Proceedings of the ICML'05 Workshop on Rich Representations for Reinforcement Learning

Reinforcement learning (RL) has developed into a primary approach to learning control strategies for autonomous agents. The majority of RL work has focused on propositional or attribute-value representations of states and actions, simple temporal models of action, and memoryless policy representations. Many problem domains, however, are not easily represented under these assumptions. This has led to work that studies the use of richer representations in RL to overcome some of these traditional limitations. This includes for example: relational reinforcement learning, where states, actions, value functions, and policies have relational representations; richer representations of action and policies that incorporate internal state, such as options and MAXQ hierarchies; and the recently introduced predictive state representations where the state of a system is represented in terms of predictive statements about future observations. The aim of this workshop is to bring together researchers working on various representational aspects of RL to explore their approaches, the relationships among them, their benefits (or drawbacks) for reinforcement learning, and the key representational challenges that remain.Only published electronicallynrpages: 74status: publishe