CAREER: Robust Strategic Reasoning for Multi-Agent Systems

University Of Texas At El Paso, El Paso TX

Investigators

Abstract

Many important decision problems in computational science involve multi-­agent systems (MAS) in which multiple decision makers must choose strategies for action from a set of alternatives. (One example is selecting a security policy --- for example by police --- to protect critical infrastructure against attackers.) This CAREER award project develops methods for analyzing MAS to select strategies for an agent to follow that will lead to desirable outcomes. The central tenant driving the project is that a good strategy should be robust in the sense that it will continue to perform well even when there is uncertainty about the outcomes or how other agents will choose their strategies. An important component of the project is a multi-purpose web platform supporting both research experiments and education on game-playing agents. The PI will design course modules that use this platform to involve undergraduates in designing game-playing agents, offering an exciting way to practice basic programming and develop critical thinking and data analysis skills. The agents designed by students also play an important role in the research; they will provide a diverse library of strategies to evaluate the robustness of reasoning methods for play against opponents with unanticipated behaviors. This platform, and the data sets generated with it will serve the broader research and education communities in computer science and MAS. Research on robust methods for strategic reasoning has the potential to significantly improve decision making in many important real-world problems, including decision support tools for homeland security operations. The availability of robust methods will also enable new applications of computational game theory in domains where confidence in the strategies, despite uncertainty, is critical. This award also supports broadening participation by developing research capacity at UTEP, a minority-­serving institution. The project's primary technical contributions are in computational game theory. Typical game-theoretic solutions are not robust to the kinds of uncertainty that arise in real applications; these include payoff uncertainty, abstraction error, and opponent modeling error. Robustness requires techniques that encompass unanticipated situations, as well as those where it is possible to precisely characterize the uncertainty. The project extends the framework of meta-games to provide a methodology for evaluating robustness in the context of different kinds of uncertainty, including the three listed above. The PI uses meta-games to evaluate new concepts for robust strategic reasoning, including methods based on Bayesian games, interval-based approaches, and approaches drawn from behavioral game theory. The web-based platform developed in the project facilitates extensive experiments with game-playing agents. This will be used to collect a diverse pool of unanticipated agent strategies (including ones designed by undergraduate students), enabling a more comprehensive investigation of robustness to unanticipated opponent behaviors.