Human-Level AI's Killer Application: Interactive Computer Games

Home , Descent 3, Dungeon Keeper (series), God game, Killer application, Madden NFL 97, Street luge

AI Magazine Volume 22 Number 2 (2001) (© AAAI) Articles Human-Level AI’s Killer Application Interactive Computer Games

John E. Laird and Michael van Lent

Although one of the fundamental goals of AI is to communication with natural language, com- understand and develop intelligent systems that monsense reasoning, creativity, and learning. have all the capabilities of humans, there is little If this is our dream, why isn’t any progress active research directly pursuing this goal. We pro- being made? Ironically, one of the major rea- pose that AI for interactive computer games is an sons that almost nobody (see Brooks et al. emerging application area in which this goal of [2000] for one high-profile exception) is work- human-level AI can successfully be pursued. Inter- active computer games have increasingly complex ing on this grand goal of AI is that current and realistic worlds and increasingly complex and applications of AI do not need full-blown intelligent computer-controlled characters. In this human-level AI. For almost all applications, article, we further motivate our proposal of using the generality and adaptability of human interactive computer games for AI research, review thought is not needed—specialized, although previous research on AI and games, and present more rigid and fragile, solutions are cheaper the different game genres and the roles that and easier to develop. Unfortunately, it is human-level AI could play within these genres. We unclear whether the approaches that have then describe the research issues and AI techniques been developed to solve specific problems are that are relevant to each of these roles. Our conclusion is that interactive computer games provide a the right building blocks for creating human- rich environment for incremental research on level intelligence. The thesis of this article is human-level AI. that interactive computer games are the killer application for human-level AI. They are the ver the last 30 years, research in AI has application that will need human-level AI. fragmented into more and more spe- Moreover, they can provide the environments Ocialized fields, working on more and for research on the right kinds of problem that more specialized problems, using more and lead to the type of incremental and integrative more specialized algorithms. This approach research needed to achieve human-level AI. has led to a long string of successes with important theoretical and practical advancements. Computer-Generated Forces However, these successes have made it easy for us to ignore our failure to make significant Given that our personal goal is to build human- progress in building human-level AI systems. level AI systems, we have struggled to find the Human-level AI systems are the ones that you right application for our research that requires dreamed about when you first heard of AI: HAL the breadth, depth, and flexibility of human- from 2001, A Space Odyssey; DATA from Star Trek; level intelligence. In 1991, we found computer- or CP30 and R2D2 from Star Wars. They are generated forces for large-scale distributed sim- smart enough to be both triumphant heroes ulations as a potential application. Effective and devious villains. They seamlessly integrate military training requires a complete battle all the human-level capabilities: real-time space with tens if not hundreds or thousands of response, robustness, autonomous intelligent participants. The real world is too expensive interaction with their environment, planning, and dangerous to use for continual training,

and even simulation is prohibitively expensive port characters that act just like humans. The and cumbersome when fully manned with AI characters can be part of the continual evo- humans. The training of 4 pilots to fly an attack lution in the game industry toward more real- mission can require over 20 planes plus air con- istic gaming environments. Increasing realism trollers. The military does not even have a facil- in the graphic presentation of the virtual ity with 20 manned simulators, and if it did, the worlds has fueled this evolution. Human-level cost in personnel time for the other pilots and AI can expand the types of experiences people support personnel to train these four pilots have playing computer games by introducing would be astronomical. To bypass these costs, synthetic intelligent characters with their own computer-generated forces are being developed goals, knowledge, and capabilities. Human-lev- to populate these simulations. These forces el AI can also recreate the experience of playing must integrate many of the capabilities we asso- with and against humans without a network ciate with human behavior—after all, they are connection. Current players of computer simulating human pilots. For example, they games are driven to networked games because must use realistic models of multiple sensing of the failings of the computer characters. In modalities, encode and use large bodies of massively multiplayer online games, human- knowledge (military doctrine and tactics), per- level AIs can populate the worlds with persis- form their missions autonomously, coordinate tent characters that can play the game along- their behavior, react quickly to changes in the side humans, providing opportunities for Human-level environment, and dynamically replan mis- interesting interactions that guide players in sions. Together with researchers at the Univer- the game and enhance the social dynamics AI systems are sity Southern California Information Sciences between players. Our hypothesis is that popu- the ones that Institute and Carnegie Mellon University, we lating these games with realistic, human-level you dreamed set off to build human-level AIs for military air characters will lead to fun, challenging games missions (Tambe et al. 1995). In 1997, we suc- with great game play. about when cessfully demonstrated fully autonomous simu- From the AI researcher perspective, the you first lated aircraft (Jones et al. 1999), and research increasing realism in computer games makes and development continues on these systems them an attractive alternative to both robotics heard of AI: by Soar Technology, Inc. Although computer- in the real world and homegrown simulations. HAL from generated forces are a good starting application By working in simulation, researchers interest- for developing human-level AI, there are ed in human-level AI can concentrate on cog- 2001, A Space extremely high costs for AI researchers to par- nitive capabilities and finesse many of the Odyssey; ticipate in this work. It requires a substantial pesky issues of using real sensor and real motor DATA from investment in time and money to work with systems; they must still include some sensor the simulation environments and to learn the modeling to get realistic behavior, but they Star Trek; or extensive background knowledge, doctrine, tac- don’t have to have a team of vision researchers CP30 and tics, and missions. Furthermore, much of the on their staff. They can pursue AI research in current funding is for building and fielding sys- worlds that are becoming increasingly realistic R2D2 from tems and not for conducting research. simulations of physical and social interactions, Star Wars. without having to create these worlds themselves. Computer games are cheap ($49.95), Computer Games reliable, and sometimes surprisingly accessible, In late 1997, we started to look for another with built-in AI interfaces. Moreover, computer application area, one where we could use what games avoid many of the criticisms often lev- we learned from computer-generated forces eled against simulations. They are real products and pursue further research on human-level and real environments on their own that mil- intelligence. We think we have found it in lions of humans vigorously interact with and interactive computer games. The games we are become immersed in. Finally, unlike military talking about are not Chess, Checkers, Bridge, simulations, we do not need to hunt out Othello, or Go, which emphasize only a few experts on these games; they surround us. human capabilities such as search and decision Another reason for AI researchers to work in making. The types of game we are talking computer games is that if we don’t start work- about use the computer to create virtual worlds ing in this area, the computer game industry and characters for people to dynamically inter- will push ahead without us (Woodcock 2000). act with—games such as Doom, Quake, Tomb Already there are at least five AI Ph.D.s working Raider, Starcraft, Myth, Madden Football, Diablo, in the industry (Takahashi 2000). AI researchers Everquest, and Asheron’s Call. have the opportunity to team with an aggres- Human-level AI can have an impact on these sive, talented, and caffeine-charged industry in games by creating enemies, partners, and sup- the pursuit of human-level AI. Here is a list of

16 AI MAGAZINE Articles reasons for AI researchers to take the computer puter games research,1 and some of the biggest game industry seriously (Laird 2000a). computer game companies (for example, Elec- First, computer game developers are starting tronic Arts in England and Sony) have started to recognize the need for human-level AI. Syn- research centers that include research in AI. thetic human-level characters are playing an More funding could become available as more increasingly important role in many genres of game developers discover they need help with computer games and have the potential to lead the AI in their products to push for a competi- to completely new genres. tive advantage. Much of the research could get Second, the computer game industry is high- done in nontraditional ways, with the involve- ly competitive, and a strong component of this ment of undergraduates, game developers, and competition is technology. AI is often men- game players. Thus, we can move AI research tioned as the next technology that will im- out of the labs and into the hands of millions. prove games and determine which games are hits. Thousands of new computer games are written every year with overall development Related Research time averaging nine months to two years, so on Computer Games technological advances sweep through the Other researchers have argued that great game industry quickly. Already, many computer play comes from “believable” agents. These games are marketed based on the quality of agents don’t necessarily have to be human lev- their AI. This field is one in which AI will have el in their intelligence, as long as they have a a significant impact. façade of intelligence supported by great per- Third, game developers are technologically … sonality. Joe Bates’s (1992) OZ research group at savvy, and they work hard to stay current with computer Carnegie Mellon University and Barbara Hayes- technology. AI programmer is already a com- Roth’s group at Stanford University (Hayes- game mon job title on game development teams. Roth and Doyle 1998) have worked on devel- Fourth, the game industry is big. In terms of hardware is oping believable agents for interactive fiction gross revenue, the computer game industry is and related computer games. Their research going to bigger than the movie industry (Croal and Toti- emphasized personality, AI agent-to-human lo 1999). provide cheap, interaction, and shallow but broad agents. We Fifth, computer game hardware is going to think these aspects are important but want to high-end provide cheap, high-end computation power emphasize that computer games provide an computation for AI in computer games in the next five years. arena for attempting to also build knowledge- The newest PC 3D video boards and the next- rich, complete, integrated AI that incorporates power for AI generation consoles, such as Sony’s Playstation many “deep” capabilities. in computer 2 and Microsoft’s X-box, move the entire John McCarthy has also argued that interac- graphics pipeline off the increasingly powerful tive computer games should be considered as a games in the central processing unit, freeing it for AI. It is topic of study for AI, where we can study how next five not at all unthinkable that in five years, there an AI system could play a game (his example is years. will be dedicated AI processors in game con- LEMMINGS, JR., a real-time scheduling and soles—we just have to tell them what we need. resource-allocation game) and solve problems Sixth, computer games need help from aca- that a human attempts.2 Other researchers demic AI. The current emphasis in computer have used other computer games such as Pengi game AI is on the illusion of humanlike behav- (Agre and Chapman 1987) and SIMCITY (Fas- ior for limited situations. Thus, most, if not all, ciano 1996). Our extension is to propose of the current techniques that are used for con- research on the AI characters that are part of trolling game AIs (such as big C functions or the game. Clearly, these efforts are related finite-state machines) will not scale up. Howev- because human-level AI characters often er, just as computer game graphics and physics require the skills of human players. One advan- have moved to more and more realistic model- tage of creating game characters is that we can ing of the physical world, we expect that game influence how games are made and played. The developers will be forced into more and more current emphasis on violence in computer realistic modeling of human characters. More- games is partially owed to the inability of the over, as researchers, we can get a step ahead of AI in these games to support more interesting the game designers by using their environ- social interactions. Although computer games ments for research on human-level AI. will probably always include violence, human- One thing that is missing in the computer level AI in games will give the game designers game field is significant research funding. freedom to explore other forms of player-char- Some of the military funding to support com- acter interaction. puter-generated forces is spilling over to com- RoboCup (Asada et al. 2000) is another relat-

SUMMER 2001 17 Articles

Game Genres AI Entity Roles AI Research Problems AI Research Areas

Action Tactical enemies Interact with environment High-level perception Role playing Partners Fast response Commonsense reasoning Adventure Support characters Realistic sensing Natural language Strategy games Story directors Adapt to environment Speech processing God games Strategic opponents Interact with humans Gesture processing Team sports Units Adapt to human player Planning & counterplanning Individual sports Commentators Difficulty Cognitive modeling adaptation Plan recognition Strategic adaptation Soft real-time response Interact with other AIs Reactive behavior Coordinate behavior Teamwork Navigation Scheduling Use tactics and strategies Path planning Allocate resources Spatial reasoning Understand game flow Temporal reasoning Humanlike responses Opponent modeling Reaction times Learning Realistic movement Knowledge acquisition Emotions Personalities Low computational overhead Low development overhead

Figure 1. AI Roles in Game Genres with Illustrative Links to Their Associated Research Problems.

ed project where competitors develop AI sys- Finally, we review the areas of AI that are tems to defeat other AI systems in both real applicable to these problems. This information robotic and simulated soccer games. In is collected together in ﬁgure 1. RoboCup, the goal is to build the best soccer- Although we list specific genres, the genres playing robots, not to create the best game play are fuzzy concepts, with many games being or humanlike behavior. RoboCup is stimulat- hybrids, incorporating components of multiple ing the development of integrated systems but genres. For example, there are strategy games none with the variety of capabilities we expect (Dungeon Keeper) that allow the human to to see in interactive computer games. “jump in the body” of one of their units and play as if it is an action game for a while. Also, there are action games where you must also Computer Game Genres manage resources and multiple units (such as In this section, we review the major genres of Battlezone). Although there will be a continual computer games to which human-level AI is blurring of the genres, the basic roles for AI stay relevant. There are other game genres, such as the same: enemies, partners, support characters, hunting games, fishing games, and lifelike strategic opponents, units and commentators. creatures games (Stern 1999), where deer-level, Action Games ﬁsh-level, or dog-level AI is necessary. For each of the genres in this section, we discuss the dif- Shortly after landing on an alien surface, ferent roles that human-level AI can play: ene- you learn that hundreds of your men have mies, partners, support characters, strategic been reduced to just a few. Now you must opponents, low-level units, and commenta- fight your way through heavily fortified tors. Other roles are possible, but these are the military installations, lower the city’s most common. In the following sections, we defenses, and shut down the enemy’s war go through these roles and discuss how AI machine. could improve the games and how these games —Quake II provide research problems for human-level AI. Action games involve the human player con-

18 AI MAGAZINE Articles

Figure 2. A Screenshot from the Popular Action Game Half-Life from Valve. (Reproduced with permission from Valve Software.) trolling a character in a virtual environment, Role-Playing Games usually running around and using deadly force Immerse yourself in a…world, where to save the world from the forces of evil. These nations hang in the balance of your games vary in the perspective that the human actions, dark prophecies test your resolve, has of his/her character, be it first person, where and heroic dreams can be fulfilled at last. the human sees what the character would see, —Baldur’s Gate or third person, where the player looks over the In role-playing games, a human can play differ- shoulder of the character. Popular examples ent types of character, such as a warrior, a magi- include Doom, Quake, Descent, Half-Life (figure cian, or a thief. The player goes on quests, col- lects and sells items, fights monsters, and 2), Unreal, and Tomb Raider. In pure action expands the capabilities of the character (such games, AI is used to control the enemies, which as strength, magic, quickness), all in an extend- are invariably alien monsters or mythical crea- ed virtual world. Example games include Bal- tures. Realism in graphics has been the point of dur’s Gate, Diablo, and Ultima. Recently, mas- competition for these games; however, the sively multiplayer role-playing games have graphics race seems to have run its course, with been created where thousands of people play better AI becoming the point of comparison. and interact in the same game world: Ultima Online, Everquest, and Asheron’s Call. In both Recent games, such as Rainbow Six, have types of role-playing game, AI is used to control extended the genre so that the human player enemies, as with action games; partners who can be part of a team, which includes either travel and adventure with the players; and sup- human or AI partners. porting characters, such as shopkeepers. The

SUMMER 2001 19 Articles

Figure 3. A Screenshot from the Upcoming Adventure Game Rubu Tribe by Outrage Entertainment. (Reproduced with permission from Outrage Entertainment.)

massively multiplayer games provide an addi- games include the Infocom series, King’s Quest, tional opportunity to use AI to expand and and many games from Lucas Arts, such as Full enhance the player-to-player social interac- Throttle, Monkey Island, and Grim Fandango as tions, perhaps with AI-controlled kings who well as Rubu Tribe (figure 3) from Outrage. AI war by sending player-controlled knights to can be used to create realistic supporting goal- battle each other. driven characters that the player must interact with appropriately to further their progress in Adventure Games the game. One of the Holy Grails of interactive Aye, ‘tis a rollicking piratey adventure ﬁction is to have a computer director who can that’s sure to challenge the mind and shiv- dynamically adjust the story and plot based on er a few timbers! the actions of the human. The majority of —The Curse of Monkey Island games have ﬁxed scripts and use many tricks to Adventure games, and the related genre of force the human player through essentially lin- interactive fiction, move further from action ear stories. However, a few games, such as Blade games as they deemphasize armed combat and Runner, have incorporated some autonomy and emphasize story, plot, and puzzle solving. In dynamic scripting into their characters and these games, players must solve puzzles and story line (Castle 1998). interact with other characters as they progress Strategy Games through an unfolding adventure that is deter- mined in part by their actions. Early adventure Players must successfully construct and games, such as Adventure and Zork, were totally rule their medieval empire while they text based, but more recent games sport 3D engage in real-time tactical warfare over graphics (sometimes using the graphics land, sea, and air. engines developed for action games). Example —Warcraft

20 AI MAGAZINE Articles

In strategy games, the human controls many individual players. Usually, the human con- units (usually military units, such as tanks, or trols one key player, such as the quarterback, the ever-present alien war machines) to do bat- while the computer controls all the other tle from a god’s eye view against one or more members of the team (figure 5). A second role opponents. Strategy games include reenact- is the strategic opponent, which, in this case, is ments of different types of battle: historical the opposing coach. One unique aspect of (Close Combat, Age of Empires), alternative real- team sport games is that they also have a role ity (Command and Conquer), fictional future for a commentator, who gives the play-by-play (Starcraft), and mythical (Warcraft, Myth). The and color commentary of the game (Frank human is often faced with problems of allocat- 1999). ing resources, scheduling production, and organizing defenses and attacks (Davis 1999). Individual Sports AI is used in two roles: (1) as a control for the Rip up the course on inline skates, speed detailed behavior of individual units that the on the street luge, pull serious air on the human commands and (2) as a strategic oppo- skateboard, and shred courses on the nent that must play against the human. The AI mountain bike. needs of the individual units differ from the —ESPN Extreme Games enemies and partners of action and role-play- For individual competitive sports, such as dri- ing games because they are not meant to be ving, flying, skiing, and snowboarding, the autonomous but are meant to be good soldiers computer provides a simulation of the sport who “follow orders.” from a first- or third-person perspective. The God Games human player controls a participant in the Building game who competes against other human or human-level You’re in charge of creating an entire city computer players. The computer player is more from the ground up—and the sky’s the like an enemy in an action game than a strate- enemies for limit. gic opponent or unit from a strategy game these games because the game is usually a tactical, real-time —SimCity 3000 requires God games give the player godlike control over competition. Individual sports can also require a simulated world. The human can modify the commentators. solving many environment and, to some extent, its inhabi- Roles general AI tants. The entertainment comes by observing the effects of his/her actions on individuals, soci- In each of the game genres described earlier, AI problems and is used in a variety of different roles to populate ety, and the world. SimCity is the classic example the game environment. integrating of a god game where the human acts as mayor, and the AI controls individual units or citizens of Tactical Enemies the solutions the simulated city. The Sims is probably the most In early games, the tactics of the computer- into coherent intriguing example (figure 4). The player creates controlled enemies were generally limited to individual characters (units) that have signifi- running directly at the player. Later enemies systems. cant autonomy, with their own drives, goals, were scripted or controlled by simple finite- and strategies for satisfying these goals, but God state machines. In these early games, the ene- (the human player) can come in and stir things mies were made more challenging, not with up by managing both the individual characters improved intelligence but with bigger guns, and their environment. tougher hides, and superior numbers. They also usually “cheated” by being able to see Team Sports through walls or out the back of their heads. More recently, games such as Half Life (Birdwell Welcome to Madden NFL 97, the game 1999), Descent 3, Quake III,3 and Unreal Tourna- that captures the excitement of a 30-yard ment have incorporated path planning and touchdown pass, the strategy of a well- many tactics that make these enemies more executed scoring drive, and the atmos- humanlike. Our own research (Laird 2000b; phere of a crisp autumn afternoon in the Laird and van Lent 1999) has concentrated on stadium. building enemies for Quake II that have the —Madden NFL 97 same strengths and weaknesses as human play- Team sports games have the human play a ers. To beat them, you have to outthink them combination of coach and player in popular as much as you have to outshoot them. Our sports, such as football (Whatley 1999), basket- SOAR QUAKEBOT is essentially a real-time expert ball, soccer, baseball, and hockey. AI is used in system that has multiple goals and extensive two roles that are similar to the roles in strategy tactics and knowledge of the game. It is built games: The first is unit-level control of all the within the SOAR architecture and has over 800

SUMMER 2001 21 Articles

Figure 4. AI Characters, Called Sims, in the God Game The Sims. (Reproduced with permission, courtesy Electronic Arts.)

rules. While it explores a level, it creates an reasoning. As they advance, they will also need internal model of its world and uses this model models of high-level vision that have the same in its tactics to collect nearby weapons and strengths and weaknesses as humans. One com- health, track down an enemy, and set ambush- mon complaint among game players is that the es. It also tries to anticipate the actions of enemy AI is cheating, which destroys the game- human players by putting itself in their shoes playing experience. For example, if the human (creating an internal model of their situation is in a dark room, the AI would be cheating if it garnered from its perception of the player) and could easily sense, identify, and locate the projecting what it would do if it were the human. However, if the human is backlit by a human player. bright hall, the AI enemy should be able to eas- Building human-level enemies for these ily sense and locate the human but possibly not games requires solving many general AI prob- identify him/her. This element is important for lems and integrating the solutions into coher- game play so that the same tactics and behav- ent systems. The enemies must be autonomous. iors that work well with humans work well with They must interact with complex dynamic AI enemies. environments, which requires reactive behav- There are many other applications of AI to ior, integrated planning, and commonsense building intelligent enemies. Because of the

22 AI MAGAZINE Articles

Figure 5. A Team Composed of AI-Controlled Players in Madden NFL™ 2000 Football. (Reproduced with permission from Electronic Arts Inc.) extended geography of the environment, they tion and natural language processing and even must navigate and use path planning, spatial gesture recognition. The partner AI must coor- reasoning, and temporal reasoning. As the dinate its behavior, understand teamwork, games become more complex, the enemies will model the goals of the human, and adapt to need to plan, counterplan, and adapt to the his/her style. Building such partners can build strategies and tactics of their enemies, using on previous research in AI in these areas but plan recognition, opponent modeling tech- within the context of all the other cognitive niques, and learning. Their responses need to activities involved in playing the game. be within the range of humans in terms of Support Characters reaction times and realistic movement. One Support characters are usually some of the least can even imagine adding basic models of emo- sophisticated AI characters in computer games, tions, where the enemies get “mad,” get “frus- but they have the most promise to improve trated,” and change their behavior as a result. games and are the most interesting in terms of Partners developing human-level AI. They currently Creating AI-controlled partners involves many have sets of canned responses that they spit of the same research issues as tactical enemies. back to the user based either on menu-selected However, enemy AI systems emphasize auton- questions or keywords. The most complex omy, and partners emphasize effortless cooper- ones, such as in Blade Runner (Castle 1998) ation and coordination between the human have some autonomy and some simple goals, player and the AI partner. Current games but they are extremely narrow goals with lim- restrict the human to using speciﬁc commands ited sets of behaviors for achieving these goals. to interact with partners, such as defend, Adding other AI-controlled support charac- attack, and follow me—commands much more ters could help populate the games with inter- limited than used in human-to-human interac- esting opportunities for interaction that guide tions. In the extreme, the need for full human- the player along various plot lines. Because to-human interaction brings in speech recogni- these characters need to exist in a virtual world

SUMMER 2001 23 Articles and generally play a human role in this source allocation involves scheduling commentator is to create a natural lan- world, they provide a useful first step production and temporal reasoning guage description of the ongoing toward human-level AI. In this role, about when the resulting units will be action in the game. The description support characters must interact with, available. The strategic opponent must can include both the moment-to- and adapt to, the environment; inter- also issue commands to the newly cre- moment action as well as key tactical act with, and adapt to, human players ated individual units, causing them to and strategy events that can require and other support characters; and pro- carry out the battle plan. Controlling a complex plan recognition and a deep vide humanlike responses, possibly large force of units with only a single understanding of the game. including natural language under- mouse is a significant part of the chal- Resource and Development Issues standing and generation. Because of lenge for human players. Because of A constant issue for game developers is these requirements and because these the human’s limited input ability, the the need to meet the limited computa- support characters are most directly strategic opponent must enforce tional power, in both memory and playing the role of embodied virtual humanlike limitations, such as reac- processing power, available in the humans, they require a wide range of tion times and realistic movements, average home computer or video game integrated AI capabilities, including when issuing commands to make the console. These resource issues can be everything from natural language to battle fair. finessed within the academic research path planning to teamwork to realistic Units community when the goal is just to do movement. In strategy games, god games, and research on human-level AI indepen- Strategic Opponents team sports games, AI is used to con- dent of the commercial applications. When creating strategic opponents for trol individual units. Generally, these However, we encourage researchers to strategy games and team sports games, units are given high-level commands take resource issues seriously because most game developers have had to from either the human player or the the more accessible our research is, the resort to “cheating” to make the oppo- strategic opponent and need to carry more likely it is that game developers nent challenging. Often, strategic out these commands. Units are usually and other industries will understand opponents are given extra units or controlled by finite-state machines (or the need for research on human-level resources or additional information large C functions) that are augmented AI and AI techniques in general. Our about the map or the human player’s with special routines for path plan- experience with the SOAR QUAKEBOT has position, or they play the game by a ning (Cavazza 1999) and path follow- driven us to research on comparisons different set of rules. Even with these ing. In addition to following orders, of SOAR with other architectures (Bhat- advantages, most strategic opponents units often need some ability to act tacharyya and Laird 1999; Wallace and are predictable and easily beaten once autonomously. For example, a platoon Laird 1999) and the overall efficiency their weaknesses are found. Strategic of marines moving from one position of SOAR. The SOAR QUAKEBOT requires 3 opponents for team sports games face to another should not ignore an ene- megabytes of random-access memory an additional difficulty in that their my tank. Instead, they should auton- and 10 percent of the processing pow- style of play must match a real-world omously choose to attack if appropri- er of a 400-megahertz WINDOWS NT PEN- team about which the human players ate or else find a new path. This TIUM II. are likely to be knowledgeable. semiautonomous behavior involves An additional constraint is that The tasks a strategic opponent must commonsense reasoning and perhaps these AI systems must be developed at perform can be divided into two cate- coordination with other units. Be- moderate cost. A game company will gories: (1) allocating resources and (2) cause there can be hundreds of units not be able to spend more than one issuing unit-control commands. In- active in a game at one time, the issues person-year on development of the AI volved in both of these tasks is the of computational and memory over- for a game. We need to develop tech- development of a high-level strategy. head are particularly important for niques for quickly building and cus- Creating this strategy, which is where unit AI (Atkin, Westbrook, and Cohen tomizing human-level AI systems. current strategic opponents are weak- 1999). Research on software engineering, est, involves integrated planning, Commentators knowledge acquisition, and machine commonsense reasoning, spatial rea- The role of the commentator is to learning will definitely play a role. soning, and usually plan recognition observe the actions of the AI and the and counterplanning to react to the human and generate natural language Conclusion human’s attack. One of the most comments suitable to describe the important aspects of strategy creation action (Frank 1999). In the RoboCup From a researcher’s perspective, even if is the coordination of multiple types competition, there is a separate com- you are not interested in human-level of a unit into a cohesive strategy. Once petition for commentator agents (Bin- AI, computer games offer interesting the plan is decided, the strategic oppo- sted 1998). Although sports games, and challenging environments for nent must determine how to best use both team and individual, are the many, more isolated, research prob- limited resources (mined minerals or most obvious genres for commenta- lems in AI. We are most interested in substitute players on a team) to com- tors, they can also be found in some human-level AI and wish to leverage pose an attack force appropriate to action games, such as Unreal Tourna- computer games to rally support for implement the battle plan. This re- ment. The obvious challenge for a research in human-level AI. One

24 AI MAGAZINE Articles attractive aspect of working in com- Bhattacharyya, S., and Laird, J. E. 1999. cal Report SS-00-02. Menlo Park, Calif.: puter games is that there is no need to Lessons for Empirical AI in Plan Execution. AAAI Press. attempt a “Manhattan Project” ap- Paper presented at the Sixteenth Interna- Laird, J. E., and van Lent, M. 1999. Devel- proach with a monolithic project that tional Conference on Artificial Intelligence oping an Artificial Intelligence Engine. (IJCAI-99) Workshop on Empirical AI, 31 attempts to create human-level intelli- Paper presented at the Game Developers’ July–6 August, Stockholm, Sweden. gence all at once. Computer games Conference, 15–19 March, San Jose, Cali- Binsted, K. 1998. Character Design for Soc- fornia. provide an environment for continu- cer Commentary. Paper presented at the al, steady advancement and a series of Stern, A. 1999. AI Beyond Computer RoboCup Workshop, 2–3, 9 July, Paris, Games. In Papers from the AAAI 1999 Spring increasingly difficult challenges. Just France. Symposium on Artificial Intelligence and Com- as computers have inexorably gotten Birdwell, K. 1999. The CABAL: Valve’s puter Games, 77–80. Technical Report SS-99- faster, computer game environments Design Processing for Creating Half-Life. 02. Menlo Park, Calif.: AAAI Press. are becoming more and more realistic Game Developer 6(12): 40–50. Takahashi, D. 2000. Artificial Intelligence worlds, requiring more and more com- Brooks, R. A.; Breazeal, C.; Marjanovic, M.; Gurus Win Tech-Game Jobs. The Wall Street plex behavior from their characters. Scassellati, B.; and Williamson, M. 2000. Journal, March 30, 2000, B14. Now is the time for AI researchers to The Cog Project: Building a Humanoid Tambe, M.; Johnson, W. L.; Jones, R. M.; jump in and ride the wave of comput- Robot. In Computation for Metaphors, Anal- Koss, F.; Laird, J. E.; Rosenbloom, P. S.; and er games. ogy, and Agents, ed. G. Nehaniv, 52–87. Lec- Schwamb, K. 1995. Intelligent Agents for ture Notes on Artificial Intelligence 1562. Interactive Simulation Environments. AI Acknowledgments Berlin: Springer-Verlag. Magazine 16(1): 15–39. The authors are indebted to the many Castle, L. 1998. The Making of Blade Run- Wallace, S., and Laird, J. E. 1999. Toward a students and staff who have worked ner, Soup to Nuts! Paper presented at the Methodology for AI Architecture Evalua- Computer Game Developers’ Conference, tion. In Intelligent Agents VI, eds. N. R. Jen- on the SOAR/Games project, most 4–8 May, Long Beach, California. nings and Y. Lespérance, 117–132. Berlin: notably Steve Houchard, Karen Coul- Cavazza, M.; Bandi, S.; and Palmer, I. 1999. Springer-Verlag. ter, Mazin Assanie, Josh Buchman, Joe “Situated AI” in Video Games: Integrating Whatley, D. 1999. Designing around Pit- Hartford, Ben Houchard, Damion NLP, Path Planning, and 3D Animation. In falls of Game AI. Paper presented at the Neff, Kurt Steinkraus, Russ Tedrake, Papers from the AAAI 1999 Spring Symposium Game Developers’ Conference, 15–19 and Amy Unger. on Artificial Intelligence and Computer Games, March, San Jose, California. Notes 6–12. Technical Report SS-99-02. Menlo Woodcook, S. 2000. Game AI: The State of 1. See, for examples, www.stomped.com/ Park, Calif.: AAAI Press. the Industry. Game Developer 7(8): 24–32. published/jcal979197050_1_1.html. Croal, N., and Totilo, S. 1999. Who’s Got 2. John McCarthy, 1998, Partial Formaliza- Game? Newsweek 134(11): 46. John Laird is a professor tions and the Lemmings Game. www- Davis, I. 1999. Strategies for Strategy Game of electrical engineering formal.stanford.edu/jmc/ lemmings.html. AI. In Papers from the AAAI 1999 Spring Sym- and computer science at 3. G. Keighley, 1999, The Final Hours of posium on Artificial Intelligence and Computer the University of Michi- Quake III Arena: Behind Closed Doors at id Games, 24–27. Technical Report SS-99-02. gan. He received his Software, GameSpot. Available at www. Menlo Park, Calif.: AAAI Press. Ph.D. in computer sci- gamespot.com/features/btg-q3/index.html. Fasciano, M. J. 1996. Real-Time Case-Based ence from Carnegie Mel- Reasoning in a Complex World. Technical lon University in 1983. Report, TR-96-05, Computer Science De- He is a fellow of the References partment, University of Chicago. American Association for Artificial Intelli- Agre, P. E., and Chapman, D. 1987. Pengi: Frank, I. 1999. Explanations Count. In gence and a founder of Soar Technology, An Implementation of a Theory of Activity. Papers from the AAAI 1999 Spring Symposium Inc. His e-mail address is [email protected]. In Proceedings of the Sixth National Con- on Artificial Intelligence and Computer Games, ference on Artificial Intelligence (AAAI-87), 77–80. Technical Report SS-99-02. Menlo Michael van Lent is a 268–272. Menlo Park, Calif.: American Park, Calif.: AAAI Press. postdoctoral fellow at Association for Artificial Intelligence. Hayes-Roth, B., and Doyle, P. 1998. Ani- the University of Michi- Asada, M.; Veloso, M.; Tambe, M.; Noda, I.; mate Characters. Autonomous Agents and gan. He received his Kitano, H.; and Kraetzschmar, G. K. 2000. Multi-Agent Systems 1(1): 195–230. Ph.D. from the Universi- Overview of RoboCup-98. AI Magazine Jones, R. M.; Laird, J. E.; Nielsen, P. E.; Coul- ty of Michigan in 2000. 21(1): 9–19. ter, K. J.; Kenny, P. G.; and Koss, F. V. 1999. His research interests Atkin, M. S.; Westbrook, D. L.; and Cohen, Automated Intelligent Pilots for Combat include machine learn- P. R. 1999. Capture the Flag: Military Simu- Flight Simulation. AI Magazine 20(1): ing and AI in computer games. His e-mail address is vanlent@ lation Meets Computer Games. In Papers 27–42. umich.edu. from the AAAI 1999 Spring Symposium on Laird, J. E. 2000a. Bridging the Gap Artificial Intelligence and Computer Games, between Developers and Researchers. Game 1–5. Technical Report SS-99-02. Menlo Developer 7(8): 34. Park, Calif.: AAAI Press. Laird, J. E. 2000b. It Knows What You’re Bates, J. 1992. Virtual Reality, Art, and Going To Do: Adding Anticipation to a Entertainment. Presence: The Journal of Tele- QUAKEBOT. In Papers from the AAAI 2000 operators and Virtual Environments 1(1): Spring Symposium on Artificial Intelligence 133–138. and Interactive Entertainment, 41–50. Techni-

SUMMER 2001 25 Articles

Artiﬁcial Intelligence and Mobile Robots Case Studies of Successful Robot Systems

Edited by David Kortenkamp, R. Peter Bonasso, and Robin Murphy

he mobile robot systems described in this book were selected from among the best available implemen- tations by leading universities and research laboratories. These are robots that have left the lab and Tbeen tested in natural and unknown environments. They perform many different tasks, from giving tours to collecting trash. Many have distinguished themselves (usually with first or second-place finishes at various indoor and outdoor mobile robot competitions.

Each case study is self-contained and includes detailed descriptions of important algorithms, including pseudo-code. Thus this volume serves as a recipe book for the design of successful mobile robot applications. Common themes include navigation and mapping, computer vision, and architecture.

6 x 9, 400 pp., ISBN 0-262-61137-6

To order, call 800-356-0343 (US and Canada) or (617) 625-8569. Distributed by The MIT Press, 55 Hayward, Cambridge, MA 02142

26 AI MAGAZINE