Abstract:

Learning is crucial aspect of any intelligent agent. The bulk of this talk with focus on our results in multi-agent learning, where agents must learn to adapt in environments populated with other adaptive, autonomous agents. I'll also spend some time briefly describing new projects in teachable agents, where agents can learn more rapidly by receiving interactive human instruction, and adaptive tutoring systems, where the tutoring system must learn to adapt to differing student capabilities and styles. In multi-agent environments, learning must account for the adaptive nature of the other agents. Traditional models such as MDPs, POMDPs, and game theoretic equilibria each have their shortcomings in this domain: e.g. the environment is not Markov, or the other agents may not be entirely rational. Regret is a principled framework for evaluating the performance of multi-agent learning algorithms, and regret-minimizing algorithms offer a good approach to this domain, one that does not need to make strong assumptions regarding expected types of opponents. I'll describe an algorithm that exhibits good performance against a wide range of possible opponents, and guarantees low regret against any arbitrary opponent.