Pinball: Planning and Learning in a Dynamic Real-Time Environment

摘要

We are interested in the automation of planning and learning in dynamic physical environments. Recently we have begun to consider the game of pinball as a robotics testbed. Our intention is to build an au-tonomous learning agent that controls the game and improves its performance over time. We believe that pinball will give us a structured environment in which to investigate issues of planning and learning, while providing the real-world constraints of uncertainty and real-time control. Our work to date has led to a sim-ple, physically realistic pinball simulation, which we are now using to test learning agents. In the future we will have a physical game interfaced to our com-puter system. Why Pinball? In the past, we have studied learning problems in robot