Improving the performance of complex agent plans through reinforcement learning

Agent programming in complex, partially observable and stochastic domains usually requires a great deal of understanding of both the domain and the task, in order to provide the agent with the knowledge necessary to act effectively. While symbolic methods allow the designer to specify declarative knowledge about the domain, the resulting plan can be brittle since it is difficult to supply a symbolic model that is accurate enough to foresee all possible events in complex environments, especially in the case of partial observability. Reinforcement Learning (RL) techniques, on the other hand, can learn a policy and make use of a learned model, but it is difficult to reduce and shape the scope of the learning algorithm by exploiting a priori information. We propose a methodology for writing complex agent programs that can be effectively improved through experience. We show how to derive a stochastic process from a partial specification of the plan, so that the latter's perfomance can be improved solving a RL problem much smaller than classical RL formulations. Finally, we demonstrate our approach in the context of Keepaway Soccer, a common RL benchmark based on a RoboCup Soccer 2D simulator.