considering which describes a goal show group the( the). method actions A suitability team in of per- or-30 doing oft10 these,t this,aWe11each recipe,tsent refer12 in timesent,taddition each used13Problem steptoThe each,t play-segmentteam14 is by rest,t play-segmentthen to a15 behaviorsexperiments teamof used,t Formulation16 the to top populatepaper{= achievep, =pGedikli, which0, describes,...,pGedikli,p aseachGedikli,0,...,p aGedikli, short,−Gedikli, S.; goalentry. showTGedikli,} S.;Siles, S.;Gedikli,1TGedikli, Siles,the(S.;Siles, observable,? S.;Gedikli,1 the F.;Gedikli,). Siles,Gedikli,S.;,Gedikli, Siles,method Siles,Durus, AtheF.; quantizedF.;S.; suitabilitySiles,S.; Durus,team Durus,F.; quantizedSiles, Siles,S.; F.; S.;F.; Durus,S.;M.; F.;S.; Siles,Durus, Durus,Siles, seg-of F.;per-Siles,Durus,M.; F.;M.;Siles, and doingDurus, Durus,M.;F.;and of Lames,M.;F.; M.;F.; F.; Durus,M.; andthese Durus,Lames, and Durus, M.;and this,and Lames, M. Lames, andM.; Lames, Lames,M.; M.2009. M.; in Lames,and M.and 2009. M. and AS-M.Lames, 2009. Lames, 2009.4 Lames, M.AS- 2009.4 AS- 2009. M.AS- M. AS-M. AS-M. 2009. 2009. AS- 2009. 2009. AS- AS- AS- AS- researchrecently.recently. interest The The significance significance into MAPR of of considering has considering grown group quite group actionssubstantially actions in in or- or- 10cerafinedWe data.11research recipetrecently.,t refer12 as,t used an13 interest,tto The orderedteam,t14 by significance,ta into15 behaviors team,t sequence16 MAPR,t to of,t achieve considering, has,t of, grownas team short,a goal group quitebehaviors observable (?? actionssubstantially). A team3 describing in seg-or-25 per-252525725fined2525afined25Aball recipet25 description250 as2525,t position25 as111 an,t used an2 orderedsenteach,t orderedof3 at,t eachby time this each4,t aprocesssequence step play-segment515 teamtime,t sequence6 is,t stepthen is to7,t given usedachieveis of8A,t{19 thenp description teamof9 international to{=Figure populateusedteamp a behaviors0,...,p goal to− 4. behaviorsofpopulate Journal each Given}− this (T?}). entry. process1 describingof A, each Computer the team describingis quantized given per- Science in Figure in Sport 4. Given8(1).5 4 cerresearch data.derrecently. to isolateinterest The team significanceintoFigure plans, MAPR 3: ratherFigure of has considering grown than 3. a quitesequential group substantially actions process in or- of finedaformingresearch recipe0der asrecently.We1 anto refer isolateinterestused a2 ordered group The3 toby team4 significanceintoteam aofFigure sequence5 teamRelated theseplans, MAPR6 behaviors 3: to7 plans ratherFigure ofachieve has of8 consideringWorkB team9 grown, to than 3. achieve, a as abehaviors goal quitesequential short, group ( a? substantially). major observable actions A describing process team goal in per-or- of (e.g. seg- entry.aforming recipeAWeball descriptionAballeach positionrefer useddescription aaddition groupposition time toby ofatstepteam aof eachthis ofat team isto these eachthis thenprocess behaviorsexperiments time process totime used plans step achieve is stepto givenBPOGAMO: is populate, to thenPOGAMO:givenPOGAMO: is achieve,{ inwhich thenaPOGAMO: asPOGAMO: usedFigure goal inPOGAMO:each short, usedAutomatedFigurePOGAMO:POGAMO: to showAutomated ( Automateda 4.entry.?POGAMO: populate toPOGAMO:). major−GivenPOGAMO:Automated observable 4.POGAMO: Automated populateAutomated A} GivenAutomatedthe Sports team Automated Automated goalSports eachSportssuitability Automated AutomatedGameSports each per-Automated Sports Automated Sports(e.g. seg-Game SportsGame AnalysisSports SportsGame Game Game Analysis GameSports of Sports Sports AnalysisGame these Analysis Models. Analysis Analysis Game Game Models. AnalysisGame Models.6 Models. AnalysisIn-5 Analysis Models. AnalysisIn-5 Models.In-In- Models.In- Models.In- Models. Models.In-In- In-In-In-In- der to isolate teamRelated plans, rather Work than a sequential process oft ,tThere,tformingrecently.,t der has,ttasks,t to a been Theaddition groupisolate,t we significance,t an mentioned. ofteam,t explosion tothese,t plans,experiments plans of ratherconsidering We in to thefirst thanachieve which interest describe a groupsequential a show major in ouractions live-sport the dataset.process goal suitability in (e.g.or-30 of30t30030,t30301mentsTherethe of,t30formingaforming302entry. possession30recipe these,t3030303 ofhas,t304tasks coordinated a,t aused beenball group5 group string,tA we6 position description by,t an mentioned.a of7 of a,t explosion= these team8movement,t ata9 each of0,...,a plans tothisthetime We achieve in process possession13 to to thefirstCohen, stepandB achieveshown, achieve interestis describe actiona P.,given thengoal string andwe a a usedin major firstLevesque, inmajorexecuted (a ourFigure?). live-sport=to Adataset. populate goal goala 4. team H.0,...,a Given by (e.g.1990. (e.g. eachper-a13 Intentionshown, is we Choice6 first with 5 recently.der to isolate The significance team plans,Related of rather considering than Work a sequential group actions process in or- of 0 1mentsformingarecently.2 recipederrecognizingder3We of to4 to coordinatedisolate areferThe used isolate group5 significance6 toplans byteam team of7team a these ofplans,team8 movement plans, individual behaviors9 ofRelated plans torather considering rather achieve to than agents. and, thanB achieveWork, a actionaas group sequential sequential goalOutside short, a actions majorexecuted (?Table). observable of process process A the in2: goal teamTable or- sport by of (e.g.of showing per-a seg-a recipe theformingwinningAmentsthe identificationdescriptionWeentry.the possession usedentry. possession referaof aA group match), descriptioncoordinated by rateof to string athis ofteam by stringcan these breakinga process of{= behaviorstoa bethis movement plans= achievea said process0 is,...,aa0 giventernational,...,a to},ternational achieveternationalisbe13 a andgiven in,ternationalgoal13 employingternationalternationalshown, as Figureternational Journal actionshown, short, international (ternational a Journal? Journal Figure major).ternational weternational Journal4.ternational A of Journalternational Journal we executedobservablefirstGiven JournalComputertactics teamof 4.of goalfirst{ Journal JournalComputer ComputerGiven of of ofJournal Computer Journal ofper- Computer(e.g. Journal. Computer Journal Science Computer ofby of Computer ScienceComputer seg-Science of a} of Scienceof Computer of in Science Computer ComputerScience Computer ScienceSport in in Science Science Sport in in8(1). Sport Science in in Sport Science8(1). Science Sport in78(1).6 Sport8(1). in8(1).6 in inSport in Sport8(1). Sport Sport8(1).8(1).8(1).8(1). recently.recognizingder to The isolate significance plans team of plans, individual of considering rather agents. thana group sequential Outside actions of process the in or- sport of a recipeformingwinningments used aof a group match), coordinatedby a of teamcan these to be movement plans achieve said to achievebe a and goal employing action ( a? major). A executedtactics team goal per- (e.g.. by a break theAPartial description fieldentry.thetasks up possessioninto Team ofwe a thisgrid mentioned.process of string Tracing4{ break{a5 isand=given WetheCommitment. then froma field0} first,...,a in vectorize} upFigure describe Ballinto13Artificialshown, a4. it grid ActionGiven our ofIntelligence we4 dataset. first5 Dataand42. then vectorize7 it 6 winningWithder torecognizingPartial the isolate aobvioustasks match), team Team planswe applications can plans, mentioned. of beRelated Tracing individual rather said to to sport than We beWork agents. from firstemploying anda sequential military describe Outside Ballthetactics domains,Actionprocess of plans our the into. dataset. sport of different DatawhichWeteam timewinningformingwinning referbreak hasseries (e.g.break to createdthethe and apassteam a the fieldmatch),groupmatch), possession the field from fieldup behaviors a intoupdemand of into cana agent into thesestring a= differ- grid be a, grid A saidaforof plans×, to=,...,a4of as real-time toagent to4a short,5 to be0 beand,...,aB5 achieve employing employingand B) then observable13 .statistics then A vectorizeshown,plan a vectorize majortacticstacticscan it andwe seg- itgoal befirst. vi-. de- (e.g. Withderrecognizing torecognizing the isolate obvious plans team plans applications plans,of of individualRelated individual rather to sport than agents. Work agents. anda sequential Outside military Outside of domains,process of the the sport sport of whichWeteamwinningformingder referrecognizingrealm, has (e.g.torecognizingWith toisolate created apassteam a most match), groupthe team fromplans obvious behaviors ofa plans demand of thisplans, can agentof these ofapplications individualwork be individual rather, A saidfor plans has, to as real-timethan agentto focussedagents. short, to agents.tobeB a achievesport sequentialemploying B) observable Outside .statistics Outside and on A plana dynamicmilitary major process of oftacticscan theand theseg- domains, goal teamsbe sport sportof vi-. de- (e.g.formingthetowinning givementsteam possessionIdeally, quantized aA (e.g. group adescriptionofAbreak match), we description coordinated pass ball string of would the positions. from these field can of this oflike upbe agent plansthis At movementinto processsaidB to each0 process A givea recognizeto× togridCohen,{ to time achieveis quantized×beCohen,{ agentCohen, of given13 step,is employingCohen,4 P.,andCohen, givenshown, andCohen,P., theB)P.,in team5 a ballCohen,} andactionCohen,and P.,majorLevesque,Figure quan- .in P.,} ACohen, andpositions.Levesque,Cohen, P.,FigureLevesque, thenandCohen, andtactics weplanCohen,tactics P.,and P.,Levesque,4.executedLevesque, goalLevesque, vectorizeand firstand P., H.Levesque, GivenP.,can 4. P., P.,andLevesque,At 1990.H.Levesque,for and.H.Given (e.g. and and× beeach 1990.Levesque, H.1990.Levesque, soc-H. it Levesque,H. IntentionLevesque,de- 1990.byH. time1990. 1990.Intention H. 1990. a step,Intention1990. H.Intention H.is Intention IntentionH. 1990.Choice the1990. is Intention1990. Choice quan-is is Intention Choice10 withIntention isChoice Intention7 Choice with is7 Choice with with is is with withChoiceis is Choice Choice withChoice with with with with derrealm, torecognizingWith isolate most the team obvious of plansRelated thisplans, of applications work individual rather Work has than focussed agents.to a sport sequential Outside and on dynamicmilitary process of the domains, teams sportof formingwinningresearchmentsteamIdeally,realm, a (e.g. group aof interest match), we most coordinated pass of wouldRelated ofinto from these can this MAPR like beagent plans work movement saidB Work to hasA has recognizeto to to achieve grownbe focussed agent employing and quite B) team aaction on major . A substantially dynamic tacticsTableplanenttacticsTableTableexecutedgoal quantizationTable 2:Tablecan 2:Table Table 2:for teams.Table Table (e.g.2:Table Tablebe 2:Table Table 2:showingTable soc- Table 2: de-areas.showingTablements byTable 2: showing Table 2:sualizations. showing2: Table 2: showing aTablefined showingTheTable the2: Table showing theof Tableshowingthe thetoidentificationresults showingIdeally, showingtheidentificationcoordinated asgive showingthe showingtoidentification possession the showing giveidentificationan theidentificationbreak are quantized identificationthe Due the orderedidentification thequantized wethe identificationthe Athe therate identification tracethe identificationdescription towouldrate identification stringidentification ratefield ballby movementidentification the ofrate bysequencebreaking ball rateby positions. up the breakingratea likedifficulty bybreaking positions.into rateby team= rate breakingbyof breaking{ ratebreaking torate aby rate thisbreakingbyagridrate At and recognize0 ofbreakingby,...,a breakingbyby each Atprocessby ofassociatedbreakingteamHervieu, breaking actionbreaking each4 breaking time13} time5 behaviors is step, teamandshown, A., executedgiven step, andthen the with tacticsin theBouthemy, quan- we vectorize Figuredescribingtracking quan- firstby for a 4. P.it soc- Given 2010. Understanding10 sports 7 ments ofrecognizingrealm,With coordinated the most obvious plans of movement this of applications individualwork has and focussed agents.to action sport Outside and executedonmilitary dynamic of the by domains,4 teams sport a 8 breaktizedwinningteamIdeally, ballt thethe1012 position,tthe (e.g. field possession11 aPartial we possessionmatch),,tup pass12 is would,tinto used13 from string Team16 can,ta to string gridlike14 populate agent be,ta of15 to Tracing={ saidatized,t4 recognize= Aa2016a.Commitment. ballto0 Using to5,...,aCommitment.aCommitment. beand0 agent position,...,a×Commitment. fromCommitment.T employing}× then13Commitment.= team B)Commitment.13 10Commitment. isArtificialshown, vectorize, Ballused.Commitment.Artificial weCommitment.Artificialshown, AtacticsCommitment.Commitment.Artificialplan toArtificialArtificialtactics Intelligencewe ActionArtificial populateIntelligence weitIntelligenceArtificial forfirstcanArtificial Intelligence firstArtificial. Intelligence soc-ArtificialIntelligence beArtificial IntelligenceArtificiala42. Data.Intelligence Usingde-Intelligence42. Intelligence Intelligence42. Intelligence42.T42.42.=42. 10,15 we42.42.1042.10 researchrecognizingrealm, interest most plans ofinto of this MAPRindividual work has has agents. grown focussed Outside quite on substantially dynamic of the sport teams sualizations.finedwinningrecognizing(i.e.Ideally,t asrealm,10research,t anwhere11 aPartial Due ordered mostmatch),we,t plans12 individual interest towould,t of of13the Teamsequencethis can,t individual likedifficulty into14 work agentsbe,t toMAPR15 Tracing said,t hasrecognize of agents. can16to associatedteam focussed has leavebe grownfrom Outsideemploying behaviors team and on withquite joinBall dynamic tactics oftherecognition.the the teamsdescribingsubstantiallytracking plansthetacticstheAction plansthe sport plans forthe plans intoteamsthe plans overintothe plans soc- into.the differentplansthe intothethe plans different into plans intodifferentDatathe plans plans plansdifferent into different into planswinning different time into into different timeinto intocer. different timebreak intodifferentfinedtized seriesIdeally, timedifferent differenttime differentseries However, timetized series different timea ball seriestheandto seriesas time match), series and time ball givethe andtoposition timefield seriesthe wean time time time series giveand the and timeposition series fieldthe quantizedpossessionand ordered serieswould asup series fieldandseries series the thequantized field can theandseries isintosoccer into and the field used intoand field is andinto the and andbe differ- likethe fielda ballused andsequence intothe differ-intogrid fieldto thestringsaid intodiffer- the thefieldball is positions. thepopulate intoto field to differ- differ-fieldlow-scoring, fieldof differ-into positions. intopopulate fieldrecognizeto differ-{a4 into× into differ-be into differ-= ofvideo intoa At5 differ- employing. differ-team differ- Usinganda eachAtdiffer-.0 using Using,...,a each team thencontinuous time} behaviorsT players time= vectorizeT13 step, tactics 10=tactics step,, 10shown, we the trajectories., wethedescribing quan-forand it. quan- we soc- com- first In Zhang, J.; Shao,15 L.; 10 recognizingrealm,(i.e.Withrealm,research where the most most plans obvious individual interestof of thisof this individualapplications work into work agents MAPR has has agents. can focussed to focussed has sportleave grown Outside and and on onmilitaryquite dynamic join dynamic of the teams substantially sportdomains, teams teams over Withwinningcer.recently.teamfinedIdeally, theIdeally, However,(i.e. obvious a (e.g. as match),The where wean pass significanceProblemapplications ordered would as individual can soccer from be like like sequence agent said FormulationFigureis of toagents to to low-scoring, consideringsport recognizeto recognize A 3:be can toof and Figure employing agent team leave military group team team continuous3. behaviorsandB) .domains, actions tacticsjoin Atacticstacticsplan teamsdescribing in forand. forcan or- over soc- com- soc- beteam de-atothen (e.g. recipecer. give chunk However,passbreak quantized useda into from the byN fieldas ball agent= a soccer (14up team positions. into A10+1) to is a low-scoring,agentgrid achieve At/10 of eachHervieu, ={ B)4Hervieu,Hervieu, 5 a .play-segmentstimeHervieu,5 A goalHervieu, A.,and continuousplanHervieu, step, A., andA.,Hervieu,Hervieu, then(} andA., andHervieu, Bouthemy,can). A., theHervieu,Hervieu, A.,andHervieu,vectorize Bouthemy,A Bouthemy, and and quan-beA., A.,and team Bouthemy, and Bouthemy, Bouthemy,de-and A.,and A.,Bouthemy, P.A., A., com-andBouthemy, itBouthemy,per-2010. and P. P. and and 2010.Bouthemy, P.Bouthemy, P.Bouthemy, Bouthemy,Understanding P.2010. P. 2010. 2010.Understanding 2010. P. Understanding2010. Understanding P. P.Understanding Understanding P.2010. 2010. Understanding2010. sports Understanding sports Understanding Understanding sports sports sports sports sports sports sports sports sports Problem FormulationFigure 3: Figure 3. realm,(i.e.research where most interest individual of this into work agents MAPR has can focussed has leave grown and on quite joindynamicentent teams substantially quantizationentent quantizationent quantizationent quantizationteamsent overquantizationent quantizationent quantizationententent quantization areas. quantizationent quantizationareas. quantization quantization areas.players quantizationareas. The areas. areas.Thecer. The areas. resultsfinedthen Theareas.Ideally, results The areas.However,and resultsThe areas. areas. areas. chunkThe resultsThere resultsarebreak areas.Thetizedt resultsas the10 areThe results theThearetized The,t Thean weresultsa theareball, ballresults arePartialThe11 tracetheintohas areresults orderedasaresults resultstraceball,t wouldthe arethe traceposition fieldresults theN12 aresoccerof databeen are thetrace position,t of=the traceareN the ofare areup the13 trace(14 the= areteam theTeamlikecontaining the oftrace issequence,t thean intotheof trace(14is teamthe usedthe14 teamoftrace is the10+1) tracelow-scoring, traceexplosion tothen,tof usedthea team traceofteam team10+1)15 gridto the ofrecognize theTracing team of chunkof,t populate× to/ the teamofZhang,10 theteam16of the populatethis the/ team{ =104 teama team team5ininto information=L.;4aplay-segments5 continuous. 5the teamfrom UsingandaN behaviors5.? Using= Jones,interest then (14}T tactics Ball=T vectorize G., 1010+1)is= and eds.,, describingin 10still weforAction, com-live-sport weIntelligent/10 itsoc- = 5 Dataplay-segments Video Event2015 Anal-15 Withrecently.realm, the(i.e. obvious mostThe where significanceapplications of individual this work of toagents has consideringsport focussed can and leave military group on and dynamic domains, actions join teams in teams or- overteamplayersa (e.g. reciperealm,cer.Ideally, and(i.e. However,passrecently.There mostusedthe where from we ball,Partial ofby has Theas individual wouldagent this a soccerdata beenteam significance work A Teamlikecontaining an to is agents has low-scoring,agent achieveexplosion to focussed of recognize Tracing can considering B) this leave a . A goal in informationon continuousplan andthe team from dynamic ( group? joincan). interest tactics A Ball teamsbeactions team andteams is de- in still forAction overcom- per-in live-sport or- soc- DataIdeally,cer.toa giverecipe However,then quantized we chunkbreak used would as theinto by soccerball field likea positions.− team up is top low-scoring, into recognize to,..., At achieve avideo grid eachpvideovideo× ofusingvideo time×video team a using continuoususingvideo goal step, playersvideo usingplay-segmentsvideo using tacticsplayers playersandvideo usingvideo ( thevideo? usingplayersvideo using). trajectories.players thenplayers quan-− playersA using trajectories. usingandtrajectories. for using players usingplayers team vectorize trajectories. trajectories. soc-com- trajectories.players playerstrajectories. players players per- trajectories.In trajectories. Zhang,it In Intrajectories. trajectories. trajectories. trajectories.Zhang, In In InZhang, In J.; Zhang, Zhang, Zhang,In Shao, J.; Zhang,20 J.;Shao,In J.;In InL.; Shao,Zhang,J.; Zhang,Shao, Zhang, L.;Shao, J.; L.;15 Shao, L.; J.; J.; L.; L.;J.; J.; Shao, Shao, L.; L.; Shao, Shao, L.; L.; L.; L.; realm,(i.e.research where most interest of individual this work into agents MAPR has focussed can has leave grown on and dynamic quite join teams substantially teams over researchcer.Ideally,derfined However, to interest isolatet as we10,t an would11 intoteam orderedas,t12 soccer MAPR plans,,t like13 sequence,t isto rather14 has low-scoring, recognize,t grown15Figure than,t16 of a quite team3: sequential team Figure continuous substantially behaviors tactics 3.recognition. processrecognition.recognition.recognition. forrecognition. andrecognition. describingrecognition. soc-ofrecognition. com-recognition.recognition.recognition.recognition.finedrecognition.formingtizedresulting as an ballto ordered in give ato position five group givethen quantized play-segments quantized chunk sequence of isa used thesea ballintoN ball to positions.− plans=N of populate positions.0 (14resulting−= team (14 to10+1) Atysis achieve4 behaviorsa10+1) in each. At andshown. Using five/ each10 Understanding time/× play-segments10 = a timeUsingT major5 = step, describing= 5 step,play-segments 10 the goal,. we the quan- Springerp quan- (e.g.0,..., Berlinp4 shown. / Heidelberg. Using2520 (i.e.recently. where The individual significance agents of can considering leave and group join teamsactions over in or- Incer. this(i.e.arecently. recipeHowever, paper, where used we Theindividual as are bysignificance soccer looking a team agents is tolow-scoring, to ofautomatically can achieve considering leave a continuousand goal answering group join (?). teamsactions A and the team over com- in or- per-scarce.cer.tizedaresulting Mostwhichrecipe However,resulting ballthen ofintoThere positiont hasused chunk five give inthe,t asfive created play-segments quantizedbydata has soccer is,t play-segmentsinto useda been collected team,t a{ tois balldemand,t an populate low-scoring, top positions. explosion0,t achieve,..., isp0,..., via} fora,tp.4 Using Atan real-timep ashown.4 in eachcontinuous goalarmyshown. theT play-segments time= Using (interest of?10 statisticsUsing). step,, human weA and team the incom- quan- live-sport and per- vi- 20 20 researchInder this to paper, interest isolate we intoteam are MAPR looking plans, rather has to automatically grownFigure than a quite 3: sequential Figure substantially answering 3. process the of finedscarce.forming as(i.e. an Most where orderedder a group toof individual isolate the sequence of data theseteam agents collected plans, plans of teamcan rather to isleave achieve behaviors via than and an a sequentialaarmyjoin major describing teams of goal human process over (e.g. of cer.thenthis However,forming process, chunktizeda weball ainto10 as can group position soccerN get11= play-segments of12 (14 is theseis used low-scoring,1310+1)this toplans{14 populate− process,{Zhang, from/−15Zhang,10 toZhang, all achieve =Zhang, L.;16Zhang, continuous wea} the 5. L.;Zhang, and L.; Usingplay-segments can} posses-Zhang, andZhang, L.;and Jones, L.; getaZhang, andL.;Zhang,Jones,Jones, majorTZhang,and andZhang, play-segmentsL.; andL.;G.,= Jones, Jones,and Jones, and L.;andG.,10 eds.,G., Jones,L.; L.; goal L.;,{ andJones, eds.,G.,com-Jones,eds., weand G.,Intelligent and G.,and G.,eds., Jones, eds.,(e.g.Jones,IntelligentIntelligent eds.,Jones, G.,Jones, G.,eds., fromIntelligentIntelligent eds., G.,Intelligent G.,}Intelligent Video G., all eds.,Intelligent eds.,Intelligent Video theeds., Event VideoIntelligent Video posses-Intelligent EventIntelligent Video Video25 Anal- Event Video Event Anal- Event Video Anal- Video Event Anal- Video Video Anal- Anal- Event Event Anal- Anal- Event Event Anal- Anal- Anal- Anal- (i.e.(i.e.recently. where where individual Theindividual significance agents agents can of can considering leave leave and and join group join teams teamsactions over over in or- recently.cer.followingcer.recognizing However,awhich recipe However, The tactical significanceTheret hasused10 as plans,t socceras created questions11 byhas soccer,t of a12 individual been of isteam,t aconsidering low-scoring,13is for demand,t an low-scoring, to soccer:14 explosion,t achieveagents.Figure15 for,t group16 continuous Outside 3:real-time a in Figurecontinuousactions goal the ofcan (interest 3.? in the statistics).and look or- A sport and at com- team inthe com- short-termlive-sport anda per- recipeannotatorswinning vi- behaviorthenthis usedthis process, chunk awhotizedresulting byofprocess, match),tizedresulting aa team balllabel weinto team ball in we can position fromN can five inall can position get to= five a play-segmentsactions be get(14partic- achieveplay-segments isplay-segments said play-segments used is10+1) used thatto to a beHess, populategoal to/ occur10p from employing populate0,...,p=R., from (0? all,..., 5). andaroundaplay-segmentsp the. all A4 Using Fern,ap the. posses-team4shown. Usingtactics posses-shown. A.theT 2009.per-= UsingT ball. 10= Using, Discriminatively10- we, we Trained3025 Par-25 der to isolate team plans, rather than a sequential process ofa recipeforming usedder to by isolate a a group team team of to plans,these achieve plans rather a goal to than achieve ( a). sequential A a team major process per- goal (e.g. of sionformingwinning stringssualizations.thenwhich to chunk represent a match),groupTherea hasinto Due acreated of team’sNhas can these= to− been be(14 behavior.the asion said plans demand difficultyan10+1) stringsysis toysis explosion{ Weysis toand be{/ysis anddoto achievefor10 andysis Understandingemploying representysis this and= Understandingassociated Understandingreal-time andysisysis 5 andas Understanding}play-segmentsin Understandingysis Understanding andweysis and aysis Understanding}ysis thea major and Understandingteam’s andUnderstanding.tactics and andSpringer statistics interest Understanding with. Understanding. SpringerUnderstanding Springer Understanding goalbehavior.. Springer.. Springer Berlintracking. Springer Springer (e.g.inand. Berlin. Berlin Springer Springer live-sportWe / Berlin. Heidelberg. Berlin. Springervi- BerlinSpringer.do. / /Berlin Springer Springer Heidelberg. Heidelberg. this /Berlin Berlin Heidelberg./ Heidelberg. / / as Berlin Heidelberg.30 Heidelberg. Berlin we/ Berlin Heidelberg. / / Heidelberg. Heidelberg./ / Heidelberg. Heidelberg.25 recently.followingrecognizing The tactical significance plans questions of individual of considering for soccer: agents.Figure group Outside 3: Figureactions of 3. in the or- sport annotatorsderwinning to isolatesualizations.recognizing awho match), team label plans, canplans Due allProblem ratheractions be toof said individual the than thatto difficulty Formulation a be sequential occur employingagents.? associatedaround Outside processulartactics the ofregion. of withball the. By sport - tracking notforming chunkingresultingsion a into groupsion strings inthenplaythis strings fivethis ofprocess, segments,chunk to process, theserepresent play-segments to representa weinto plansthisa we can aN localteam’s can aget−=N to team’s get(14 play-segments= achieve behavior.s play-segments(140,..., behavior.ticle10+1)10+1) Filterss a We4 major/ We10 fromdoshown. for/ fromthis=10 do Complex all 5goal this= asplay-segments allthe 5Using we as the(e.g. posses- we Multi-Object posses- Tracking. In CVPR30 30. derrecognizing to isolate team plans plans, of individual rather than agents. a sequential Outside of process the sport of realm,formingwinningrecognizing mostwhich a match),group ofThere has plans this created of haswork can of these individual been be has a said plans demand focussedan to explosion agents. to be achievefor employing onreal-time Outside dynamic in acan the majorcancan oflooktacticscan lookstatistics interest teamscan the lookcan lookatcan goalatlook thecan looksport at. thecan atcan look short-termthecancan atlook the short-term at look(e.g.incan thelook short-termand atlookthe lookwhich short-term at theshort-termlive-sport look at short-term thecanat at atvi- the short-term behavior theIdeally, theshort-termat theresultingwinning lookbehavior short-termthey behavior theshort-termplayers short-term short-term behavior behavior at short-term behavior ofsualizations. callthe behavior we of a inbehaviorthen of teama short-termbehavior a and behaviorfiveaball ofmatch), teamwouldof behavior behavior teamof chunk a abehaviorfrom of ateam theplay-segmentsteam fromactions ofteam a from of team a oflike behaviorball, aof frompartic-Due teamof can a fromintoteam a aof apartic- a fromteam partic- team atoteam. a afrom databe toateam fromThe partic- partic-can partic-recognize aof from−{ saidfromthe partic-from a a alookfromcontainings− F24partic-Hess, team partic-0 a difficulty,..., aHess,to aspartic-Hess, partic-at a partic-,..., R.,soccer frombepartic-Hess, theHess,s R.,team and4R.,} employingHess, short-terms a and R.,shown.Hess,andFern,Hess, partic-this R., associated data andtacticsR.,Fern,Hess, Fern, andHess,Hess, R.,A. informationHess,R.,and Fern, feedUsingFern, behaviorFern,A. 2009.andR., A.andplay-segments Fern,R.,tactics R., forR.,2009. A.2009.and Fern, Fern,A.and col-A. withand Discriminativelyand 2009.A. soc- Fern,2009. of2009.Fern, DiscriminativelyA. DiscriminativelyA.2009.Fern,. Fern, a tracking isDiscriminatively2009. teamA. Discriminatively A.Discriminatively stillA. A.Discriminatively 2009. 2009. from 2009.Discriminatively Trained Discriminatively aDiscriminativelyDiscriminativelyTrained partic- Trained Par-Trained Trained Par- Trained Par- Par-30 TrainedPar- Par- Trained Trained Trained Par- Par- Par- Par- Par- der to isolate team plans, rather than a sequential process of forming1. What a grouprealm, type of mostplaying these of plans this style work to can achieve we has expect focussed a major from on agoal dynamicbehavior team? (e.g. maybe teams lost asthis we willcan process,Ideally,resultingcan see lookresultingsion in look at thesion wewhich stringsthe at wenext in canstrings the short-term five in would subsection. short-termto hasget five play-segments represent to play-segmentscreatedrepresent play-segments like behavior behavior a to team’s{ a recognize ofteam’s demandHess, a− of0 behavior. team from as behavior.R.;0 team,...,} from Fern,all forteam4 from Wes the a4shown. real-time A.;We partic- do tacticsa posses-shown. partic-anddothis this Using Mortensen,as forweUsing as statistics we soc- E. 2007.and vi- Mixture-of- realm, most of this workProblem has focussed Formulation on dynamic teams whichrecognizingIdeally,(i.e. they where call planswe individualball would of individualactions like agents to.Problem The agents. recognize can F24 leave Outside soccer Formulation and team joinof data tactics theular teamsular sportfeedular region.ular region.ular forular region. over region. col-ular region.ular By soc- region.ular Byular region.ularular notBy region. Bynotwinningular region. Byregion. notchunking region. By not chunking Bynotregion.cer. chunkingular not By chunking Bynot chunking Bythis chunkingnotBy ByHowever,region.a intonot chunking Bynot match),into notchunking not process, into chunking play not chunking intochunking chunking playinto By intoplay chunkingresulting segments, into notplay play segments,asintocan weplay segments, into playchunking intosoccer segments,segments,into caninto play segments,be play into in thissegments, play playget saidplay segments, this five segments,play this local is segments,play-segmentsinto segments, localthissegments, thislow-scoring,to play-segmentslocal this segments, playthis belocal local local this this localemployingticle segments, this localthis{ thislocalticleticle this Filters local local from localticle Filtersticle localticle Filters continuousticles this for Filters0 all,..., Filtersticleticletactics} for Complex Filtersfor local theticleticle Filters Complex forFilters Complexticlesticle for forposses-4 Filters Complexfor. Filters Complex andFiltersComplexshown. Multi-Object Filters for forComplex Multi-Object Multi-Object Complex forComplex com- for for Multi-Object for Complex UsingMulti-Object Complex Multi-Object Complex Multi-Object ComplexTable Tracking. Multi-Object Multi-Object Tracking. Tracking. Multi-Object Multi-Object 2: Multi-ObjectTracking. Multi-Object Tracking. Table Tracking.In Tracking.CVPR Tracking.In CVPR Inshowing Tracking. In. Tracking.CVPR InTracking. Tracking.CVPR..CVPR In.CVPR. the In. In. InCVPR InCVPR identification..CVPRCVPR. . .. rate by breaking recognizingrealm, most plans of this of individual work has agents. focussed Outside on dynamic of the teams sport winningrealm,playersIdeally,Insualizations. this mostwhich a paper,and wematch), of thewould has we this ball, Due canare created worklike looking databe to to saidthe has a containingrecognize demand todifficulty focussed to automatically be employing forteam onthis associated real-time dynamic informationtacticsa answering=tacticsa60 with for,a teams statistics15,a. soc-2 the tracking,a is9lected3,a stillsion49 and,a5ular,a stringsfor9scarce.Ideally,this6 vi-ular,a region. the7this9canplayers,a process, region. to8can look,aEnglishprocess,9 Most representBy we9 look,a By notat14 and10 we would the,a atof not chunking we can1115the Premier short-term the,a achunking can team’sgetshort-term1215 ball, like,a data get play-segments13 intoular12,adatato play-segments League behaviorbehavior.into collected14 play region. recognize12 behaviorParts playcontaining segments,12{Pictorial of Bysegments, (EPL){ We froma of10 isnot team a fromused teamvia teamthis chunking Structures all this by from} this an theallthislocal from}tactics Opta information thealocalarmy posses- ap- partic- into fora posses- partic- ( Objects? play for of) humanTable soc-segments, iswith still 2:Variable Table this localPart showing Sets. the identification rate by breaking recognizing1. What type plans of playing of individual styleProblem can agents. we expect Outside Formulation from of the a team?Figure sport 2:winning (Fromrealm, the a most match),(i.e. XML of where this can feed, individualwork be said we has tocan focussedagents be infer employing can on theleave dynamictactics and joinbehavior teamsbehavior.behaviorbehavior teamsbehavior{behavior maybebehavior maybebehavior maybebehavior overmaybebehaviorbehaviorbehavior maybelost maybe lostbehavior maybe lost asIdeally, maybe lost as wemaybe lostbehavior maybeas lost maybewe maybe as will lostwe asmaybesioncer. lostwillas we welostwill seeas welost we lostwill lostas stringssee However,maybewillwe as inseewill lost weas wouldas inwewillas thesee see weinthiswillas seesualizations.we thewe will nextthe losttoin seein wewill inwillnext process,willsee therepresent the next see inwillas thesubsection.like as see in seenextthe seesubsection.we in nextsubsection.soccer thesee in thenext toin willin subsection. thewe next insubsection. the thenexta recognize subsection. theDuesee team’snext can next subsection.nextis subsection. next in low-scoring, getsubsection. subsection. tothe subsection. behavior. subsection.Hess, play-segments} thenextHess, teamHess, R.; difficultysubsection.Hess,Hess, R.;Fern,R.; WeHess, tactics continuous Fern,Hess,R.; Fern,Hess, R.;used A.; R.;Hess,Fern, fromHess, Fern, Fern,Hess, A.;R.;A.; associatedandHess, forR.; thisFern, A.;and Fern,R.;and allFern, Mortensen,A.;R.; soc-A.;ap- R.; R.;A.; and theFern,andMortensen, Mortensen,and Fern, and A.;Fern, A.;Fern, andMortensen, posses- Mortensen,com- andA.;Mortensen, and withA.; Mortensen, A.;E. A.;the andMortensen, and E.Table2007.and tracking plans Mortensen, E.Mortensen,Table2007. E.Mortensen, E. E.2007.2:Mixture-of-2007. into E.Table2007. 2:Mixture-of- Table2007. differentMixture-of- E. Mixture-of-E. E.showing Mixture-of- 2007. 2007. showing2007.Mixture-of- time Mixture-of- Mixture-of- the Mixture-of- Mixture-of- theseries identification identification and the rate field rate by into bybreakingbreaking differ- (i.e. whereIn this individualpaper, we are agents looking can leave to automatically and join teams answering over thelected2.cer. Are However, forscarce. there theplayers any EnglishIn Most as this areas soccerand ofpaper,ofthe Premier the theis ball, we low-scoring,data field areProblem data League collected lookingwhich containing they (EPL) to continuousFormulation isautomatically tend via this by to an Optautilize information armyand answeringcom- (? of) humanis is theproach astill goodbehaviorsionbehavior as exampleularscarce. it strings maybeallowsular region. maybe region. to Most lostus of represent By lost asto this. By not ofwe analyzeas not thechunkingwewill The a chunking willbehavior team’s datasee the F24 see in intoIn collectedshort-term the behavior.inICCV maybeinto data thenext play play next. subsection.segments, lostis We is segments,subsection. behavior a as viaused time we an this will this codedthis armyof local see ap- local in of thethe human next plans subsection. intoTable different 2: Table time showing series the and identification the field into rate differ- by breaking realm,(i.e. where most of individual this work agents has canfocussed leave andon dynamic join teams teams over cer.(i.e.Ideally,following However, wheresualizations. we individual tactical as would soccer questions like Due agents is low-scoring, to to canrecognize for the soccer: leave difficulty continuousand team join associateda tacticsa= teamsa=a=a and=a0a,aa0=a,a=a1 forover0a,a,aa com-01=a,a,a witha2=01a,a0,aa=1 soc-2a,a=,a31=02aa,a,a1,a032a,a=,aa,a40213 trackinga0,a,a2,a10,a403,aa,a,a,aa,a51324cer.,a10,a3,a=215,a1,a4,a,a6annotators24,a35,a1,a24,a3a26,a,a25 However,,a,a730,a54,a6,asion,a235,a473,a36,a,a8465,a17,a,asion,a63,a4players58,a4,a4 strings7,a9,a5768,a2,a,a,a7,a45,a69,a5,a strings58,a106,a87,a9,a3 who,a,a8,a56,a,a7 as,a10,a6,a6,a97810,a to,a411,a,a96 and7,a8 soccer,a,a7,a7,a represent810119label,a to,a5,a11107128,a9,a8,a,a represent9810 the,a,a11,a12116,a89,a121110,a13,a,a9,a,a109 all,a,a is11,a7 ball,912,a131012,a,a13101211 a1410low-scoring,,a,a11,a actions8,a,a1012 team’s,a,a13,a141113 a,a14131112 data,a11912,a team’s,a,a1311,a,a14,a1214,a141312101213,a behavior.14 containing12,a,a that,a13,a1413 behavior.111413,a13,a,a,a14 occurcontinuous,a14121414,a We13 We used,athis around14 used thisinformation and this ap-the com- ap- ballthe - isplans still into different time series and the field into differ- realm, most of this work has focussed on dynamic teamsball position andIdeally,(i.e. possession where we individual would at like every agents to timerecognize can stepleave team(solid and join tactics teamsEntropy for over soc-{ { Maps{{ {{ {cer.{ {{{ However,{ proacha = behavior asabehavior it,a soccer allows,a maybe,a maybe is us,a low-scoring, tolost,a analyze lost as,a we as,aa we will} the=Parts,a will}}Parts seecontinuous short-termaParts,a0 Pictorial},a see inParts},aParts Pictorial1 Pictorial} the,a inParts},a Pictorial2 the}Parts next,aPictorial Parts Structures Pictorialbehavior}3 nextParts},a Structures,a Structures}Partsand subsection.Pictorial}Parts Pictorial4Parts,a Structures subsection.,a StructuresPictorial Structures5 Pictorialcom- forPictorial,aStructures Pictorial of,a6 Structures forObjectsStructures for,a7 Objects forStructuresObjects,a Structuresfor Structuresfor Structures8 Objectsfor,a with Objectsent Objects for9Objects for,a with Variable quantization Objects10 for withthe for,a withVariable for Objectswith with11 Objectsthe plansVariable Objects,a Variable Part withVariable plans12 Part,a Sets. withinto Variable with13 Part withareas. Sets.Part into,aVariable different Part Variable14 Sets. Variable Sets. differentPart Sets.The Sets. Sets. Part Part timeresults Part Part Sets. time Sets. Sets. seriesSets. are series andthe and tracethe the field of fieldthe into into team differ- differ- following tactical questions for soccer: a team{proach overaannotators= a0 asparticulara01 it,a allows12,a2 who3,a region, us34,a label to45,a analyze56 which,a all67,a actionsIBM78 themaybe,a89 SlamTracker. short-term,a910,a that lost1011,a otherwise.occur1112 behavior,a 2012.1213 around,a}www.australianopen.com/1314 of,a14 theent ball quantization - areas. The results are the trace of the team 2. Are(i.e. therefollowing where anyIn this areas individual tactical paper, of thequestions we agents field areProblem lookingwhich can for soccer: leave they to Formulation automatically and tend join to utilize teams answering over is the a goodmorecer.annotators However,than examplescarce. others? Mostwho asof soccer this. label of the The all is data low-scoring, actions F24 collected data that is occurcontinuousis a via time anaround coded army and the of com- humanballfeed -T thata= teamwhich 10 listsproach over{ aproach all they= a particular playerasa callit,a as allows it,aball allowsaction region,,a us,a actions to us events,a which analyze toIn,a analyze{.ICCVIn,aIn The maybe withinICCVICCV theIn,aIn.InICCV theF24InICCV short-term,a.. lostInICCVIn short-termthe,a.ICCV.ICCV soccer.In otherwise.In.In,a gameICCVInICCV.ICCV.ICCV behavior,a. data. behavior. with.,a} feed of,arecognition. of col-entent quantization quantization areas. areas.} The The results results are are the the trace trace of the of the team team (i.e. where individual agents can leave and join teamslines over and dotscer. However, are annotated,1. What asplayers soccerIn type this dotted ofis and paper, low-scoring,playing the we lines ball, style are lookingdata are can continuouscontainingwe in- to expect automatically andEntropy fromEntropy thisEntropy com-EntropyEntropy aEntropy information MapsEntropy team?Entropy Maps answering MapsEntropyEntropyEntropy MapsH MapsEntropy Maps(X Maps)= Maps Maps Maps Maps the is Maps stillTpa(x team=T) log10= overscarce.{p 10(ax)= a0 particulara0 Most1,a12,a23 of(1),a region,34 the,aT45,a= data whichen_AU/ibmrealtime/index.html56 10,a67 collected,a maybe78,a89,a lost910,a is10 otherwise.11 via,a1112,a an1213 army,a13}recognition.14,a of14 human. ent quantization areas. The results are the trace of the team more than others?following tactical questions for soccer: annotators who label all actions that occur around thea ball player,ps0 - =s team,6a,which5 team,...,Ta team= event over{ 1015 they{ over a type, particularcall a particularball minute region, actions region,IBM andIBMIBM which SlamTracker..IBMsecond The SlamTracker.whichIBMIBM SlamTracker.IBMmaybe SlamTracker.IBM SlamTracker.F24IBM maybeSlamTracker.IBM forIBM SlamTracker. 2012.IBM SlamTracker. soccer lostIBM 2012. SlamTracker. 2012.each SlamTracker.lost SlamTracker.otherwise.www.australianopen.com/ SlamTracker. 2012. 2012. 2012.www.australianopen.com/otherwise. datawww.australianopen.com/ 2012. ac-www.australianopen.com/ 2012. 2012.www.australianopen.com/www.australianopen.com/ feedwww.australianopen.com/ 2012. 2012.}www.australianopen.com/ 2012. 2012.recognition. col-}www.australianopen.com/www.australianopen.com/www.australianopen.com/ feed thatwhich lists all they1.following playerWhat call typeball action tactical of actions playing events questions. style The within can for F24 wesoccer: the soccer expect game datafrom with afeed team? col-x Xlectedp00{s= 6annotators for, 5T,...,= the} 1015 English who label Premier all actions League that (EPL) occur by around Opta the (?)recognition. ballrecognition. - In this paper, we are looking to automaticallyferred). answering3. theIn scoringscarce. situations Most where of are the their data strengths collected and is weak-via an armyHH(XH(X)=H( ofXH)=(HX)=(X humanH(ps)=X∈H1)=()=XHp=(H(X)=H(xpHX(())=p5(XxHp( log()=X,X0)x0p)=9(p log{))=(X)=,...,s=(px logx()=)xp)p log() log(6ppxx log,(12(p)x)5x(p logp),...,x)(p(p)x logx(( logx)px)()} log)px15 log logp()x( logpx)p(p)x((px)x())x) Intille, S., and Bobick, A. 1999. A Framework for Recog- 1. What type of playing style can we expect from a team? ￿1 sp =lectedp500,{9s=,...,0 for6, 125 the,...,} English15(1)(1)(1) (1)(1)(1) Premieren_AU/ibmrealtime/index.html(1)en_AU/ibmrealtime/index.html(1)en_AU/ibmrealtime/index.html(1)en_AU/ibmrealtime/index.html(1)en_AU/ibmrealtime/index.html(1)(1)(1)en_AU/ibmrealtime/index.html Leagueen_AU/ibmrealtime/index.htmlen_AU/ibmrealtime/index.htmlen_AU/ibmrealtime/index.htmlen_AU/ibmrealtime/index.htmlen_AU/ibmrealtime/index.htmlen_AU/ibmrealtime/index.html (EPL) by Opta. (?.)...... a player,2. team, Arewhich there event they any type, call areasball minute of actions the and field.second The which F24 they for soccer tendeach todata ac- utilize feedtion. col- Eachis11{ asp event1 good= p5 has0, 9= example,...,} a6, series512,..., of15 ofEntropy this. qualifiersnizing The Maps Multi-Agent F24 describing data Action is it. a Ev- fromtime Visual coded Evidence. In AAAI. nesses?lected for the English Premier League (EPL) by Opta (?x)psxX2xX=xXxxX9X,xX19{,...,xsX1xwhichXxXxxX12xX{X X{ they} call} ball} actions. The F24 soccer data feed col- 3. In scoring1. situationsfollowing What type where tactical of playing are questions their style strengths can for wesoccer: and expect weak- from a team? annotators2. Are there who any areas label of all the actions field which thatAnalysis occurthey tend Using around to Entropy utilize the￿∈￿ ball∈ Maps￿∈sp￿2∈2￿{∈￿= -∈is￿∈p￿9 a∈￿1,∈{9￿ps good=∈￿,...,1￿∈∈￿=∈5}, 1295 example,...,, 9,...,}12 12 of this.Intille,Intille,Intille, TheIntille, S.,Intille,Intille, S., andIntille, S., F24Intille, andIntille, S.,and Bobick, S., S.,Intille, and S.,Bobick,Intille, Bobick, anddataIntille, andIntille, S., andS., Bobick, Bobick,A. Bobick, and S.,and Bobick,is S., A.S., 1999. A. S., and Bobick, Bobick,a and and1999.A. 1999.and timeA. Bobick, A. Bobick,A A.1999. Bobick, Bobick,1999. Framework 1999. A. AA. A1999. coded Framework Framework 1999.A 1999. A. A A. Framework A.A A. A Framework 1999. 1999. Framework Framework 1999.A for Framework Recog- forA A FrameworkA for Recog-Framework for Framework forRecog- Recog- for Recog- Recog- Recog- Recog- for for for for Recog- Recog- Recog- Recog- more1. than What others? type of playing style can we expect from a team?ery eventp = collected9sp,229,...,= 91,12{9 by,..., Opta12 for} a given match is listed within tion. Each eventlected has for a the series English of qualifiers Premier describing League (EPL) it. Ev- by Opta (?s3) feed that{sp2lected= lists9,{9 for,..., all} the player12 English} actionnizingKim,nizingnizing Premier eventsMulti-Agentnizing K.;nizing Multi-Agent Multi-Agentnizing Grundmann,nizing Multi-Agentnizing Multi-Agent withinLeaguenizing Multi-Agentnizingnizing Actionnizing Multi-Agent Multi-Agent Action Action Multi-Agent Multi-AgentM.; the Multi-Agent ActionMulti-Agentfrom (EPL) Action Action Shamir, Actionfrom gamefrom Visual Action Action from from Visual Visualby from Actionfrom A.;Action with Evidence.Visual Action Action fromVisualOpta from VisualMatthews, Evidence.Visual from fromVisualEvidence. fromEvidence. ( Evidence. Evidence.In Visual) VisualAAAI VisualEvidence. I.;In AAAI Hod-In Evidence.. In Evidence.AAAI Evidence.In AAAI.. AAAIAAAI In. .AAAI. In. In InAAAI In..AAAIAAAIAAAI. . .. 2. Are there any areas of the field which they tend to utilize is a goodmore example than others? of this. The F24 data is a time coded sp33{ = 92,{9ps,...,2 =} 129, 9,...,} 12 H(X)= p(x) log p(x) H?(X,(1) Y )= p(x, y) log p(x, y) (4) nesses? We definewhich a team they as a setcall ofball agents actions having. The a shared F24AnalysisDeterminingAnalysisAnalysisAnalysis soccerAnalysis ob-Analysis UsingAnalysisAnalysis UsingAnalysis UsingAnalysisAnalysisUsingParameters data UsingEntropyAnalysis Using Entropy Using Entropy Using feedEntropy UsingEntropy Using EntropyUsing p Maps4 Entropy UsingEntropyMaps= col- MapsEntropy Entropy EntropyMaps9 Entropysp Maps,feed3 Maps3Entropy9{,...,= Maps Maps Mapsthat92 Maps,10 Maps{9 Maps,..., Maps lists}12 all player} actiongins, J.; events and Essa, within I. 2010. the Motion game Fields withH to( PredictX, Y )= Play p(x, y) log p(x, y) (4) 2.1. Are What there type any of areas playing of the style field can which we expect they tend from to a utilize team? is a good example of this. The F24 data is a timethis coded feed.ap player, The= sp93 type3, 9=,..., team,9 of,{109,..., events event12 listed type,} Kim, minuteare:Kim,Kim, K.;Kim, goals,Kim, K.; K.;Grundmann,Kim, and Grundmann,K.;Kim, Grundmann,Kim, K.;K.; shots,Grundmann,Kim, secondKim,Grundmann, Kim, K.; K.;Kim, Grundmann, M.;K.;Grundmann, Grundmann, K.; K.;passes K.;M.; M.; Grundmann,Shamir, Grundmann,for Grundmann, Grundmann,M.; Shamir, Shamir,M.; M.; eachM.; Shamir, Shamir,A.; M.;Shamir, Shamir, A.; Matthews, ac-M.; Shamir, M.; A.; Matthews,M.; A.; Shamir, A.; A.;Shamir, Matthews, Shamir,Matthews,H A.;Matthews, I.;(X, Matthews,Hod- A.; I.; A.; A.;Y Hod-I.; Matthews,)= I.;Matthews, Matthews,Hod- I.; Hod-− Hod- Hod-I.; Hod- I.; I.; I.; I.;Hod- Hod- Hod- Hod-p(x, y) log p(x, y) (4) more than others? tion. Each eventery has event afeed series collected that of2. qualifiers lists Are by allOpta there player describing forany a areas action given it. of match This theevents field is within which listed theywithin the game tend to with utilize 4{ is{ps3 a3 = good} 9, 9,..., example} 12 of this. The F24x dataX is a time coded H(X,− Y )=Hx(YX)=y Y p(px,(y y)) log logpp((yx,) y) (4) (3) 3. In scoringlected forsituations the English where are Premier their strengths LeagueDeterminingHowDeterminingDeterminingDetermining discriminativeDetermining and (EPL)DeterminingDeterminingDetermining weak-DeterminingDetermining ParametersDetermining ParametersDetermining by Parameters Parameters areParameters OptaParameters Parameters our Parameters Parameters Parameters representations? Parameters Parameters(? Parametersp)a4{ player,=p =9,{99,...,, team,9 That,...,}10 is,10 event} do type,Evolution minute in Dynamic and second￿∈ Sports for Scenes. each In ac-CVPR. x−X∈y Y∈ more than others? jective andfeed a shared3. that In scoring lists mental all situations state player (?). action whereAs we events are are their dealing within strengths with the and game weak-(with withEntropy start/endtion. Maps Each4 point),{ event{ tackles, has} a seriesclearances,} gins,gins,gins, of J.;gins, qualifiersgins, J.;and J.;gins, cards, and andJ.; Essa,gins, J.;gins, and Essa,J.; Essa,andgins, andgins, J.;I.gins, and J.; freegins,Essa, describingEssa, Essa,2010.and I. andI.J.; Essa, J.; J.; 2010.J.;2010. kicks,andI. Essa, Essa,and I. I.and and2010.Motion I. 2010.Essa, 2010. Essa, Motion2010.I. Essa,MotionI. Essa, 2010.2010.Motion it. I. Fields Motion I. Motion I. I.Motion 2010.Ev- Fields2010. 2010. 2010. Motion toFields Fields Predict MotionFields toFields Motion Motion Predict toFields to Predict toPlay Predict Fields Fields Predict toFields Play Predict Play toPlay￿ to∈ Playto toPredictPlay￿x Predict− Predict￿ PredictPlay ∈X￿y PlayY Play− Play Play We define a team as a set of agents having a sharedtype ob- of data isthis currently feed. used Thenesses? for type themore of real-time events than others? listed online are: visual- goals, shots,teams passes have unique stylesEntropy ofEntropy play?Entropy Maps Andtion. Mapsfeed ifps Each4 so,4Maps{= that can9 event, we9 lists,..., detect has} all10 playera seriesLaviers, action ofK.; qualifiers eventsSukthankar, describing within G.; Molineaux, the it. game Ev- M.; with and Aha, D. ￿∈ x ￿X∈ y yY Y 2. Are there any areas of the field which they tend to utilizesoccer,a eachplayer, agentis ateam,nesses? good is permanently event example type, fixed of minute this. to one The and team. F24 secondHowHow UnlessHowHow data discriminativeHow discriminativeHow discriminativeforHow discriminativeHow is discriminativeHow discriminativeHow eachHow adiscriminative discriminativeHow time discriminativecorners, discriminative are discriminative ac-areThe discriminative are ourare codedour areEntropy ourinformationare representations? ourare offsides,representations? our arerepresentations?our areEntropy ourrepresentations? are representations? are ourare representations? Mapsour are representations? our our ourrepresentations? representations? substitutionsourcontent representations? representations?Maps representations?That{ representations? That That is, Thatof Thatis, dois, Thata do is,ThatdorandomAnalysis is, That}is, and doThatis, do Thatdo ThatEvolution is, Thatis,do stoppages.EvolutionEvolution is,do discretedois, is,UsingEvolution is, doEvolution do do inEvolution do DynamicEvolution inEntropyEvolution in variable, Dynamic DynamicEvolution inAnEvolution in inEvolution DynamicEvolution inDynamic Dynamic Sportsexample Dynamic in inMaps Sports SportsDynamicX Dynamic in in Scenes.Sports in Dynamicin Sports Dynamic Sports DynamicScenes. SportsScenes.Dynamic SportsScenes. Sports InScenes. Scenes. Scenes. SportsCVPR InSports Sports Scenes.CVPR In In. CVPR Scenes. In In Scenes.CVPR Scenes..CVPR In.CVPR. In. In InCVPRCVPR.CVPR. . .. ￿∈ ￿∈￿∈ 3. In scoring situations where are their strengths and weak- a player, team, event type, minute andteamswhen secondteamsteamsteams theyhaveteams haveteams haveteamsdeviate have uniqueteams for have uniqueteamsteams haveteams uniqueteams haveunique eachteams have uniquefrom styles unique have styleshave haveunique have styles uniquethis styleshaveac- of unique styles unique styles ofuniqueThe unique play? style of styles unique play?ery stylesofEntropy play? of informationstyles ofandAnd stylesplay? play?stylesevent styles styles of play?And stylesof Andquantify play? ifa of play?And of And so, ifplayer, of play?of of AndMaps ifcollected play?so, of play?And can play?so, play? if if Andcontent play?it? canif And so, can we ifso, And And Andteam, Andwe ifso,can detect we if canAnd so, if detectso, canby weofif wedetect if if weso, can ifso, can so,a wedetectOptaevent detect so, detectcanrandom we can can we detect can we detect we detectfor2009. we type, we detect detect detectdiscrete a detect Improvinggiven minute variable, match Offensive and is secondX listedX Performance withinfor each through ac- Opponent p(x, y) jective and a shared3. In scoring mental situations state (?). whereAs we are are their dealing strengthsizations with of and events, weak-(with as well start/endtion. as post-analysis Each point),3.In event scoring tackles, has for situations prominent a seriesclearances, of where tele- qualifiers cards, are their free describing strengths kicks, and it. Ev- weak- TheEntropyery information eventH(X)= Maps collected contentp(x) byof log aOptapLaviers, random(xLaviers,)Laviers, forLaviers,Laviers, K.; adiscreteLaviers, given K.; K.;Sukthankar,Laviers,Laviers, K.;Sukthankar, Sukthankar,Laviers, K.;(1)Laviers,Laviers, matchvariable,K.; Sukthankar,Laviers, Sukthankar, Sukthankar, K.; K.; Sukthankar, G.; K.;Sukthankar, Sukthankar,K.; is K.; G.;K.; G.;Molineaux, Sukthankar,listed Sukthankar, G.;Sukthankar, Molineaux,Sukthankar, Molineaux,X G.; G.; G.; Molineaux, Molineaux,within G.;Molineaux, Molineaux, M.; Molineaux,G.;I G.; M.;( G.;andX Molineaux, Molineaux,M.;; andM.;Molineaux, Aha,Y M.;)= and Aha,and M.; D. and Aha, Aha, D.andM.; Aha,M.; M.; D. Aha,D.and and D. andD. and Aha, D. Aha, Aha, Aha, D.p D.( D. D.x, y) logp(x,p( yx,) y) more than others? an agent hasWefeed been define that dismissed a team lists asall from a player set the of match, actionagents each havingeventswhen teamwhenwhenwhen awithin theywhen sharedthey al-when theywhen deviatetheywhen deviate theywhen theydeviatewhen thewhen deviate they ob- deviatetheywhen from deviate theyoffrom gamethey deviate they from deviatewith thisfrom thethey deviate fromdeviate this deviatefrom thiswith style from deviatewithdata thisa stylethis from styleprobability thisfrom anda from stylethis from stylefromThe andprobability feedthisstylefeed. andfrom quantify this styleEntropy this andquantify thisandinformationstyle this quantify and style this isstyleand Thequantify quantifystyle Hstyle distributionquantifyand it? given styleand( quantify distributionXit? and Maps it?andtypequantifyMaps and)= quantify and it?quantify quantifyit?content quantifyin quantifyit?of Figure it?p it?Determining events(p it?x( it?of) it?x) logwasit?)was aModeling. 5(a).p random listed( definedx defined) This Parameters are: In discrete by byAAAI type goals,Shannon Conference(1) variable, of shots, data on passes ArtificialI(X Intelligence; Y )=H for(X, Y )= p(x, y) logp(x, y) logp(px,(x, y(5)) y)(5) (4) nesses? vision and newspaper entitiestion. (e.g. EachWe ESPN, define event a hasThe team a Guardian). seriesas a set of of qualifiers agents having describing a shared it. ob- Ev- withthis a probabilitytion. feed. Each TheHx(X distributionX event type)= of hasp events(2009.px( a)x2009. log2009. series)was Improving2009.p listed2009.( Improvingx Improvingdefined2009.) of2009. Improving2009. Improving are: qualifiers2009. Improving Offensive2009.2009. by Improving2009. ImprovingOffensive Offensivegoals, Shannon Improving Offensive Improving Offensive(1)Improving OffensiveImproving Performance Offensive describing shots, PerformanceOffensive PerformanceOffensive Performance Offensive PerformanceOffensive PerformanceOffensive Offensivepasses Performance through Performance it. through PerformanceI Performance Ev- through( Performance OpponentXthrough through through; Y Opponent through)= Opponent Opponent through Opponent through through through− Opponent Opponent Opponent Opponent Opponent− p(x, y) log p(x)p(y) (5) soccer, each agentnesses? is permanently fixed to one team. Unless corners, offsides, substitutions and stoppages. AnFigureMutual example Information 3: Given(?) a possessionwith a probability string,∈H(xX distributionX)= we breakInteractivep(px()x log)was itp( Digitalxup defined) Entertainment by Shannon(1) . I(X; Y )=− −x X y Yx X yp(Yx, yp()x logp)p(x(y)p)(y) (5) waysery hasjective eventthe same and collectednesses? a number shared by mentalof Optaagents state for (11).a ( given). We As refer we match are to team dealing is listed with withinis currently(?) used for the￿ real-timeModeling. onlineModeling.Modeling.Modeling.Modeling.Modeling. In visualizationsModeling.AAAIModeling. In InAAAIModeling.AAAI InModeling.Modeling. InConference InModeling.AAAI InAAAIAAAI Conference ConferenceAAAI In In ConferenceAAAIAAAI ConferenceIn Conference In ConferenceIn ofAAAIIn onAAAI ConferenceAAAI ConferenceAAAI Artificial on on Conference Artificial Conferenceon Conferenceon Conference onArtificial on Artificial Intelligence Artificial Artificialon Intelligence Artificial on onIntelligence Intelligenceon Artificial Artificial Intelligence Artificial for Intelligence for Intelligencex for Intelligence for Intelligence IntelligenceXx for∈ foryX forYy∈ Y for∈ for for for∈ erya eventjective player, collected and team, a shared byevent Optamental type, for state? a minute given (?). As match and we are second is dealing listed for within with each(with ac-(?) start/endTheery information event point), collectedH￿∈( contentXxHow tackles,)=X by discriminative of Optap a clearances,(x random) for log p a(x given discreteare) ourcards, match representations?variable, free(1) is listed kicks,X within That is, do ￿∈￿−￿∈ ￿￿ ￿ p(x)p(y) 3. In scoring situations where are theirEven strengths though and this weak- data has been widely used, there are no intoMutualMutualMutualMutualMutualNMutual InformationMutual InformationequalMutual InformationMutual InformationMutualHMutual Information Information(MutualX Information)= chunksInformation Information Information Informationcan Information bep( andx((with)? log) usep(x start/end) the quantized point),(2)￿∈ x tackles,X Laviers, ball clearances, K.; po- Sukthankar, cards, G.; Molineaux, free kicks, M.; and Aha, D. ￿∈ x ￿X∈ y Y an agent hasWe been define dismissed a team as from a set the of match, agents each having team a shared al- ob-ofbehaviors the datathis, feedfeed., as is The short, given type observable in of Figure events segments 5(a). listed This of are: coordinated type goals, of shots, data passesevents,Analysis− ascan wellcan Usingbe with be as Entropy post-analysis a probability Maps distribution for￿∈ Interactivex prominentXInteractiveInteractiveInteractivepInteractiveInteractive(xInteractive Digital)wasInteractiveInteractive Digital Digital televisionInteractive definedInteractive Digital EntertainmentInteractive Digital DigitalInteractive Digital Entertainment Entertainment Digital Digital Entertainment by Entertainment Entertainment Digital and Entertainment DigitalShannon Digital Digital Entertainment Entertainment. . Entertainment. Entertainment Entertainment Entertainment...... == HH(X()+X)+￿H∈ (HY￿∈()Y )H(X,H( YX,) Y ) We define a team as a set of agents having a shared ob- soccer,tion.soccer, eachWe Each agent each define event isagent apermanently team has is permanently asa series a set fixed of of fixedagents to qualifiers one to havingteam. one team.describing Unless a shared Unless ob- it.xAnalysisXcorners, Ev- Using offsides, Entropy substitutions Mapsteams∈ haveand unique stoppages. styles of play? An And example if so, can we detect = H(X)+H(Y ) H(X, Y ) systems which use this datathisB or feed. data The like typethis for of events automatic listed are: goals, shots,HH(XH(X)=H passes(XH)=(HX)=(XH()=XTheH)=()=XH∈(HX)=H(pX( information()=(XHx(p)=XX()p)=(corners,x log()=)=X)=)xpcanp log)()=(thispx logx()xp)pbe log() log(ppxx log((p feed.)x)x(p logoffsides,p) content,x)(p(p)x logx(( logxp)x)() log)px The log logp()x( logpx(2))p()p namelyx(( type(2)substitutionspx)x(2)())x)(2)(2)(2) of￿(2)2011. (Shannons) events(2)(2)(2)(2) Improving(2)(2) and listed stoppages. entropy Offensive are: goals,of Performance An shots,example through passes Opponent − − p(x, y) ways has the samenesses? number of agents (11). We refer to team is currentlymovement used and action for the executed real-time bya online team(e.g. visualizations passsition from values of asDetermining− our−￿−−The− play-segmentAnalysis−The−Analysis information−−(? information−Parameters−)− Using− Using content, Entropy content, representation,Entropy namelywhen Maps namely MapsLaviers, theyLaviers, (Shannons)Laviers, deviateLaviers, (Shannons)Laviers, K.;Laviers, K.; K.;Sukthankar,Laviers,Laviers, s.from K.;Sukthankar, Sukthankar,Laviers,entropy K.;Laviers,Laviers, K.; Sukthankar,Laviers,this entropy Sukthankar, Sukthankar, K.; K.; Sukthankar, style G.; K.;Sukthankar,of Sukthankar, K.; K.; G.;K.; G.;Molineaux, andofSukthankar, Sukthankar, G.;Sukthankar, Molineaux,Sukthankar, Molineaux, G.; quantify G.; G.; Molineaux, Molineaux, G.;Molineaux, Molineaux, M.; Molineaux,G.; it? G.; M.; G.;and Molineaux, Molineaux,M.; andM.;Molineaux, Aha, M.; and Aha,and M.; D. and Aha, Aha, D.andM.; Aha,M.;= M.; D. Aha,D.and and D. andHD. and Aha,( D. Aha,X Aha, Aha,)+ D. D. D.H D.(Y−) H(X, Y ) jective and a shared mental state (?). As we are dealing with (withan agent start/endan has agent been point), has dismissed been tackles, dismissed from clearances, fromthe match, the match, cards, each each team free team al- kicks,newspaper al-a discretex xXxofXxXx thex entitiesX randomXxXThexX dataxXxXx informationxX (e.g.xX variableX feedX ESPN, is X given content, that The has in namelyModeling. Guardian). Figurea probabilitydistribu- (Shannons) 5(a). In International Even This entropythough type Joint of of Conference data on ArtificialI(X; Y )= p(−x, y) log (5) jective and a shared mental state (?). Astactical we are dealing analysis. with (witheryjective start/end event and collected point), a shared tackles,by mental Opta state clearances, for a (? given). As wecards,match are dealingfree isH( listedY )= kicks, with within￿∈aDetermining￿∈ discrete￿∈ap￿( discrete∈￿y∈￿of)∈ log￿∈Analysis randomthe￿(with∈p￿∈(￿y random∈￿￿)∈ dataParameters∈￿∈ start/end variable Using feed variable Entropy(3)is X point), given that X that2011. has Maps2011. intackles,2011. has a Improving Figure2011.probabilitydistribu-2011. a Improving Improving2011. probabilitydistribu-2011. Improving2011. clearances, Improving 5(a).2011. Improving Offensive2011.2011. Improving2011. ImprovingOffensive Offensive ThisImproving Offensive Improving OffensiveImproving OffensiveImproving Performance Offensive cards, type PerformanceOffensive PerformanceOffensive Performance Offensive PerformanceOffensiveof PerformanceOffensive Offensive Performancefree datathrough Performance kicks,through Performance Performance through Performance Opponentthrough through through Opponent through Opponent Opponent through Opponent through through through Opponent Opponent Opponent Opponent Opponent− p(x)p(y) behaviors, , as short, observable segments of coordinated agent A to agent B). These behaviors are observed from par- How− discriminativeDeterminingDeterminingcan beare Parameters our Parameters representations?Intelligence That Conferenceis, do . x X y Y soccer, each agent is permanently fixed to one team. Unlessevents, ascorners,ways well has aswayssoccer, offsides, post-analysis the has same each the substitutions number same agent for number is of prominent permanently agents ofand agents (11). stoppages. television fixed (11). Werefer to We one An andrefer to team.team example to thisteam UnlesstionpX datay Yis has=(p1,...,pn)isthendefinedas: currentlyais discrete been currently widely random used used for used, variable for the there the real-timeModeling. XModeling. thatreal-timeModeling. areModeling. hasModeling. noModeling. Ina onlineModeling.Modeling. systems probabilitydistribu-International In onlineInModeling.Modeling.InternationalInternational InModeling.Modeling. In InInternational visualizationsInInternationalInternational visualizationsInInternational whichIn International JointInternational In In In JointInInternational JointInternational ConferenceInternationalInternational Joint Joint Conference ConferenceJoint Joint Conferenceof Joint JointConference of Conference Conference on Joint Joint Conference Joint Artificial on Conference onArtificial Conference on Conference Artificial on Artificial Artificialon Artificial Artificial on on on on Artificial Artificial Artificial Artificial∈ ∈ soccer,We each define agent a team is permanently as a set of fixedagents to havingFor one this team. a work shared Unless we used ob- thecorners, 2010-2011this feed. offsides, EPL The season typesubstitutions ofF24 events Opta and listed stoppages. are: goals, AnHH( shots,exampleYH()=YH(YH)=(HY)=(YH()=Y passes)=H(tionpXHow)=∈YH(HY)=pH(HtionpXY()=yp(YH()=Y)( discriminativeYp)=y log( =(p1,...,pn)isthendefinedas:)=)=))=Yypp log)()=(pcorners, logy =(p1,...,pn)isthendefinedas:y())yp log() log(ppyy log((p))yy( logp))y(p(p)y logy(( log)yp)y areoffsides,() log)py log logp()y( logourpy)(3)p()py(((3)py)y(3)( representations?)Mutual)y)(3) substitutions(3)(3)(3)(3)(3) Information(3)(3)(3)(3) That and is, stoppages. do An example ￿ ￿ B teams−−￿−− have−How− −tionpX unique− discriminative−Determining−−−The− =(p1,...,pn)isthendefinedas: styles information of play? are Parameters our And content, representations?Intelligence ifLi, so,IntelligenceIntelligence R.,namely canIntelligenceIntelligenceIntelligence andIntelligence we Conference Chellappa,IntelligenceIntelligence(Shannons) detect Conference Conference ThatIntelligenceIntelligenceIntelligence ConferenceIntelligence Conference Conferenceis, Conference. R. Conference Conferencedo. 2010.. entropy Conference. Conference. Conference Conference. Group. .. of Motion. . .. SegmentationSummary and Future Work movementan agent and action has been executed dismissed by a from team the (e.g. match, pass each from team al-newspapertial spatio-temporalbehaviors entitiesbehaviors, (e.g. tracings,, as, ESPN, short,, as which short, observable The inGuardian). observable this casesegments refers segments Even of to though coordinated ball of coordinateduse thisyInteamsyY [85],yevents,In dataY yYy [85],Iny haveYHowevents,Y ShannonyY or[85],yY Shannon asy uniquedataY discriminativeyYyy Shannon wellY asyYY like usesYwell styles usesas this probability post-analysisuses as of probability arepost-analysis play?for probability our automatic And representations? theory theory if forso, theory for tocan tacticalprominentto model prominentmodel weto model detect That analysis. infor- televisionis,infor- television do and and SummarySummary and= and FutureH( FutureX)+ WorkH( WorkY ) H(X, Y ) anjective agent and has a been shared dismissed mentalfrom state (the?). match, Asfeed, we which areeach dealing team consists al- with ofof 380 theof games the(with dataan data agent feed andstart/end feedhas more is given isbeen point), thangiven dismissed in 760,000 intackles,Figure Figure from 5(a). clearances, 5(a). the This match, ThisH type cards,( typeX, each Y of)= of team free data datawhen￿ al-kicks,∈￿∈ they￿∈￿p∈￿(∈￿ deviatex,∈￿ y∈aHow￿of)∈ discrete￿ log∈￿∈ thefrom￿￿∈p∈ discriminative￿(∈x, data thisy random) style feed(4) and variable are is quantify givenourUsing X representations? that it? ina Spatio-Temporal has Figure a probabilitydistribu- 5(a). That Driving This is, do typeForce Model. of data In CVPRSummary. and Future Work− movementmovement infermovement fromB and hand-annotated actionB and action executed executed ball-action by a by team a datachunk team (e.g. (see (e.g. pass each pass from possession from−Formationmationwhen ourmation string,teams sources, they workteams sources,In have deviatesources, [85],o have i.e.,we, unique intoi.e., Shannon fromtheunique used i.e., the adata styles this thedata set thestyles usesproducedstyle data of ofproduced 2010-2011 play? of play-segmentsprobabilityandproducedLi, play? quantifyLi, AndR.,Li, by by R., AndandLi, R., aLi, if aby andsourceH R.,andLi, so,Chellappa,source theoryEPLR., ifit? a(Li, andXChellappa,R.,Li, so, canChellappa,source and)= R.,Li, andR.,Chellappa, canLi,is seasonis wetoChellappa,Li, Chellappa,Li, and R.,and treatedChellappa, R.,R. model weisR., detect R., andChellappa, 2010.Chellappa,treated andR. R.and detect andp 2010.Chellappa, 2010.F24 R. Chellappa,( infor- R.xChellappa, GroupChellappa,R. 2010.) R. 2010.log 2010.Group Group 2010.R. R.TopTo Motion Group( 2010. Groupx R.DoTo GroupR.)Do MotionGroup R. 2010. Do2010. GroupMotion 2010.Segmentation Motion Motion Motion GroupSegmentation Group MotionGroup Segmentation Segmentation(2) SegmentationMotion Motion Motion Segmentation Segmentation Segmentation Segmentation Segmentation agent Aways to agent has B).the These same number behaviors of are agents observed (11). fromWeevents. refer par- to Eachteam teamthis data plays has 38 gamesbeenways widely each, has the which used, same corresponds there number are of no agents systems (11). whichH WeH(X,H(X,H refer( YX,H()= YHX,( YX,)=H(X, Y)=H to( YX,)=H( YH)=X,teamxH()= YX,((X,X(H Y)=X, Yy)=( YX,newspaper)= Y Y)=p)=)=( Yx,pnewspaper)=(px, y()x,tionpXteamspypis log((p) yx,x,( log) currentlyx,pp logy y((p)x,)yentitiesx, havep(p log) log( =(p1,...,pn)isthendefinedas:px,( y yx,logp(x,)p)x,(p yentitiesp logx,(y(y) uniquex,p( y)x,) logx,( y) logx,p y) y y() log(e.g.used(4))x,p) y) logp log()x,( y(4) logstylesx,p(4) (e.g.)p( ypx,( y)( ESPN,px,(4) for)x,(4)((4) yx, y) ofyESPN, (4))) y the play?)(4)(4) The(4) real-time(4)(4) AndThe(4) Guardian). if Guardian). so,− online can we Even detect Evenvisualizations though though of wayssoccer, has each the same agent number is permanently of agents fixed (11). to We one refer team. to team Unlessis currentlyis currentlycorners, used offsides, used for for thesubstitutions the real-time real-time online and online stoppages. visualizations visualizations Anas example of￿∈ of a￿∈ randomwhenmationwhen they variable. they deviatesources, deviate from Thei.e., from this the information this style data styleUsing andLi,producedUsingUsing andC.; quantifyaUsing Spatio-TemporalUsing content,aManya, quantify aUsing Spatio-Temporal Spatio-Temporal byUsing aUsing a Spatio-Temporalit? aSpatio-Temporal Usinga F.;Usingsource Spatio-TemporalUsingit? aUsing namelya Mohamedou,Spatio-Temporal Spatio-Temporalx a a DrivingSpatio-TemporalX ais Spatio-Temporala Spatio-Temporal Spatio-TemporaltreatedDriving Driving Driving ForceDriving Driving N.;Driving Force Driving Model.and Force Force DrivingModel.ForceTo ForceDriving Planes, Driving Model. DoForce In Model. Model.CVPR In Force J.Force Model.CVPR ForceIn In2009.. CVPR Model.InCVPR Model.. Model.CVPR In.CVPR. In.. In InCVPR InCVPR..CVPRCVPR. . .. next subsection). S = {s0,..., s−N−−−1−as}−−ofas a−− randomequal− a−−− random− length, variable. variable. where The The informationN informationis the content, total content, namely∈ namely tial spatio-temporal tracings, which in this case referswith to ball each team playingagent each Aotheragentbehaviors to agent Ateam to B). agent, twice These, as B). short,(once These behaviors observable home behaviors are observed aresegments observed from offrom coordinated par-Opta par-xMutualxXxyX feed,xXyYxthisxyXYX InformationxYXyythisx dataywhichXYxYXywhenYxXxxyY dataXInyxX hasXYyY [85],yXy they consistsYyY hasY beenY deviate Shannon been widely offrom widely uses 380 this used, probabilityExploiting stylegames used, there and there quantify andCycle theory are are more noStructures it? to no￿ systems model systems than in infor- MAX-SAT. which which In Interna- behaviorsbehaviorsan, agent, as, has short,, as been short, observable dismissed observable segments from segments the of match, coordinated of coordinated each teamuse this al-events, dataevents, or as data well as like well as this post-analysis as post-analysis for automatic for for tacticalprominent prominent analysis. television television and(Shannons)￿ and∈￿∈￿(Shannons)￿∈Mutual∈￿￿∈￿∈(Shannons)∈￿∈￿∈as￿∈￿∈∈￿ entropyInformation∈￿ a￿events,∈∈￿￿∈ entropy∈∈ random￿￿∈∈￿∈￿ entropy￿∈∈￿ of∈ as of a variable. discretewell a of discrete a discrete as Therandompost-analysis randomLi, randomLi, information C.;Li,C.; variableLi, C.;Manya, variableLi,Li, Manya,C.; Manya, variableC.;Li,Li, C.; Manya, for F.;Manya,Li, C.;X Xcontent,C.;Li, Manya,Li, F.;Mohamedou,F.;prominent thatLi, that C.;Manya, XManya, C.; Mohamedou, C.;F.;Mohamedou, C.;that F.;Manya,F.;has Manya, Mohamedou,F.; Manya, Mohamedou,Manya,namelyMohamedou, has F.; F.;Mohamedou, Mohamedou, N.;Mohamedou,F.; television F.; F.; F.; N.;Mohamedou, and Mohamedou, Mohamedou, Mohamedou, N.; and N.;Planes, N.; N.; and Planes,and N.;and and Planes, J. Planes, andN.; Planes, N.; 2009. J. N.; andPlanes, J.and2009. J.and 2009.J.Planes, Planes,2009. Planes, J.2009. 2009. 2009. J. J. J. J. 2009. 2009. 2009. 2009. B A plantialcan spatio-temporaloftial be the spatio-temporal defined data feedB as tracings, an is orderedtracings, given which insequence which in Figure this in case thisnumber of 5(a). caserefersteam This refers of to play-segments ball type to760,000 ball of dataMutual events.Mutualuse formationH thisa Information(X Eachpossession,)= Information datap sources,(x, y team or) data i.e.,p(x playsand) thelike logExploiting dataMp thistional( 38xExploitingExploiting)is produced forExploitinggames theExploiting ConferenceExploiting automaticCycleExploiting to-Exploiting Cycle Cycle byExploiting each,(2)Exploiting StructuresCycleExploiting aCycleExploiting Cycle on sourceStructures CycleStructures Theory tacticalCycle CyclewhichStructures Structures Structures CycleStructures isin Cycle Cycle CycleStructurestreated andStructures MAX-SAT.in in analysis. Structures MAX-SAT.in Applications Structures in Structures Structures in MAX-SAT. in MAX-SAT. MAX-SAT. MAX-SAT.in MAX-SAT. In in inin ofInterna- InMAX-SAT. MAX-SAT. Satisfia-MAX-SAT. InInterna- InInterna- InInterna-SummaryInterna- InInterna-Interna-Interna- In In In InInterna-Interna-Interna-Interna- and Future Work movementmovement infermovement from and hand-annotated actionB and action executed executed ball-action by a by team a data team (e.g.and (see (e.g.once pass away).pass from from TheFor teamour work namesmovement we and used ranking andthe action2010-2011 for the executed 2010- EPL by season a teamI(X F24 (e.g.; Y )= passa from probabilitydistributionpXa probabilitydistributionpXuseap probabilitydistributionpX((Shannons)x, this ynewspaper) log dataH(X or entropy)=− data entities =(p1,...,pn)isthendefinedas:(5) oflike =(p1,...,pn)isthendefinedas: ap discrete=(p1,...,pn)isthendefinedas:( thisx (e.g.) log forp ESPN, random(x automatic) H(Y The variable)= Guardian).(2) tactical X thatp(y) has analysis.log Evenp(y) though (3) ways has the same number of agents (11). We refer to teamnewspapernewspaperismovement currently entities entities infer (e.g.used (e.g. from ESPN, for ESPN, hand-annotated the The real-time The Guardian). Guardian).tal ball-action online number Even Evenvisualizations data of though play-segmentsthough (see− of asMutual for a app random((x, teampx(p Information) yx,p()x,(py x(p y)x,−()X over variable.x,p y(x,p) y)p()x,( yx,p) ap( ypx,( y) match.(px,)x,(tional y Thex, y)bility y)tional) ytional) informationConferencetional Testingtionaltional Given Conference Conferencetionaltional Conferencetional. Conference Conferencetional Conferencetional ontionaltional Conference Conference− content, onTheory on Conference Conference Theory onTheory Conference Conference on on Theory on and Theory Theorynamely on Theoryon and and Applications Theory Theory on and on Applications andApplications on on and Theory and Theory Applications Theory Theory Applications and and Applications ApplicationsTo of and Applications and andSatisfia-Do of Applications Applications Satisfia- of Applications of Satisfia- of Satisfia- Satisfia- Satisfia- of Satisfia- Satisfia- of of of ofSatisfia- Satisfia- Satisfia- Satisfia- 2011 EPL data isbehaviors givenmovement in describing Tableagent infer 1. A a In to fromrecipe agent our hand-annotated approach, usedB). These by a behaviors to team ball-action to are achieveI observed(IX( dataIX;I(YX;(IXY)=;(IXY)=(; (seeXIY)=;(I fromY;X)=(YIXcorresponds)=I;(I)=I(YX;(X(YX)=;Ix;Y par-()=;Y;Xwhere)=YY)=;y)=)=YY)=p(Forwherex, 0pa(plogx, yprobabilitydistributionpXwith()x,Forpyp logour((p) yx,x, 0=( log)x,p logy yeach(0ourp)x,) ywork(pandlog) logx,H( y logpx, =)p(p( ywork logx,(X( 0Hy) teamx,thepx,￿) log( andy)=∈( welogx, yX) y) logbase(5)) yx)= log logwethe)(5) playingused logX(5) baseof used(5)(5)(5)p the =(p1,...,pn)isthendefinedas:the((5)x ofp)(5) logarithm eachthe((5) logthex 2010-2011)(5)(5) logp logarithm(5) 2010-2011((5)x otherp)(x) determines team determines EPLy EPLY(2) twice season(2) season F24 F24 next subsection).agent Aagent to agent A to B). agent These B). These behaviors behaviors are observed are observed from from par-Opta par- feed,thisthis datawhich data has consists has been been widely of widely 380 used,games used, there there and are are more no no systems systems than which which−−￿−∈−−￿−∈where−−−−−− 0(Shannons)this− logp(= dataxp 0)(ppx( and(x)yp) has)(p entropyxy( theyxp)p)(x(p( been￿yp−∈(y) basex)yp()xp()py of()−px(p)((pxy( of)x widelya(yp))xbilityp)( discretep the)y(Li,(py)bilityy(bility))y logarithmR.; Testing)bilitybility used, Testing Chellappa, Testing randombility. Testingbilitybility Testing. thereTesting. determinesbilitybility Testingvariablebility Testing.bility R.;.. Testing.￿ are Testing∈ and Testing Testing.. XnoZhou,. that. . systems. S. has 2009. which Learning Multi- behaviors, , as short, observable segments of coordinateda goal (?). Aevents,next team subsection). performing as well as post-analysis a group of these forthat prominent plans a to team’s television= possession(onceHx(XxXx)+yX homexXyYx andHxyXYX(xYXyYyOpta stringxyXY) andxYwhereXyYxXxxyHYXyxXX onceY(feed,yX,YyXy is0YyY logY of)Y away). which= lengthH 0 and(X)= Thexconsists theTXx1 base,X team the ofp of(namesthe result-x) 380 log logarithmp( gamesx and) rankingdetermines and more(2) than analyze the tactics of anext team subsection). wetial are spatio-temporal required to knowtracings, the which in this case refers tothe￿∈ ball￿∈￿the unit,￿∈∈￿￿Opta∈￿∈the∈ unit,￿∈ e.g.￿∈￿ unit,∈￿∈∈ e.g.￿∈ feed,a￿ if￿use∈H∈￿ probabilitydistributionpX￿∈ e.g.∈ base∈( if￿￿∈Y∈￿ basethis∈￿)=￿∈ if∈ which￿ 2∈ base the 2data the 2measure the measureorconsistsp(y datameasure￿)∈ log￿− is∈Li,p islike(Modal inyLi, ofin R.;Li,)=(p1,...,pn)isthendefinedas: is bits, bits, thisR.; inLi,Chellappa, R.;380Li,Densities bits, Chellappa, ifR.;Li, Chellappa, if R.;for itsLi, itsLi,games R.;Chellappa, if Chellappa, theautomaticR.; theLi, itsR.; Chellappa,(3)Li, R.;Li,Li,onChellappa,R.; thenatural naturalChellappa, R.; andR.; R.; R.; Discriminativeand Chellappa, natural andR.;Chellappa, and Zhou,Chellappa, R.; Chellappa,R.; R.;and Zhou, Zhou,tacticalmoreand and R.; R.;and S. Zhou, Zhou, andZhou, and2009.R.; S.Zhou, R.; R.;than R.; 2009.andS.TemporalZhou, analysis. andS. and S.Learning 2009. S. Zhou,2009. Zhou, 2009. Learning2009.S. Zhou, Learning2009.InteractionS. Learning S.Multi- Learning S. 2009. 2009. Multi- Learning2009. Multi- Multi- Learning Learning Multi- Multi- Learning Learning Multi- Multi- Multi- Multi- Multi- A plantialcan spatio-temporaltial be spatio-temporal definedB as tracings, an orderedtracings, whichsequence which in this in case this of caserefersteam refers to ball to760,000 ball events.use this Each data team or data plays like this 38 forgames automatic each, tactical which analysis. − H(Y )=− p(y) logx HXp((yX,) Y )= (3) p(x, y) log p(x, y) (4) movement and action executed by a team (e.g. passachieve fromuse a this majornewspaper dataA goalplan or (e.g.datacan entities winninglike be defined this (e.g. a for ESPN,match), as automatic an ordered The caning Guardian). be tactical sequence number said analysis. of= of Even= play-segmentsfor= teamH=H(number=XH the=( thoughX)+H(=numberXH)+=(H 2010-2011X)+(H=XnumberH(=X)+(H==H)+Y e(H)+X(Hthe=760,000)(Y thenH(XH)+ e(HYH)X(H)+ unit,() thenH(XH for(Y)+YX eH(X(HY)+)( its)X,H)+ thenX()+)Y(H each e.g.EPL itsX,()+( YinH)YX,H(H events.HY)H)Y(( inits nats,ifY(H)X,(H()YYX,) datapossession base(nats,yH() inY )X,H− Y())X,H nats,( etc.) YHX,H 2Each is(etc.) YHX,( the( YX,X,) given( The etc.)X, Y TheY measure Y)) teamY) isThe term￿)∈ term in therefore: termTableplays islog log in log1/pi bits,1/pi 1. 38 1/piindicates ifindicates games its indicates the natural each, which position of the ball and whichA plan teammovementcan has be possession defined infer from as ofan hand-annotated it ordered at sequence ball-action of team data (see760,000− −−whereFor events.−−− H− our 0−( logY−H￿−)=∈ Each(− work−Y =−y)= 0Y and teamwe thepModal(yManifold usedModalp) baseModal( playslogy)Modal Densities logModalp ofthe(Modaly Densities Densitiesforp the) 38Modal(Modaly 2010-2011Densities Group) DensitieslogarithmModal− gamesonModalDensitiesModalModal Densitieson DensitiesDiscriminativeon Activity DensitiesDiscriminativeon Discriminative(3) Densities on Densitieson determinesDensities each, onDiscriminative(3) EPLDiscriminative DiscriminativeRecognition. on onDiscriminative Discriminativeon Discriminative which onseason Temporal on on Discriminative Discriminative Temporal Discriminative TemporalIn Temporal F24CVPR Temporal TemporalInteraction Temporal Interaction. Interaction Temporal Interaction Temporal TemporalInteraction Interaction Interaction Interaction Interaction Interaction behaviors describingmovement a recipe infer from used hand-annotated by a team to ball-action achieve data (see N T −T number e then its in￿∈ nats,− etc. The term logx 1/piX y indicatesY movementagent infer A to from agent hand-annotated B). These behaviors ball-action are observed data (see fromcorrespondsto par- be employingFor withFor ourbehaviors eachtacticsour work work team. describing weHowever, we playingused used as athe recipe soccer eachthe 2010-2011 2010-2011 other used is low-scoring, by team EPL a= EPL team ( twice1 season season to)+1. achieve F24the F24 Ifthe amount thethe amount possessioncorresponds amount of uncertaintyof uncertainty of uncertainty string with associated associated− is each associated smallery teamManifoldYMolineaux, withManifold withManifold inManifold withplaying theManifold the fordura-Manifold corresponding thefor correspondingGroupManifold for M.ManifoldGroup correspondingforManifold Group each2008.Manifold forManifoldManifold￿ ActivityGroupfor∈ Group Group forActivity forGroupActivity Workingother￿∈ Group forActivityGroup forActivity Recognition. forActivity for ActivityGroup GroupteamRecognition. Recognition. Group GroupSpecificationActivity Activity Recognition. Recognition. Activity Recognition. ActivitytwiceRecognition. Activity InRecognition. CVPR In Recognition. for Recognition.CVPR InRecognition. In Rush. CVPR InCVPR.CVPR In 2008.CVPR. In.. In InCVPRCVPR..CVPRCVPR. . .. every time step (i.e. everybehaviors second).thisnext datadescribing Tosubsection). has do this, been a recipe we widely infer used the used, by there a team are to no achieve systemsTo which analyzecorresponds the unit, tactics with e.g. of ifH each basea( team,Y )=y 2 teamY the￿∈ we measure playing requirep(y) log is in toeachp( bits,y know) other if its where the team natural(3) twice next subsection). a goal ( ). A team performing a group of theseSummaryT plansoutcome. to andoutcome. Futureoutcome.the It can It amountOpta canWork Italso can also feed,be of also be uncertainty viewed viewed be which viewed￿ as∈ as the associated−consists theMolineaux, as amountMolineaux, theMolineaux,amountMolineaux, amountMolineaux,Molineaux, with of M.of ofMolineaux,Molineaux, M.information 2008.380informationM. theofMolineaux,Molineaux, 2008.M.Molineaux, 2008.informationMolineaux, correspondingM. M. gamesWorking M.2008. 2008. 2008. M.Working WorkingM.2008. Working2008.M. 2008. M.Working SpecificationWorking M.and M. Working 2008. 2008. Specification 2008.Working 2008. moreSpecification SpecificationWorking WorkingSpecification Specification Working for Specification than Rushfor Specification Specification forRushSpecification for 2008 forRush Rush 2008 forRush 2008 Rush2008 for 2008for 2008 for for Rush 2008Rush Rush Rush 2008 2008 2008 2008 a goalnext (?). subsection).A team performing a group of these plans to (oncecontinuous homeOptaa goalOpta and and feed, ( oncefeed, complex). which A? away). team which due consists performing to Theconsists the team various of of a 380 names group 380 multi-agent games games andoftion these ranking and than andin- plans more more, we tothe than discardthan ball is and it.(oncenumber Towho representhome has e possessionthen and its onceeach in nats, away). play-segment ofyInterface. it etc.Y at everyThe Technical term team time log names report,step 1/pi (i.e. Knexusindicates and ranking Research Corp. tial spatio-temporal tracings, which inball thislocation case refers from to ballthe data feed.use? thisA Weplan data describecan or bedata our defined like method this as anfor ordered automatic sequenceSummary tacticalSummarySummarySummarySummarySummary ofSummary analysis.gainedSummary team andSummarySummary andSummary andSummary Future(once andgained and byFuture and Futureoutcome. and observing Future andFuture Futureandhome760,000 byWork and Future and andWorkFuture observingWork and Future ItFuture WorkFuture andFuture Workcan that Future Work events. Workalso once Work outcome. that Work Work Work be Work outcome. away). viewed EachInterface. Thus,￿∈Interface.Interface. as team The Thus,Interface. the entropyInterface.Interface. TechnicalInterface. amount team playsentropyInterface. Technical TechnicalInterface.Interface. Technical isInterface. Technical TechnicalnamesInterface.Interface.of report, Technicalmerely38 is information report,Technical report,Technical merely games report,Knexus Technical andreport,Technical report, Technical Technical Knexusreport, Knexus report, report,rankingKnexus Research Knexuseach, Knexus Knexusreport, Research report,Research report, report, Knexus Knexus Research Researchwhich Corp. KnexusResearch Research Knexus Knexus Corp. Research Corp. Corp. Research Research Corp. Research Corp. Corp. Corp. Corp. Corp. A plan can be defined as an ordered sequence of team 760,000achieve events. a major Each goal team (e.g. winningplays 38 asTo match),games= Do{s0, .each, can . . , s beT − which said1}, thegained quantizedfor bythe observing the amount 2010-2011 ballthat of position uncertainty outcome. EPL data isOpta.Thus, associated tabulated is 2012. entropygiven withwww.optasports.com in is the Table merely corresponding 1. . achieve aA majorplanmovement goalcan be (e.g. defined infer winning from as ana hand-annotated match), ordered can sequence bevia ball-action said Figure of team 2. data Atforteractions,t the0 (seethe 2010-2011760,000achieve ball labeling is passed aevents. EPL major and to data segmentinggoalEach the is location (e.g. given team winning the in plays at Table gamet1. a 38 match), 1.into gamesTo aTo DoTo seriesTo Do can DoTo DoTo each, DoTo DobeTo DoToTo DoTo said Do DoTowhichDo every Do second).forgained the 2010-2011 To by do observing this, EPL we that inferdata outcome. is the given ballThus, in location entropy Table 1.is from merely tobehaviorsFor be employing our work describingtactics we used. a However, recipe the used 2010-2011 as soccer by a is team low-scoring,EPL to season achievea statisticala statistical F24a statistical averageoutcome.corresponds average average of It of uncertainty can uncertainty of withalso uncertainty be each viewedorOpta. orRiley, information.Opta.Opta. information. orteam 2012.Opta. information.asOpta. P., 2012. 2012.Opta. the andplayingwww.optasports.comOpta. 2012.Opta. 2012.amountwww.optasports.com Veloso,www.optasports.com 2012.Opta.Opta.Opta. 2012.www.optasports.comOpta. 2012.www.optasports.comwww.optasports.com eachwww.optasports.com 2012.of 2012.M. 2012.www.optasports.com 2012.informationwww.optasports.com 2002.otherwww.optasports.comwww.optasports.comwww.optasports.comwww.optasports.com Recognizing team. .. twice. . .. Probabilis-. . . .. to be employingbehaviorstactics. describing However, as a recipe soccer used is low-scoring, by a team to achieveof plans iscorresponds extremelyt difficult. with each Hence team recognizing playing each teamat each other tac- time team step. twice An examplea statisticalTo analyze of the average description the oftactics uncertainty of is a team, given or information. we in require to know where behaviorsnext describing subsection). a recipe used by a teamThe to next achieve action labeledTo analyzecorrespondsto is be at the employingOpta4a tacticswhere goal feed,with ( a? oftactics). player each whicha A team, team. team However,takes consists we performing playing requireon an as of soccer toeach 380a know group is othergames low-scoring, where of team these and twiceplansthe more data to than feed.To analyzegained The(once method by home the observing tactics ofand doing once thatofRiley, atic outcome. thisteam,Riley, away).Riley, Opponent P.,Riley,Riley, is P.,and P.,Riley,we best andThus,Riley, TheandP.,Riley, Veloso, P.,require Movement andRiley,P., Veloso,describedRiley, Veloso,and Riley, teamRiley, P.,entropy andP., Veloso, M.Veloso, Veloso,and andP., Veloso,P.,to M.P., M. P.,names2002.and Models. Veloso, Veloso,andknow isandM. and 2002.M. 2002.in Veloso,M.merely Veloso,M. Veloso,2002. RecognizingVeloso, 2002. M.andwhere M.2002. In 2002.Recognizing Birk, M. 2002. M.Recognizing ranking M.Recognizing Recognizing Recognizing2002. A.; 2002. 2002.Probabilis- Recognizing Coradeschi, Probabilis- Recognizing Recognizing Probabilis- Recognizing Probabilis- Probabilis- Probabilis- Probabilis- Probabilis- Probabilis- Probabilis- a goal ( ). A team performing a group of these plans to continuous and complex due to the variousFigure multi-agent 3. in- the ball is andH(X who)= has possessionp(x) log p(x of) it at every time step (i.e. continuousa goal and (? complex). A? team due performing to the various a group multi-agent ofopposition these in- plans player. tothetics ball As using nothing(once iscontinuous and(once the whohome occurredMAPRachieve home has and and framework possession andcomplex abetween once major once away). goal duethe we away). of described ittimeto (e.g. Theat the everyThet winning1 various teamteam previously time names amulti-agent names step match), is and(i.e. and can ranking in- rankingFigure be said 5(b).the ballafor statistical isH and the(HX( 2010-2011)= whoX)= average hasp possession of(px( EPLuncertainty)xtic) logS.; logtic Opponentticp dataandp( Opponent Opponenttic(xticxtic) Tadorokoro,) of Opponenttic is Opponentortic it MovementOpponenttic given information. at Opponent Movementtic Opponent Movementtic everytictic Opponent Movement Opponent Movementin OpponentS., Movement Opponent Movement Models.(1) Table eds., time Movement Movement Models. Models.(1) MovementRoboCup-2001: Models. MovementstepModels.1. Movement Movement InModels. Models. Birk, In In Models. Models. (i.e. Birk, Birk, In InModels. A.; Models. Birk,In In Models. Models.Birk, A.; Coradeschi, Birk, Birk,Robot In A.; Coradeschi, Birk, A.; In A.; InCoradeschi, Soccer Coradeschi,In Birk, Birk, A.; Coradeschi, Birk, Birk, Coradeschi, A.; A.; A.; A.; Coradeschi, Coradeschi, Coradeschi, Coradeschi, A plan can be defined as an ordered sequence of team 760,000teractions, events. labeling Each and segmenting team plays the 38From game games ourinto set a each, series of play-segments whichevery second). for a givenH( ToX)= dox team,X this,S.;S.; andpS.;S( weAx andS.; and) Tadorokoro,S.;or log inferandS.; Tadorokoro, Tadorokoro, andSpS.; andS.;B( Tadorokoro,x Tadorokoro,, andS.;) andtheS.; Tadorokoro,S.;S.; S.,and Tadorokoro, andTadorokoro, balland andS.,eds., S., Tadorokoro, Tadorokoro, eds.,Tadorokoro, S.,eds., Tadorokoro, locationS.,RoboCup-2001: S., eds.,S., eds.,RoboCup-2001:RoboCup-2001:(1) eds., S., eds.,S.,RoboCup-2001:RoboCup-2001: eds., eds., S.,RoboCup-2001: S.,RoboCup-2001: from S., S., eds.,RoboCup-2001: eds., eds., RobotRoboCup-2001:RoboCup-2001: RobotRoboCup-2001: Soccer Robot Robot SoccerRobot Soccer Robot Soccer Soccer Soccer Robot Robot Soccer Robot Robot Soccer Soccer Soccer Soccer achieve a major goal (e.g. winning a match),and t4, wecan infer be said theimpossible ball wasfor without dribbled the 2010-2011 these in a labelled straight EPL plans. data line is and given in Table 1. x xXX World Cup V. Springer Verlag. teractions, labeling and segmenting the game into a series every second).teractions, Toto be do labeling employing this, andwe infertacticssegmenting the. However, ball the location game as soccer into from isa series low-scoring,every second). To∈ do∈ ￿ this,∈ we infer the ball location from achievebehaviors a major goal describing (e.g. winning a recipe a used match), by cana team be said to achievefor thecorresponds 2010-2011of plans is extremely EPL with data each difficult. is team given Hence playing in Table recognizingwe each can 1. build other team a team distribution tac- twicethe to data characterizeTo feed.analyze The￿￿ the methodthe tacticsx expectedWorldXWorldWorld of of CupWorld doingWorld a Cup CupWorld team,V. be-World Cup SpringerWorld V CupV. this.World SpringerCup SpringerWorld V weWorld V.CupWorld V Cup Springer. is.V Springer Springer Verlag.require. Cup SpringerbestVCup V Cup. Verlag.Cup. Verlag. Springer Springer VV Verlag.. described V. Verlag.SpringerV Verlag.. Springer. to Verlag. Springer Springer know Verlag. Verlag. Verlag. Verlag. Verlag. inVerlag. where to be employing tactics. However, as soccerat a isuniform low-scoring, velocityTo between overcomeof plansTo these analyze is this extremely issue,two the locations. we tactics difficult. quantize of We aHence team, a do match recognizing we intorequire equal toteam know tac- where the data feed. The methodH(X)=￿∈ of doingp(x) log thisp( isx) best described(1) in of plansto is be extremely employinga goal difficult. ( tactics). A team. Hence However, performing recognizing as soccer a group team is low-scoring, tac- of these plansthe data to feed.To analyze Theticscontinuousmethod using the the tactics andof MAPR doing complex of framework a thisteam, due is we best to we requirethe describedhavior various to inknow previouslymulti-agent in each where location.MeasuringAnalysis is in-AnalysisAnalysis WeFigure Usingthe Using do Team Using ball 5(b).Entropy this Entropy is Entropy andas Tactics follows: Maps who Maps Maps has via Givenpossession Entropy the ofMaps it at every time step (i.e. continuous? and complex due to the variousthe same multi-agent thing between in-temporalt5 chunksand(oncet6, which before home we the and use ball once to is describe passedaway). team The behavior. team names and ranking x X tics using the MAPR framework we described previously is ticsthe using ballimpossibleteractions, the is and MAPR who without labeling framework has these possession and labelled segmenting we described of plans. it at the every previously game time into step is a series (i.e. FigureDeterminingAnalysisevery 5(b). Using second). Parameters Entropy To do Maps￿∈ this, we infer the ball location from continuousachieve and complexa major goal dueto (e.g. the winning various amulti-agent match), can in-Figure be saidt 5(b).thet ballfor is and the 2010-2011 who has possession EPL data of is it given at everyobservations in Table time step1. ofPlay-Segment (i.e. aDetermining team,Determining we determine Parameters Representation Parameters the subset of play- teractions, labeling and segmenting thefrom game the into location a seriesAs at these6 toimpossible segmentsevery7. UsingofTo second). plans without do overcomethe notis same extremely Todescribe thesethisprocedure, do labelled this, issue, a difficult. method we we plans. we infer quantize ofHence achieving the recognizing a ball match locationa into team equal from tac- HowDetermining discriminative Parameters are our representations? That is, do impossibleteractions, without labeling these labelled and segmenting plans. the game into a series every second). To do this, we infer thesegments ball locationSd, which fromHowHow originated discriminative discriminativeMeasuringAnalysisthe datain quantized areUsing feed. are our our Team The Entropy representations? representations? area method Tactics Mapsd. of From doing via That That Entropy this is, do is best Maps described in ofto plans be employing is extremelytactics difficult.. However, Hence as recognizingcan soccer estimate is low-scoring, team the tac-specific ball position goal,Tothe overcome datatemporal we andticsTo do feed.analyze using team not chunks this call The the possession the issue, MAPRthem method which tactics plans, we framework for we of quantize of butthe doinguse a team,play-segments to wea thisdescribe match we described is require best teaminto. described previously equal to behavior. knowGiven in where is theteams partialHow have discriminative spatio-temporal unique styles are of play?our tracings representations? And if of so, the can team we That detect (i.e. is, do To overcomeof plans is this extremely issue, we difficult. quantize Hence a match recognizing into equal team tac- the play-segmentteams vectorsteams haveMeasuring have within uniqueDeterminingFigure uniqueS stylesd 5(b)., styles we Team of Parameterskeep of play? play? aTactics And count And if if so, ofso, via can the can Entropy we we detect Maps ticscontinuous using the and MAPR complex framework due to we the described variousremaining previouslymulti-agent timesMeasuring -We is as in- use thethetemporal these stoppages dataFigurethe play-segmentsAs feed. Team ballchunks these5(b). are Theis segments and Tactics tagged which method whoto form we doin has via theofnot use a possession library doing describe dataEntropy to describe or thisplaybook a methodof is teamMaps it best at everyof behavior. described achieving timeball step in a movement), (i.e.whenteamsPlay-Segment they have we deviate chunk unique from stylesRepresentationthis this signalstyle of play? and upquantify And into if so,discrete it? can we seg- detect temporal chunks which we use to describe team behavior. impossible without these labelled plans.locations of wherewhen thewhen they ball they deviate travels deviate from fromor this occupies this style style andand during quantify quantify these it? it? tics usingimpossibleteractions, the MAPR without labeling framework these and labelled segmenting we described plans. thefeed. previously game It is into worth is a noting seriesplay-segments,Figure hereAs these thatevery 5(b).specificsegments all= second).data goal,p1 is,p do we normalized2,p not To do3,...,p describenotdo callthis,m onto them, a we where method plans, inferm of buttheis achieving theplay-segments ball location aments. from -Play-Segment calledwhenGivenHow they play-segments. the discriminative deviate partial Representation from spatio-temporal this are We style our represent representations?and quantify tracings these it? of That play- the is, team do (i.e. Play-Segment RepresentationPTo overcome{ this issue,} we quantizeplay-segments. a match into As equal thisMutual representationteams InformationMeasuring have unique will yield styles Team a of very play? Tactics high- And if via so, can Entropy we detect Maps As theseimpossible segments without do not describe these labelled a method plans. of achievinga field a of size 100number× 100,specific of with uniqueWe all goal, usepositions play-segments wethese do play-segments notgiven call within for them teams the to plans, playbook. form but a libraryplay-segments Using or playbooksegments. Mutualof MutualGiven as Informationball an Information the occupancy2 movement), partial spatio-temporal map we which chunk describes this tracingssignal which up of intothe areas discrete team (i.e. seg- ofTo plans overcome is extremely this issue, difficult. we quantize Hence recognizing a match into team equal tac- Measuringthetemporal data feed. chunks Team The which method Tactics we of use doing via to describedimensional Entropy this is team best Maps feature described behavior. vector inMutual (Dwhen), theyweInformationH would( deviateX)= prefer from thisp the(x style) log dimen- andp(x) quantify it? (2) specific goal,To overcome we do not this call issue, them plans, we quantize but play-segments a matchattacking into. equalleft toGiven right.this playbook theWe partial useplay-segments, we these spatio-temporal can: play-segments= tracings top1 form,p2,p a3 of,...,p library them team or,playbook where (i.e.mofisof the the fieldball thements movement),Play-Segment ballH(H -X was( called)=X)= during we play-segments. chunk−p thatRepresentation(px(x)) play-segment.log this logpp( signal(xx)) We up represent into(2) discrete these seg- play- temporaltics using chunks the MAPR which framework we use to we describe described team previously behavior. is MeasuringFigureAs these 5(b). segmentsTeam Tactics do not describe via Entropy asionality method of to Maps achieving be D for visualization a purposes.−− Tox X achieve this, P { } H(Xx )=xXX ∈ p(x) log p(x) (2) We use these play-segments to form a library or playbook of ball movement),Play-Segmentnumber we chunk of unique Representationthis play-segments signal up into within discrete the playbook. seg- Using segments as an occupancy∈ ￿− map which describes which areas temporalAsimpossible chunks these segments which without we do these not use describe tolabelled describe a plans. method team of behavior. achieving a play-segments,specific goal,= wep1 do,p2 not,p3 call,...,p themm plans,,we where can butm useplay-segmentsis either the the.ments mean,MutualGiven - median, called the Information play-segments. mode,partial￿∈￿ spatio-temporaltotalx X count We or represent tracings these of the play- team (i.e. this playbookP we{ can: } of the fieldH the(Y )= ball was￿ during∈p(y) log thatp(y) play-segment.(3) play-segments, = p1,p2,p3,...,pm , where m4.is the OCCUPANCYments -Play-Segmentnumber called MAPS ofplay-segments. unique Representation play-segments We represent within the these playbook. play- Using segmentsH as(HY an()=Y )= occupancy−p(py()y map) log logpp which((yy)) describes(3) which areas As thesespecific segmentsTo overcome goal, do we not do this describenot issue, call them we a method quantize plans, of but a achievingplay-segments match into a equal. GivenWe the use partial these play-segments spatio-temporal to form tracingseven a library of an the entropyor playbook team measure (i.e.of to describeball movement), eachH−(H−Y ()=Xp()=y weSdY), chunk whichp(yp)( logthisx will) logp( signalyp)(x) up into discrete(2) seg- P { } Measuring Team Tactics via Entropy Maps y yY Y￿∈ − (3) number of unique play-segments within the playbook. UsingGiven we havesegments ball trackingthis as an playbook occupancyplay-segments, and team we can: map possession which= p infor-describes,p ,p ,...,p whichyield an areas, where occupancym is theorof team the behavioral fieldments the - ball called map. was∈￿∈ play-segments. Vectorizing during− x X that play-segment. the We represent these play- specificWetemporal goal, use wethese chunksdo play-segments not call which them we to plans, use form to but a describe libraryplay-segments or teamplaybook behavior.. of Givenball the movement), partial spatio-temporal we chunk this1 2 tracingssignal3 up ofm into the discrete team (i.e. seg- ￿ y Y￿∈ P { } ￿∈ this playbookWe useplay-segments, we theseAs can: these play-segments segments= do top not1 form,p2 describe,p a3,...,p library am methodmation, or,playbook where of we achievingm canofisof thepartition the field aball thements the movement),Play-Segment ball ballnumber - was tracking called of during we unique play-segments. chunk data thatRepresentation play-segments into play-segment. this pos- signal We within up representoccupancy into the discreteplaybook. these map, seg- play- Using gives us oursegmentsD-dimensional as an occupancy spatiotemporal map which describes which areas number of uniqueP play-segments{ within thesession} playbook. strings Using (i.e. continuoussegmentsthis movement playbook as an occupancyof we the can: ball map for which describesfeature vector whichx areas. The meanof occupancy the field the maps ball using was during entropy that play-segment. play-segments,specific goal,= wep1 do,p2 not,p3 call,...,p themm plans,, where butmplay-segmentsis the .mentsGiven - called the play-segments. partial spatio-temporalA We represent tracings these of the play- team (i.e. this playbookP we{ can: } a single team without turnoverof the or field stoppage), the ball was where duringO = that play-segment.to describe each area is shown in Figure 4 and can give a numberWe of unique use these play-segments play-segments within to form the playbook. a libraryA or UsingAplaybook ofBsegmentsBball as movement), anB occupancy we map chunk which this describes signal up which into discrete areas seg- {o0 ,..., oI−1} and O = {o0 ,..., oJ−1} refer to the pos- indication of redundant patterns. As can be seen the top this playbookplay-segments, we can: = p ,p ,p ,...,p , where m is theof the fieldments the - ball called was play-segments. during that play-segment. We represent these play- P { 1 2 3 sessionm} strings associated with each team and I −1 and J −1 teams have higher entropy over most of the field compared number of unique play-segments withinare the the playbook. number Usingof possessions.segments We then as quantize an occupancy the field map whichto the describes lower teams which - this areas gives an indication that these teams this playbook we can: into D bins where D = l × wofequal the field size areas the ball and was vectorize during thatutilize play-segment. more options (i.e. less predictable) which is intuitive the field via the columns. As the possession strings vary in as these teams normally have more skilled players. As the length, we apply a sliding window of length T to quantize or frequency counts incorporate temporal information (more

1369 7

66

(1) Manchester United (2) Chelsea (3) Manchester City (4) Arsenal (5) Tottenham

55

44

(6) Liverpool (7) Everton (8) Fulham (9) Aston Villa (10) Sunderland

33

22

(11) West Bromwich Albion (12) Newcastle (13) Stoke City (14) Bolton (15) Blackburn

11

0 0

(16) Wigan (17) Wolverhamption (18) Birmingham City (19) Blackpool (20) West Ham United

Figure 4: The mean entropy maps for each of the twenty English Premier League teams characterizing their ball movement patterns. The maps have been normalized for teams attacking from left to right. The bright red refers to high entropy scores (i.e. high variability) and the blue areas refer to low entropy scores (i.e very predictable behavior). counts, means the ball is moving quicker), we used the to- % Teams Correctly Identified tal count of entires as the entropy measure normalizes this 50 information. 40 30 20 5. STRATEGY ANALYSIS 10 0 5.1 Discriminating Team Behavior Statistics Occ Maps Combined Evaluating team strategy is a very difficult task. The ma- jor hurdle to overcome is the absence of strategy labels. But Figure 5: The identification rate for correctly iden- given we know the team identity, and assuming that teams tifying home performances using event-labeled data exhibit similar behaviors over time, we can treat the task as (i.e. no location information), occupancy maps (i.e. an identification problem. We can do this by answering using no event information, just location), and a combina- only ball movement information, can we accurately identify tion of the two. the most likely team? We model team behavior using a codebook of past per- formances. If a team’s behavior is consistent, then previous knowledge of where teams operate could boost the discrim- matches will be a good predictor of future performances. inating power. For the experiments, we used 380 games of the season and To conduct the experiments, we compared our occupancy used a leave-one-match-out cross validation strategy to max- representation, to twenty-three match statistics currently imize training and testing data. Before we investigate the used in analysis (e.g. passes, shots, tackles, fouls, aerials, difference between home and away performances, we first possession, time- in-play etc.). We also combined the two need to obtain the best possible representation. To evalu- inputs by concatenating the feature vectors. For classifica- ate this, we wanted to see how effective event-labeled data tion, we used a k-Nearest Neighbor approach (with k = 30) was in discriminating between different teams, and if having and all experiments were conducted using D = 10 × 8

1370 Exp Event-Labeled Occupancy Maps 1 18 H v H 19.26 38.79 H v A 16.09 30.08 A v A 13.98 36.41 A v H 16.36 30.34

Table 2: The hit rate accuracy of experiments which tested home (H) and away (A) models against home and away matches (e.g. HvH refers to home model tested on home matches and HvA refers to the home model being tested on away matches). Actual Team Ranking

20 0 do this, we simply subtracted the home occupancy maps 1 Predicted Team Ranking 20 from the away maps and divided by the away occupancy. The difference maps for all twenty teams is given in Figure 7 (a) Home vs Home and it makes for compelling viewing. To make it easier to quantify the difference in occupancy, we calculated the dif- Figure 6: Confusion matrices of the team identifica- ference with respect to certain areas of the field. Specifically tion experiments using the combined representation. we calculated the difference: for the whole field (W), the at- tacking half (H), and the attacking-third (T) - these values are listed below each difference map. As can be seen from spatial areas (heuristically, we found those values of k and the difference maps, spatially, nearly all the teams (18 out D gave the best performance). To reduce the dimensionality of 20) had more possession in the attacking half and prob- but maintain class separability for the spatiotemporal rep- ably more telling is that 19 out of the 20 teams had more resentation and the combined representation, we used linear possession in the attacking third. Seeing that the shoot- discriminant analysis (LDA). We used the number of teams ing proficiency is essentially the same (10% vs 9%), we can as classes (C = 20), which resulted in a C − 1 = 19 dimen- point to the observation that the more possession in the at- T sional input feature, y = W x, where W can be found via tacking third leads to more chances, which in-turn leads to solving more goals. A potential statistic to back this observation up is that Chelsea - who were the only team to have less WΣbW occupancy in the attacking third - had the smallest discrep- arg max Tr( ) (1) W WΣwW ancy between home and away shots (only 12, the next was Arsenal with 45). where Σb and Σw are the between-scatter and within-scatter matrices respectively. To test the various representations, With the absence of labels to compare against it is im- only identity experiments on home performances were tested possible to say whether this was actually this case as other (home and away comparisons are be done next). Results for factors may have contributed to the home advantage (i.e. these experiments are shown in Figure 5. From this fig- referee’s [22], shooting chance quality, game context (i.e. ure, it can be seen using spatiotemporal data greatly im- winning, losing, red-cards, key injury, derby matches etc.). proves the discriminability between different teams (19.26% However, through the use of spatiotemporal data, we can vs 38.79%), and fusing the two together boosts performances provide evidence of behavioral differences which can aid in again (46.70%). This result makes sense as the location of the analysis of performance and decision making. This ap- where teams play in addition to what they do should charac- proach can also be used to flag and predict individual team terize their behavior. The confusion matrix of the combined performances, and in the next section we show methods in representation shows that teams who play a similar style of- which these can be applied. ten get confused with each other (e.g. the top 5 teams and teams 13-15) and from viewing the entropy maps in Figure 4, 6. PRE/POST GAME ANALYSIS we can see that these teams look similar. Given a coach or analyst is preparing for an upcoming match, having a measure of how variable a team’s perfor- 5.2 Comparing Home vs Away Behavior mance is would be quite beneficial. For example, the coach If a team plays in the same manner at home as they do or analyst may have viewed a previous match and formed a away, the home model should be able to yield similar perfor- qualitative model based on their expert observation. How- mance in identifying away matches as they do to the home ever, this model is only formed by a single observation and matches. To test this theory, we used our home models to may be subject to over-fitting. Having an measure which identify away performances and our away models to identify could indicate how variable a team’s performance is would home performances. From the results (see Table 2), we can be quite useful. Given they have a feature representation of see that there is a drop in the hit-rate of the occupancy map each of the previous performances of a team, our approach representation – 8.69% for the home model tested on the could be a method of determining the performance variance. away matches and 6.07% on the reverse case. Even though To do this, the distance in feature space between each of the not excessive, this drop in performance suggests there is a past performances, y, and the mean, ˆy, can be calculated change in the spatial behavior between home and away per- where the mean is formances. M To explore this aspect further, we visualized the difference 1 X ˆy = yi (2) M in occupancy between the home and away performances. To i=1

1371 100

(1) Man United (2) Chelsea (3) Man City (4) Arsenal (5) Tottenham W=3.8%, H=7.6%, T=10.7% W=1.1%, H=2.5%, T=-0.5% W=1.5%, H=3.1%, T=8.7% W=3.9, H=4.2%, T=9.5% W=0.1%, H=3.0%, T=8.9% 50

(6) Liverpool (7) Everton (8) Fulham (9) Aston Villa (10) Sunderland W=1.8%, H=3.5%, T=9.5% W=-2.9%, H=-0.3%, T=5.6% W=-0.2%, H=1.6%, T=4.5% W=0.1%, H=0.6%, T=0.2% W=2.6%, H=7.1%, T=8.4% 0

(11) West Bromwich Albion (12) Newcastle (13) Stoke City (14) Bolton (15) Blackburn -50 W=0.6%, H=3.1%, T=9.0% W=5.2%, H=6.7%, T=0.5% W=-3.7%, H=2.1%, T=1.9% W=6.2%, H=5.4%, T=8.0% W=8.1%, H=3.6%, T=7.3%

-100 (16) Wigan (17) Wolverhamption (18) Birmingham City (19) Blackpool (20) West Ham United W=8.1%, H=12.4%, T=8.1% W=-4.3%, H=0.4%, T=4.1% W=5.0%, H=8.3%, T=13.8% W=-1.0, H=3.4%, T=11.9% W=-2.2%, H=-0.7%, T=3.6%

Figure 7: The normalized difference maps between home and away performances for all the 20 EPL teams. In all maps, teams are attacking from left-to-right and a positive value refers to a team having more occupancy in home games, while a negative value refers to more occupancy in away games. Percentages underneath each team give a value on this difference (Key: W is whole field, H refers to forward half and T refers to the attacking third.)

120 home and away performance example, we can show the vari- ation in performance for each team in the EPL by finding the 80 variance in distortion (Figure 9). As can be seen in this fig- ure, each team’s home performance has quite a low variance 40 which gives an indication that when team’s play at home

Distortion Variance 0 they do not vary their approach too much. Conversely, it 0 5 10 15 20 (a) Home Performances seems that the away behavior is quite random so forecasting 120 away performance may be unreliable. In terms of post-match analysis, a similar approach could 80 be used to see if a team’s performance was within the expec- tation range (i.e. ±σ). A good example was Fulham’s away 40 performance against Manchester United. In this match, they lost 2-0 and conceded both goals in the first half (12th and 0 0 5 10 15 20 32nd minute). As can be seen by comparing both occu- (b) Away Performances pancy maps, in their match against Manchester United they occupied a lot more possession in the middle of the field Figure 8: Variance in distortion for (a) home and then normal. This highlights the importance of context, as (b) away performances. after scoring two early goals Manchester United sat back and allowed Fulham to have the majority of possession in non-threatening regions (52% of overall possession) [6]. To and where M is the number of previous performances. A counter this, we would have to normalize for match con- distance measure such as the L2 norm could be used given text (i.e. score, strength of opposition etc.). However, this that the input space has been scaled appropriately (as is the is a major problem as we would limit the amount of data we case in our work), which generates the distance measure via would have to train our model. Future work will be focussed on clustering styles unsupervised to maximize the amount distm = kym − ˆyk2 (3) of context dependent data. where m refers to the game of interest. Returning to our

1372 60

40

20 (a) (b) Distance from Mean 0 0 5 10 15 20 Figure 10: Occupancy maps of the: (top) mean away Opponent Number performance for Fulham, and (bottom) their perfor- mance against Manchester United - in this match Figure 9: Example of the distortion for each away they lost 2-0 and conceded early in the match. performance of Fulham in the 2010-2011 season. In match 16 they played Manchester United, and the performance on the day was far different from their 8. REFERENCES other away performances (they lost 2-0 on this oc- [1] S. Ali and M. Shah. Floor Fields for Tracking in High casion). Density Crowd Scenes. In ECCV, 2008. [2] N. Allen, J. Templon, P. McNally, L. Birnbaum, and K. Hammond. StatsMonkey: A Data-Driven Sports 7. SUMMARY AND FUTURE WORK Narrative Writer. In AAAI Fall Symposium Series, Most sports analytics approaches still only use event-labeled 2010. statistics to drive analysis and decision-making despite the [3] BBC-Sports. Footballers may trial wearing microchips influx of ball and player tracking data becoming available. to monitor health. The reason why this new and rich source of information is www.bbc.co.uk/sport/0/football/21460038, 14 Feb being neglected stems for the fact that it is continuous and 2013. is extremely difficult to segment into categories which would [4] M. Beetz, N. von Hoyningen-Huene, B. Kirchlechner, enable high-level analysis (e.g. team strategy labels). The S. Gedikli, F. Siles, M. Durus, and M. Lames. emerging field of sports spatiotemporal analytics attempts ASPOGAMO: Automated Sports Game Analysis to use spatiotemporal data such as ball and player track- Models. International Journal of Computer Science in ing data to drive automatic team behavior/strategy analysis Sport, 8(1), 2009. which would be extremely useful in all facets of the sports [5] P. Carr, Y. Sheikh, and I. Matthews. Monocular industry (e.g. coaching, broadcasting, fantasy-games, video Object Detection using 3D Geometric Primitives. In games, betting etc.). In this paper, we gave an overview ECCV, 2012. of the types of sports analytics work being done both in [6] ESPNFC. http://espnfc.com/us/en/report/ industry as well as academia. Additionally, we gave a case- 292849/report.html?soccernet=true&cc=5901. study which investigated possible reasons for why the home [7] K. Goldsberry. CourtVision: New Visual and Spatial advantage exists in continuous sports like soccer. Using spa- Analytics for the NBA. In MIT Sloan Sports Analytics tiotemporal data, we were able to show that teams at home Conference, 2012. play have more possession in the attacking third. Coupled [8] A. Gupta, P. Srinivasan, J. Shi, and L. Davis. with the fact that the shooting and passing proficiencies are Understanding Videos, Constructing Plots: Learning a not significantly different, this observation can partially ex- Visually Grounded Storyline Model from Annotated plain why home teams have more shots and score more, and Videos. In CVPR, 2009. in-turn win more at home compared to away matches. Using [9] Hawk-Eye. www.hawkeyeinnovations.co.uk. our feature representation, we also showed examples where pre and post game analysis can be performed. Specifically, [10] D. Henschen. IBM Serves New Tennis Analytics At we were able to show the variation in home and away perfor- Wimbledon. www.informationweek.com/software/ mances for each team, as well as the ability to flag anomalous business-intelligence/ performances. ibm-serves-new-tennis-analytics-at-wimbl/ Our work also highlighted the importance of match con- 240002528, 23 June 2012. text and the limiting factor it could have on training ex- [11] A. Hervieu and P. Bouthemy. Understanding sports amples. In our future work, we will look at unsupervised video using players trajectories. In J. Zhang, L. Shao, methods which cluster playing similar playing styles which L. Zhang, and G. Jones, editors, Intelligent Video can enrich our training data set, without effecting its dis- Event Analysis and Understanding. Springer Berlin / criminating power. Additionally, we are looking to extend Heidelberg, 2010. this approach to focus on using player tracking information [12] S. Intille and A. Bobick. A Framework for Recognizing to discover team formations and plays. Predicting team in- Multi-Agent Action from Visual Evidence. In AAAI, teractions and subsequent performances and outcomes, es- 1999. pecially when they have not played each other is another [13] K. Kim, M. Grundmann, A. Shamir, I. Matthews, area focus of our research. As reliable high-level labels are J. Hodgins, and I. Essa. Motion Fields to Predict Play almost impossible to obtain, predicting match outcomes as Evolution in Dynamic Sports Scenes. In CVPR, 2010. our evaluation tool seems to be the best indicator of im- [14] M. Lewis. Moneyball: The Art of Winning an Unfair proved team modeling. Game. Norton, 2003.

1373 [15] R. Li and R. Chellappa. Group Motion Segmentation [26] Opta Sports. www.optasports.com. Using a Spatio-Temporal Driving Force Model. In [27] S. Pellegrini, A. Ess, K. Schindler, and L. van Gool. CVPR, 2010. You’ll Never Walk Alone: Modeling Social Behavior [16] R. Li, R. Chellappa, and S. Zhou. Learning for Multi-Target Tracking. In CVPR, 2009. Multi-Modal Densities on Discriminative Temporal [28] M. Perse, M. Kristan, S. Kovacic, and J. Pers. A Interaction Manifold for Group Activity Recognition. Trajectory-Based Analysis of Coordinated Team In CVPR, 2009. Activity in Basketball Game. Computer Vision and [17] W. Lu, J. Ting, K. Murphy, and J. Little. Identifying Image Understanding, 2008. Players in Broadcast Sports Videos using Conditional [29] Prozone. www.prozonesports.com. Random Fields. In CVPR, 2011. [30] B. Siddiquie, Y. Yacoob, and L. Davis. Recognizing [18] P. Lucey, A. Bialkowski, P. Carr, E. Foote, and Plays in American Football Videos. Technical report, I. Matthews. Characterizing Multi-Agent Team University of Maryland, 2009. Behavior from Partial Team Tracings: Evidence from [31] SportsVision. www.sportsvision.com. the English Premier League. In AAAI, 2012. [32] STATS SportsVU. www.sportvu.com. [19] L. Madden. NFL to Follow Army’s Lead on Helmet [33] Statsheet. www.statsheet.com. Sensors in Attempt to Prevent Head Injury. [34] D. Stracuzzi, A. Fern, K. Ali, R. Hess, J. Pinto, N. Li, www.forbes.com/sites/lancemadden/2012/07/16/ T. Konik, and D. Shapiro. An Application of Transfer nfl-to-follow-armys-lead-on-helmet-sensors-in/ to American Football: From Observation of Raw -attempt-to-prevent-head-injury/, 16 July 2012. Video to Control in a Simulated Environment. AI [20] R. Masheswaran, Y. Chang, A. Henehan, and Magazine, 32(2), 2011. S. Danesis. Destructing the Rebound with Optical [35] X. Wei, P. Lucey, S. Morgan, and S. Sridharan. Tracking Data. In MIT Sloan Sports Analytics Sweet-Spot: Using Spatiotemporal Data to Discover Conference, 2012. and Predict Shots in Tennis. In MIT Sloan Sports [21] V. Morariu and L. Davis. Multi-Agent Event Analytics Conference, 2013. Recognition in Structured Scenarios. In CVPR, 2011. [36] C. Xu, Y. Zhang, G. Zhu, Y. Rui, H. Lu, and [22] T. Moskowitz and L. Wertheim. Scorecasting: The Q. Huang. Using Webcast Text for Semantic Event Hidden Influences Behind How Sports Are Played and Detection in Broadcast. T. Multimedia, 10(7), 2008. Games Are Won. Crown Publishing Group, 2011. [37] Zonalmarking. www.zonalmarking.net. [23] NBA Shot Charts. www.nba.com/hotspots. [24] D. Oliver. Basketball on Paper: Rules and Tools for Performance Analysis. Brassey’s, Incorporated, 2004. [25] D. Oliver. Guide to the Total Quarterback Rating. espn.go.com/nfl/story/_/id/6833215/ explaining-statistics-total-quarterback-rating, 4 August 2011.

1374