Reports of the AAAI 2010 Conference Workshops

Reports

David W. Aha, Mark Boddy, Vadim Bulitko, Artur S. d’Avila Garcez, Prashant Doshi, Stefan Edelkamp, Christopher Geib, Piotr Gmytrasiewicz, Robert P. Goldman, Alon Halevy, Pascal Hitzler, Charles Isbell, Darsana Josyula, Leslie Pack Kaelbling, Kristian Kersting, Maithilee Kunda, Luis C. Lamb, Bhaskara Marthi, Keith McGreggor, Lilyana Mihalkova, Vivi Nastase, Sriraam Natarajan, Gregory Provan, Anita Raja, Ashwin Ram, Mark Riedl, Stuart Russell, Ashish Sabharwal, Jan-Georg Smaus, Gita Sukthankar, Karl Tuyls, and Ron van der Meyden

n The AAAI-10 workshop program was held AI and Fun Sunday and Monday, July 11–12, 2010, at the Westin Peachtree Plaza in Atlanta, Georgia. Interactive entertainment has become a dominant force in the The AAAI-10 workshop program included 13 entertainment sector of the global economy. In 2000, John Laird workshops covering a wide range of topics in and Michael van Lent justified interactive entertainment as a artificial intelligence. The titles of the work - domain of study in AI when they posited that computer games shops were AI and Fun; Bridging the Gap Between Task and Motion Planning; Collabo - could act as test beds for achieving human-level intelligence in ratively Built Knowledge Sources and Artificial computers, leveraging the fidelity of their simulations of real- Intelligence; Goal-Directed Autonomy; Intelli - world dynamics. There is an additional perspective on AI for gent Security; Interactive Decision Theory and games: increasing the engagement and enjoyment of the play - Game Theory; Metacognition for Robust Social er. This perspective is consistent with the perspective of com - Systems; Model Checking and Artificial Intelli - puter game developers. For them, AI is a tool in the arsenal of gence; Neural-Symbolic Learning and Reason - the game to be used in lieu of real people when no one is avail - ing; Plan, Activity, and Intent Recognition; Sta - able for a given role. Examples of such roles are opponents, tistical Relational AI; Visual Representations and Reasoning; and Abstraction, Reformula - companions, and nonplay characters (NPCs) in roles that are tion, and Approximation. This article presents not fun to play such as shopkeepers, farmers, and victims; cin - short summaries of those events. ematographer; dungeon master; plot writer; or game designer.

As we move down this list, the computational sys - While we are not ready to formally define the tem is charged with progressively more responsi - term fun , we can—and should—use it as a call to bility for providing a user with a fun experience. arms for an investigation into core research on But what is fun? We seem to know it when we intelligent systems that reason about and manage see it, but fun is also highly subjective. Can we the quality of human experiences both in a variety computationally model fun? Can intelligent sys - of domains—including many beyond games, such tems learn and utilize models of fun, player prefer - as education, training, and health—and in a vari - ences, storytelling, and so on, to affect human ety of computer-mediated experiences such as sto - experiences? If so, what would this enable with rytelling, interactive drama and theater, serious respect to increased engagement, enjoyment, or games, mixed-reality environments and virtual new forms of computer-mediated interactive expe - cinematography. While fun can be subjective, riences? What are the potential ways forward? To progress can be made through study of related, begin to address these questions, we invited 11 objectively measurable phenomena: engagement, research groups to present on their work and to enjoyment, immersion, flow, replayability, moti - help shed light on the questions from a number of vation, and others yet to be identified. Finally, we perspectives. We explicitly did not ask for papers, note the strong potential for societal impact but instead asked that each group present on their through the domains that can be affected as well as perspective on AI approaches to fun, to address the insight into the basic questions of what drives us questions above, and to speculate on the future. and how we can computationally model and auto - The presentations roughly clustered into four matically reason about it. themes: (1) Player modeling and learning from Mark Riedl, Charles Isbell, Ashwin Ram, and humans, featuring presentations on how to learn Vadim Bulitko served as cochairs of this workshop. models of players, customize game play experi - No technical report was issued. ence, and make inferences about what humans like to do. (2) Virtual and real humans, featuring pre - Bridging the Gap Between sentations on virtual humans in mixed-reality game environments, modeling human improvisa - Task and Motion Planning tional actors for the purposes of building better Task-level planning of the sort typically studied by interactive experiences, and agents that express the AI and planning communities has historically curiosity. (3) Storytelling and discourse, featuring been quite separate from motion planning as stud - presentations on the question of whether we can ied by roboticists. There is an increasing belief that achieve the dream of the Holodeck, how interac - it is both necessary and useful for a tighter cou - tive storytelling might reinforce cognitive percep - pling between these levels. From the perspective of tion of engagement, and how computational sys - task planning, motion planning can be viewed as tems can mediate virtual experiences through providing geometric constraints and heuristics, automated cinematography. (4) Making learning and therefore information about the feasibility and AI fun, featuring presentations on the question of costs of higher-level actions. From a motion plan - how to engage learners in AI courses with virtual ning perspective, task planners allow taking advan - worlds and robotics. tage of the rich combinatorial structure that exists In addition to the four themes, we invited three in the configuration spaces that arise when, for experts from a disparate set of industries to talk example, manipulating several objects. about challenges and opportunities for AI and fun: The workshop featured talks from diverse areas. Miguel Encarnacao, director of emerging technol - There were several presentations on particular for - ogy innovation at Humana, Inc.; Joe Marks, vice malisms for combining the levels of planning from president of Disney Research; and Bob Sottilare, the perspectives of classical planning, motion chief technology officer of the U.S. Army Simula - planning, POMDPs, search-based planning, and tion and Training Technology Center. While it temporal logic-based constraints. There were also would seem that there is little in common with presentations on robotic applications and open training warfighters, making people healthy, and source software, as well as fast GPU implementa - entertaining people in theme parks, a consensus tions of motion planners. emerged that fun is important for motivating peo - Several common themes emerged from the pre - ple to learn, proactively maintain their health, and sentations and discussion. There was a general create brand loyalty. There was also a call for more consensus that pure low-level motion planning in research into natural forms of human-computer configuration space was fairly well solved for prob - communication. Most notably, all three domains lems of moderate dimensionality, and that a prin - cited scalability—more people, more personaliza - cipled approach to coupling with higher-level con - tion, longer experience durations, longitudinal straints and goals was the next step. There were a interactions—as a primary bottleneck that required variety of opinions, however, on the form that this creative automated intelligence. coupling should take, ranging from hierarchical

96 AI MAGAZINE Reports approaches that tackle the different levels in con - shop, the presented papers address a diverse set of cert, to approaches based on finding a good dis - problems and resources. Some papers explored the cretization or summary of lower-level state, but process and result of combining different then planning independently. Finally, many par - resources, extracting and formalizing knowledge ticipants stressed the need for common bench - from structured, semistructured or unstructured marks and representations, so that the different sources, harnessing people power and learning approaches could be compared. from them how to automatically generate and The workshop was organized by Maxim enhance knowledge repositories. The two invited Likhachev, Bhaskara Marthi, Conor McGann, and talks explored in more detail two of the workshop David E. Smith. The papers of the workshop were themes. Henry Lieberman (Massachusetts Institute published as AAAI Technical Report WS-10-01. of Technology Media Laboratory) presented work developing a platform for knowledge acquisition and usage from expert users, built upon experience Collaboratively Built Knowledge with eliciting information from nonexperts in a Sources and Artificial Intelligence collaborative manner, and Lenhart Schubert (Uni - versity of Rochester) explored the usage of knowl - Until recently, the AI, and in particular the natural edge extracted from general text for formal infer - language processing (NLP), communities have ence, and the impact of ambiguity and specific relied on resources built manually by experts in constructs in language on the knowledge extracted specific areas (such as linguists, philosophers, and and the inferencing methods. cognitive linguists). User-contributed knowledge has opened up a new perspective, in that it cap - Vivi Nastase, Roberto Navigli, and Fei Wu served tures the kind of knowledge and organization that as cochairs of this event. Supporting agencies arises naturally out of the consensus of the masses, included AAAI, and HITS, gGmbH, Heidelberg, and as such represents better our collective knowl - Germany. The papers of the workshop were pub - edge. The outcome is a multifaceted and extreme - lished as AAAI Technical Report WS-10-02. ly rich source of information, revealed through embedded annotations and structural informa - Goal-Directed Autonomy tion. The first such collaboratively developed reposi - The objective of the AAAI Workshop on Goal- tory of information to be extensively used in AI Directed Autonomy (GDA) was to encourage dis - and natural language processing was Wikipedia. Its cussion and novel contributions on intelligent usefulness was demonstrated through its contribu - agents that can self-select their goals. How should tions to a wide range of tasks: text categorization, an autonomous agent behave competently when clustering, word-sense disambiguation, informa - interacting in a complex environment (for exam - tion retrieval, information extraction, and ques - ple, partially observable, multiagent, with large tion answering. decision spaces, dynamic updates, stochastic out - In recent years, more and more resources and comes, and continuous effects)? The option of collaborative endeavours have started to be incor - complete a priori domain engineering is not porated and exploited as knowledge repositories appealing due to its high cost; it would require for various tasks. Tags associated with images in planning for all possible contingencies due to Flickr, question-answer collections in Yahoo! opportunities or plan execution failures. Alterna - Answers are a few examples of such information tively, the agent could be given the ability to sources. Amazon’s Mechanical Turk gives re - decide what goals it should pursue at any point in searchers access to “human computation” power, time, which would increase its level of autonomy and is being used more and more as a solution to by relaxing the assumption that its assigned goal is the difficult problems of large-scale evaluations the only one that it should pursue throughout its and data annotation, both crucial for the continu - lifetime. This capability of goal reasoning could ous development of the AI and NLP fields. dramatically affect the types of tasks that these The Collaboratively Built Knowledge Sources agents can perform. and Artificial Intelligence workshop took place on As demonstrated by the workshop’s attendees, July 11, 2010, in Atlanta, Georgia, immediately goal reasoning is of interest to researchers studying preceding the 24th AAAI Conference on Artificial cognitive architectures, game AI, multiagent sys - Intelligence. This workshop is a successor to the tems, planning, robotics, and other topics in Wikipedia and Artificial Intelligence: An Evolving which competent agent behavior is desirable in Synergy workshop organized at AAAI 2008 and the complex environments. To our knowledge, this User Contributed Knowledge and Artificial Intelli - was the first workshop that focused on goal rea - gence: An Evolving Synergy workshop organized as soning. We named the workshop GDA because it is part of IJCAI 2009. the name of a recent conceptual model for goal Consistent with the original aim of the work - reasoning whose components address problems

WINTER 2010 97 Reports

relevant to other such models. These include (1) experiences with the applicability of techniques detecting situations that may trigger goal reason - from one field to problems from the other, and to ing, (2) explaining why a detected situation identify the key issues to be addressed in increasing demands attention, (3) deciding how to respond to the convergence between security and AI. such situations (for example, through goal(s) for - This is a fertile area for research, and has been mulation), and (4) managing the current set of attracting an increasing amount of interest in both pending goals, which may involve tasks such as communities. Prior to this workshop there was a goal interruption, transformation, resumption, 2009 ICAPS workshop on the topic, as well as two and/or deletion. workshops held in conjunction with the ACM The workshop began with a survey on goal rea - Conference on Computer and Communications soning, given by Matthew Klenk (NRL). Felipe Security (CCS), and so organized primarily from Meneguzzi (Carnegie Mellon University) discussed the computer security community. work on motivations, and Matt Molineaux AI and security is a large and growing area, both (Knexus Research) discussed progress on a recently for research and for applications. Our increasingly evaluated system that builds on this foundation networked world continues to provide new oppor - for goal formulation. In the context of cognitive tunities for security breaches that have severe con - architectures and meta reasoning, Dongkyu Choi sequences at the personal level (identity theft, and (Stanford University) discussed how goal reasoning resulting financial losses), for businesses (theft of is embedded in the ICARUS, Randy Jones intellectual property, or business plans, or costly (SoarTech) described the use of appraisal theory for responses to the theft of customer data), and for implementing GDA in Soar, and Ashok Goel (Geor - governments. Computing and the Internet have gia Institute of Technology) described its relation become crucial parts of the infrastructure of almost to metareasoning. every significant commercial or governmental In a robotics context, Jeremy Baxter’s (QinetiQ) enterprise. Turning off the computers or discon - presentation focused on user interaction, while necting from the network has become tantamount Nick Hawes (University of Birmingham) described to turning off the power. a goal-directed reasoning framework for control - The use of techniques drawn from AI is increas - ling a mobile robot. Héctor Muñoz-Avila (Lehigh ingly relevant as the scale of the problem increas - University) and Ben Weber (University of Califor - es, in terms of the size and complexity of the net - nia, Santa Cruz) described different roles of case- works being protected, in terms of the variety of based reasoning for GDA, and Russell Knight sur - applications and services provided using that infra - veyed applications of CASPER, a real-time structure, and with the sophistication of the embedded planner scheduler. Finally, Mike Cox attacks being made. Filtering the faint signals of (DARPA) monitored a panel on the relation of goal intrusion from a flood of data related to normal reasoning to plan adaptation (Muñoz-Avila), operations can be viewed as data mining. Learning replanning (Ugur Kuter, University of Maryland), methods can be applied to generate classifiers for and planning and uncertainty (Daniel Bryce, Utah this process, or to detect the presence of new State University). means of attack. AI planning methods can be used There were many interruptions and lively dis - to generate compact representations of possible cussion on topics such as the approaches for attacks, which can then be used to deploy coun - designing the components of a GDA model (for termeasures. Plan and intent recognition are example, when, how, and what new goals should important areas of research as well and are the be formulated?), the relative benefits of alternative focus of a growing number of researchers. The models, how to manage concurrent goals, the rela - detection of anomalous operations or network traf - tion of goal reasoning to automated subgoaling fic can be viewed as a component of many securi - and opportunistic planning, and the ability of cur - ty functions, including both intrusion detection rent methods to scale. and plan recognition. Another recent topic is David Aha, Matthew Klenk, Héctor Muñoz-Avila, improving anomaly detection using the ubiqui - Ashwin Ram (Georgia Institute of Technology), and tous and increasingly powerful graphics processors Daniel Shapiro (ISLE) served as organizers of this in our computers. Because of the distributed nature workshop. The papers presented were not published. of computer networks, they are susceptible to attack that comes from multiple directions, which Intelligent Security can be mounted by an individual in a single loca - tion. Thus, the issue of information fusion (com - The purpose of the Intelligent Security workshop bining indications drawn from separate data series was to bring together researchers with an streams) is an important tool, as well. interest in both security and AI. The goals were to Mark Boddy, Stefan Edelkamp, and Robert P. tease out common themes and differences, identi - Godman served as cochairs of this AAAI workshop. fy common problems and their solutions, share The papers presented were not published.

98 AI MAGAZINE Reports

Interactive Decision Theory gaining, opponent modeling and Monte-Carlo search in poker, trust models applied in supply and Game Theory chain management, and a computational deci - The Interactive Decision Theory and Game Theo - sion-theoretic approach for interactive assistants. ry workshop is a continuation of a series of work - Still other papers were devoted to teamwork and shops on decision and game theories held over coordination under model uncertainty and to sig - previous years. These topics remain active research naling games as models for conversational ground - areas since game and decision theories proved to ing. Overall, we were impressed with the technical be powerful tools with which to design maturity and high level of formalism presented in autonomous agents and to understand interac - the papers, and the impressive selection of appli - tions in systems composed of many such agents. cations the researchers are looking at. Decision theory provides a general paradigm for Piotr Gmytrasiewicz, Prashant Doshi, and Karl designing agents that can operate in complex Tuyls served as cochairs of this event. The papers of uncertain environments and can act rationally to the workshop were published as AAAI Technical maximize their preferences. Decision-theoretic Report WS-10-03. models use precise mathematical formalism to define the properties of the agent’s environment, Metacognition for the agent’s sensory capabilities, the ways in which the agent’s actions change the state of the envi - Robust Social Systems ronment, and the agent’s goals and preferences. The one-day Workshop on Metacognition for Agent’s rationality is defined as behavior that max - Robust Social Systems was a sequel to a series of imizes the expectation of the degree to which the successful workshops on the topic of metareason - preferences are achieved over time, and the plan - ing beginning in 2007. The focus of this workshop ning problem is identified as a search for the was on design considerations, issues, and chal - rational, or optimal, plan. lenges in using metacognition to improve the Game theory adds to the decision-theoretic robustness of social systems that include purely framework the idea of multiple agents interacting artificial entities or both humans and software within a common environment. It provides ways agents. The workshop had both full paper and to specify how agents, separately or jointly, can four-page position papers, some of which built on change the environment and how the resulting results from the previous workshops—an encour - changes affect their individual preferences. Build - aging sign of the formation of a research commu - ing on the assumption that agents are rational and nity in this area. The papers were categorized self-interested, game theory uses the notion of under three themes: (1) metacognition in human- Nash equilibrium to design mechanisms and pro - machine social systems, which included discus - tocols for various forms of interaction and com - sions on modeling human behavior, reproducing munication that result in the overall system humanlike interactions, understanding and deal - behaving in a stable, efficient, and fair manner. ing with conflicts in human input, and the role of Recent research has sought to merge advances in modeling self and emotions in human-machine decision and game theories to build agents that social systems; (2) metacognition in multiagent may operate in complex uncertain environments systems, including domains suitable for decentral - shared with other agents. This research has inves - ized metacognition and a common platform for tigated the problems of Nash equilibrium as a solu - evaluating multiagent cognition to the implemen - tion concept, focused on epistemological advances tation of decentralized metacognition in multia - in game theory and expressive ways to model gent systems; and finally (3) metacognitive archi - agents. Alternative solution concepts have been tectures, a theme that included classification of investigated with the aim of designing autono - metacognition under higher-level, theory-based mous agents that robustly interact with other, metacognition and lower-level, experience-based highly sophisticated agents in both cooperative metacognition, metarepresentational theories of and noncooperative settings. metacognition applied for theory-based metacog - Papers presented at the workshop spanned the nition, control theories of metacognition applied spectrum of theoretical issues as well as emerging for experience-based metacognition and frame - application areas. There were papers on learning to works for implementing scalable metacognitive cooperate, computation of steady states in two- architectures. player extensive games, improved fast computa - In addition to the presentations, the sessions tion of Nash equilibria, and the maximum entropy were followed by an interactive panel where each approach to computing correlated equilibria. author presented views and comments regarding Papers that included applications were ones on topics put forth by a selected moderator. These ses - cognitive hierarchies applied to the lemonade sions also gave the audience an opportunity to ask game, updating higher order beliefs during bar - questions or make comments on issues that

WINTER 2010 99 Reports

spanned individual paper presentations. Some of range of modal logics, and there are now model the issues discussed were the complexities and checkers whose specification language is able to issues involved in modeling agents situated in a express modalities such as knowledge, belief, and social system that includes humans (versus all so on, as well as time. Such modalities are of par - machines); the role of emotions in social systems; ticular interest when dealing with autonomous the feasibility of having models of self and models and multiagent systems. of others; criteria for evaluating metacognitive sys - In principle, a model checker conducts an tems; deadline with overhead costs associated with exhaustive examination of the state space. Under - metareasoning; and finally modeling emotions to pinning the success of the area is a range of sophis - potentially help with metacognitive processing. ticated optimization techniques and heuristic algo - A special highlight of the workshop was the rithms that enable this computation to be invited talk. Ashok Goel of the Georgia Institute of performed efficiently. In this regard, model check - Technology spoke of model-based metareasoning ing has benefited from a range of ideas from artifi - for self-adaptation. He first presented the idea of cial intelligence, including search heuristics such proactive, goal-directed reconfiguration of reason - as A*, and planning approaches to counterexam - ing processes. He used an assembly task as a moti - ple construction. vating example to discuss how reasoning mecha - Several themes were touched by the papers pre - nisms of one task can be analogically transferred to sented at the workshop. Stefan Leue (University of other tasks. Professor Goel then discussed localized Konstanz) presented an algorithm for finding the k reinforcement learning in the context of free-civ, a shortest paths in graph, a problem that is relevant, turn-based strategy game and showed how model- among others, for stochastic model checking. Ste - based metareasoning would provide reward signals fan Edelkamp (University of Bremen) spoke on the for different parts of the search space. He conclud - use of the graphical processing unit (GPU) for ed the talk with his thoughts on retrospective, fail - external memory breadth-first search. Hector ure-driven repair of domain knowledge. This was Geffner (University of Barcelona) gave an invited followed by a lively question and answer session talk on planning with incomplete information, on the role of models in metareasoning. stressing that while logic and theory are needed in The cochairs of this workshop were Anita Raja planning, the bottom line is heavily empirical. (University of North Carolina at Charlotte) and Two further presentations were also devoted to Darsana Josyula (Bowie State University). The planning. Siddharth Srivastava (University of Mas - papers were published as AAAI Technical Report sachusetts) spoke on computing applicability con - WS-10-04. ditions for plans with loops, with various results conerning termination and other behaviours of Model Checking and transition systems applying not just to a particular planning formalism, and hence of interest to the Artificial Intelligence model checking community. Stefan Edelkamp Model checking is an approach to verification spoke on action planning for automated program based on representing some system as a model, a verification, using approaches from planning for semantic structure that supports the truth or falsi - verification of C code. ty of a formula of a logic. Typically, the model Multiagent systems and epistemic logic was describes all the possible behaviors of a system over another workshop theme, building a strong bridge time, and the logic is a modal logic that allows one between AI concepts and model checking. Abdur to specify some desired or undesired behaviors in a Rakib (University of Nottingham) spoke on bound - concise way. A model checker is a software system ed model checking for temporal epistemic logic. that takes as inputs representations of the system Francesco Russo (Imperial College, London) spoke and its specification in modal logic, and computes on automatic data abstraction in model checking whether the specification holds in the system. multiagent systems. Ron van der Meyden present - The interactions between model checking and ed an extension of CTL with epistemic operators. artificial intelligence are rich and diverse. Model Finally, Kaile Su (University of Beijing and Univer - checking originated in the 1980s as an approach to sity of Brisbane) gave a presentation on the Herbi - the verification of concurrent hardware processes vore protocol, which involves knowledge and and computer network communications protocols. anonymity. These days, model checking is applied by During the workshop discussion period, Stefan researchers working on computer software such as Edelkamp aimed at bridging the gap between plan - hardware device drivers and cryptographic proto - ning and model checking and raised the question cols. AI applications include planning, stochastic of whether the knowledge aspect is really essential process models, autonomous robots, and other for the applications just mentioned. forms of multiagent systems. This broadening of Not surprisingly, the workshop was geared more application area has also led to a broadening of the toward AI than model checking, and there was

100 AI MAGAZINE Reports some support of the idea that to reach into the model-checking community, the next workshops should move away from AI conferences for a change. The Feigenbaum Prize Ron van der Meyden and Jan-Georg Smaus served as cochairs of this workshop. The papers of this workshop were not published. The ﬁrst biennial AAAI Feigenbaum Prize will be awarded in 2011 at the 25th anniversary of the AAAI Conference in San Francisco, California. The AAAI Neural-Symbolic Learning Feigenbaum Prize recognizes and encourages out - and Reasoning standing Artiﬁcial Intelligence research advances that are made by using experimental methods of comput - AI faces huge challenges in its quest to develop tru - er science. The “laboratories” for the experimental ly intelligent systems. These systems are required work are real-world domains, and the power of the to learn and adapt to changes in the environment research results are demonstrated in those domains. they operate, and to reason about commonsense The Feigenbaum Prize may be given for a sustained knowledge in ways that can control the accumula - record of high-impact seminal contributions to exper - tion of errors. Current developments in the area of imental AI research; or it may be given to reward sin - neural-symbolic computation bring an opportuni - gular remarkable innovation and achievement in ty to integrate well-founded symbolic AI reasoning experimental AI research. and inference systems with robust neural comput - The prize is $10,000 and is provided by the Feigen - ing machinery and learning to help tackle some of baum Nii Foundation and administered by AAAI. these challenges. Neural-symbolic systems combine the statistical nature of learning and the logical nature of rea - For complete details about how to submit nomina - soning. Over the years, researchers have built tions for this prize, please see sound neural-symbolic models that are able to www.aaai.org/Awards/feigenbaum.php learn several forms of reasoning, including tempo - ral, modal, epistemic, fuzzy, intuitionistic, and relational (first-order, predicate) logics. In a nut - shell, neural-symbolic computation offers a methodology for integrating reasoning and learn - ing in intelligent systems and a rich model for cog - developments, implementations and challenges, nitive computation. These features allow the inte - including complexity and first-order learning grated study of symbolic and connectionist AI. issues. Further, neural-symbolic computation seeks to Leo de Penning presented a paper introducing provide explanations to certain important ques - tions in cognitive science, such as the nature of an integrated neural-symbolic cognitive agent reasoning, knowledge representation, and learn - architecture for training, assessment, and feedback ing, following the computational theory of mind. in simulators. De Penning illustrated the practical The workshop contained a mix of invited talks use of neural-symbolic computation in a virtual and submitted presentations. Invited talks instruction platform, reporting initial results on a spanned the foundations of (logical) reasoning and real simulator environment. The proposed archi - neural and statistical learning to the application of tecture combines temporal logic reasoning and neurosymbolic technology in the aerospace indus - learning in recurrent restricted Boltzmann try, training simulators, and vehicle control. machines. The workshop started with a keynote address by Sihle Wilson presented a paper that contains the philosopher Paul Thagard who presented exciting initial design of a neuro-fuzzy plug-in that would new ideas on how brains make mental models. allow vehicles autonomously to retrieve a driver Thagard brought the role of abductive reasoning to from a nearby location. The vehicle would use neu - the foreground and discussed the importance of ral networks to help it avoid collisions, but also attention and creativity to this reasoning task, fuzzy logic to help it make suboptimal decisions in with particular emphasis on how emotions can case a collision cannot be avoided. drive one’s attentional focus. Dragos Margineantu gave an invited talk report - Gadi Pinkas gave an invited talk on how to rep - ing on years of experience at Boeing and DARPA resent first-order logic in symmetric networks, par - on testing adaptive and learning models. Margin - ticularly Boltzmann machines. The effective inte - eantu presented a robust methodology for evaluat - gration of first-order logic and artificial neural ing learning systems. Particular emphasis was networks has been a challenge for decades. Pinkas placed on the evaluation of high-risk low-proba - revisited his proposed model and discussed recent bility events, the difficulties associated with small

WINTER 2010 101 Reports

or zero counts, and the challenges relating to high- shop brought together researchers from these cost anomaly detection. diverse areas to share ideas and developments. Pedro Domingos’s talk advocated the use of a The workshop included two invited talks. Hec - foundational approach to the research in the area. tor Geffner spoke on plan recognition as planning. Various methods used by the logical and statistical Tanzeem Choudhury spoke on activity recognition approaches to AI were reviewed, with emphasis on and social network identification using small Markov logic networks (MLNs) as a proposed syn - mobile platforms. Both invited speakers, along thesis toward a unified theory of learning and cog - with Gita Sukthankar, were panelists on a very nition. MLNs integrate probabilistic inference and interesting panel discussion of how to go about first-order logic in novel ways and have been learning the structures that are to be recognized. shown useful in a range of applications. As the speakers used very different models, this Kristian Kersting’s invited talk on statistical rela - resulted in some very interesting and diverse ideas tional mining discussed the exponential growth of and lively debate. Among the major themes of the scientific data and exciting new directions on effi - workshop were a more formal understanding of cient large-scale inference and knowledge extrac - the limits of some approaches to plan recognition, tion. learning plans for recognition by demonstration, Jim Prentzas and Ioannis Hatzilygeroudis pre - and activity and behavior recognition for use in sented a paper that tackles the problem of knowl - home environments edge extraction from neural networks (and knowl - Christopher Geib, Gita Sukthankar, David Pyna - edge extraction from graphical models or complex dath, and Hung Bui served as cochairs of this work - networks, in general) — a crucial part of NeSy sys - shop. The papers of the workshop were published tems but also an important research area in its own as AAAI Press Technical Report WS-10-05. right. The discussion session was coordinated by Artur d’Avila Garcez. D’Avila Garcez listed a large num - Statistical Relational AI ber of challenges and areas for further research. Much has been achieved in the field of AI, “the sci - The discussion then focused on two main issues: ence and engineering of making intelligent endowing agent systems with emotions and repre - machines” as John McCarthy defines it, yet much sentations for learning, including nonclassical log - remains to be done if we are to reach the goals we ic and deep networks. Other issues included han - all imagine. One of the key challenges with mov - dling inconsistency, model checking and the ing ahead is closing the gap between logical and modeling of beliefs and intentions. statistical AI. Logical AI has mainly focused on Artur S. d’Avila Garcez, Pascal Hitzler, and Luis complex representations, and statistical AI on C. Lamb served as cochairs of this event. The uncertainty. Clearly, however, intelligent machines papers of this workshop were not published. must be able to handle the complexity and uncer - tainty of the real world. The last decade has seen real progress toward Plan, Activity, and closing the gap. Nowadays, we can learn proba - bilistic relational models automatically from mil - Intent Recognition lions of interrelated objects. We can generate opti - Plan recognition, activity recognition, and intent mal plans and learn to act optimally in uncertain recognition all involve making inferences about environments involving millions of objects and other actors from observations of their behavior, relations among them. Exploiting shared factors that is, their interaction with the environment and can speed up message-passing algorithms for rela - with each other. This area of research combines and tional inference but also for classical propositional unifies techniques from user modeling, machine inference such as solving SAT problems. We can vision, intelligent user interfaces, human or com - even perform lifted probabilistic inference avoid - puter interaction, autonomous and multiagent sys - ing explicit state enumeration by manipulating tems, natural language understanding, and first-order state representations directly. So far, machine learning. It plays a crucial role in a wide however, the researchers combining logic and variety of applications including assistive technolo - probability in each of the AI subfields have been gy, software assistants, gaming, computer and net - working mostly independently. work security, behavior recognition, coordination The Statistical Relational AI workshop convened in robots and software agents, e-commerce and col - researchers driving forward work in this area laborative filtering. This diversity of applications through the interplay between addressing AI tasks and disciplines, while producing a wealth of ideas and using statistical relational techniques to solve and results, has contributed to fragmentation in the them, forming a common core of problems and field, as researchers publish relevant results in a wide ideas, and ultimately starting to explore what we spectrum of journals and conferences. This work - call “statistical relational AI”: the science and engi -

102 AI MAGAZINE Reports neering of making intelligent machines that act in papers presented at the workshop was the primacy noisy worlds composed of objects and relations of spatial reasoning, especially in such nonobvious among the objects. domains as blind-map navigation and the diagno - The 19 papers and posters at the workshop cov - sis of certain mental illnesses. Generally, these ered a wide range of topics including Bayesian papers built upon the major theme of the AAAI abductive reasoning, lifted inference, lifted plan - 2009 fall symposium on multirepresentational ning and planning by probabilistic programming, architectures for human-level intelligence. The lifted SAT, relational learning, probabilistic pro - workshop also included an invited talk, given by gramming for natural language processing, rela - Paul Rosenbloom (University of Southern Califor - tional data integration, cognitive architectures, nia) on the development and research direction of and killer applications, among others. SOAR 9, and its focus on the integration of first- In the first part of the workshop, researchers pre - class perceptual processing. sented new technical contributions. Bart Selman’s The workshop concluded with a lively discus - and Josh Tenenbaum’s invited talks provided a sion and debate among the participants concern - synthesis of probabilistic and logical inference and ing the place of visual and spatial reasoning in the a statistical relational AI perspective on acquiring pantheon of artificial intelligence research. While commonsense theories respectively. The second the utility of purely visual representations in tradi - part of the workshop was a lively poster session tional physical symbol systems was apparent to that encouraged the participants to discuss the those in attendance, considerable discussion commonalities and need for differences among the occurred regarding the realization of perceptual various AI tasks that can be addressed by statistical symbol systems. relational techniques. The group reached a gener - Keith McGreggor and Maithilee Kunda served as al consensus that statistical relational AI is an cochairs of this workshop. The papers of the work - exciting emerging area requiring more investiga - shop were published as AAAI Press Technical tion. The topic of efficient and lifted inference Report WS-10-07. found particular interest. Kristian Kersting, Stuart Russell, Leslie Pack Kael - Abstraction, bling, Alon Halevy, Sriraam Natarajan, and Lilyana Mihalkova served as cochairs of this workshop. Reformulation, The papers of the workshop were published as and Approximation AAAI Press Technical Report WS-10-06. It has been recognized since the inception of arti - ficial intelligence that abstractions, problem refor - Visual Representations mulations, and approximations (ARA) are central and Reasoning to human commonsense reasoning and problem solving and to the ability of systems to reason Visual representations and reasoning play an effectively in complex domains. ARA techniques important role in human problem solving, model - have been used in a variety of problem-solving set - ing, and design. Although the ability to think like tings and application domains, primarily to over - a human long has been a goal of AI, today’s AI come computational intractability by decreasing agents nonetheless are limited in their visual rea - the combinatorial costs associated with searching soning. Advances in this area may enable more large spaces. In addition, ARA techniques are also extensive autonomous reasoning in visual useful for knowledge acquisition and explanation domains, foster deeper computational support for generation. and understanding of human problem solving, The aim and scope of this one-day workshop modeling, and design, and promote more intense were similar to an independent and bigger sympo - use of visual representations in human-machine sium series called SARA (Symposium on Abstrac - interaction. These technological goals raise basic tion, Reformulation, and Approximation). Since theoretical issues such as the precise role of visual the early 1990s, the SARA symposia have provided reasoning in intelligence and the relationship an opportunity to bring together participants with between visual reasoning and perceptual processes. diverse backgrounds, leading to a rich and lively This interdisciplinary workshop aimed to describe exchange of ideas, allowing the comparison of and discuss the latest scientific research that may goals, techniques, and paradigms, and helping inform and influence progress toward these goals. identify important research issues and engineering The workshop brought together participants hurdles. This workshop continued to do the same, from diverse research communities such as AI, cog - filling in a gap that was felt in the recent years as nitive science, learning science, and design sci - SARA has lately been organized every other year. ence, in addition to researchers in the fields of phi - The workshop provided a somewhat more infor - losophy, public policy, health systems, and mal setting than the symposia, without formal human-computer interaction. One major theme of published proceedings but with an AAAI Press

WINTER 2010 103 Reports

Technical Report documenting the papers present - Charles Isbell is an associate professor in the School of ed at the workshop. Interactive Computing, Georgia Institute of Technology. This was the first year for this workshop. With Darsana Josyula is an assistant professor of computer sci - 11 papers grouped into three themes—ARA for ence at Bowie State University. constraint-satisfaction problems, ARA for machine learning, and ARA for reasoning in combinatorial Leslie Pack Kaelbling is a professor of computer science and engineering at the Massachusetts Institute of Tech - search spaces—the workshop brought together an nology. exciting range of topics to the table, especially those of interest to AAAI conference participants. Kristian Kersting heads the Statistical Relational Mining The talks touched upon topics such as satisfiability group at Fraunhofer IAIS and is affiliated with the Com - solvers, knowledge compilation, state space puter Science Department at the University of Bonn, Ger - abstraction for search, reinforcement learning, many. inductive logic programming, and answer set pro - Maithilee Kunda is a computer-science doctoral student gramming. The active presence of senior at Georgia Institute of Technology, Atlanta, Georgia. researchers in the area such as Robert Holte (Uni - Luis C. Lamb is an associate professor of computer sci - versity of Alberta, Canada) helped provide a broad ence at the Federal University of Rio Grande do Sul, perspective throughout the workshop, especially Brazil. during the open discussion session held toward the end. Bhaskara Marthi is a research scientist at Willow Garage, Inc., in California. The participants emphasized strong interest in the general area of abstraction in computer science Keith McGreggor is a computer-science doctoral student and indicated that having this workshop fill the at Georgia Institute of Technology, Atlanta, Georgia. gap between the 2009 and 2011 SARA symposia Lilyana Mihalkova is a member of the LINQS group at was highly useful in maintaining a feeling of con - the Computer Science Department, University of Mary - tinuity in this research community. land College Park. Gregory Provan and Ashish Sabharwal served as Vivi Nastase is a research associate at the Heidelberg cochairs of this workshop. The papers of the work - Institute for Theoretical Studies in Heidelberg, Germany. shop were published as AAAI Press Technical Report WS-10-08. Sriraam Natarajan is a post-doctoral research associate at the Department of Computer Science in the Universi - ty of Wisconsin Madison.

Gregory Provan is a professor in the Department of David W. Aha leads the Adaptive Systems Section at the Computer Science at University College Cork, Ireland, Naval Research Laboratory in Washington, D.C. and directs the Cork Complex Systems Lab.

Mark Boddy is a member of the technical staff at Adven - Anita Raja is an associate professor of software and infor - tium Labs, Minneapolis, Minnesota. mation systems at the University of North Carolina at Charlotte. Vadim Bulitko is an associate professor in the Depart - ment of Computer Science, University of Alberta. Ashwin Ram is an associate professor in the School of Interactive Computing, Georgia Institute of Technology. Artur S. d’Avila Garcez is a reader in computing at City University London, UK. Mark Riedl is an assistant professor in the School of Interactive Computing, Georgia Institute of Technology. Prashant Doshi is an associate professor at the Universi - ty of Georgia. Stuart Russell is a professor of computer science and engineering at the University of California, Berkeley. Stefan Edelkamp is a researcher and lecturer at the Uni - versity of Bremen, Germany Ashish Sabharwal is a research associate in the Depart - ment of Computer Science at Cornell University, USA, Christopher Geib is a research fellow in the School of and works at the Institute for Computational Sustain - Informatics at the University of Edinburgh. ability.

Piotr Gmytrasiewicz is an associate professor at the Uni - Jan-Georg Smaus is a Privatdozent at the University of versity of Illinois at Chicago. Freiburg, Germany.

Robert P. Goldman is a principal scientist at Smart Infor - Gita Sukthankar is an assistant professor in the School mation Flow Technologies LLC in Minneapolis, Min - of Electrical Engineering and Computer Science at the nesota. University of Central Florida.

Alon Halevy heads the Structured Data Group at Google Karl Tuyls is an associate professor at Maastricht Univer - Research. sity, The Netherlands.

Pascal Hitzler is an assistant professor of computer sci - Ron van der Meyden is a professor at the University of ence at Wright State University, Dayton, Ohio. New South Wales and NICTA, Australia.

104 AI MAGAZINE ICWSM Call for Papers

WINTER 2010 105 ICWSM Call for Papers

106 AI MAGAZINE IAAI-11 Call for Papers

WINTER 2010 107 IAAI-11 Call for Papers

108 AI MAGAZINE