Efficient reinforcement learning with relocatable action models

Abstract

Realistic domains for learning possess regularities that make it possible to generalize experience across related states. This paper explores an environment-modeling framework that rep-resents transitions as state-independent outcomes that are common to all states that share the same type. We analyze a set of novel learning problems that arise in this framework, providing lower and upper bounds. We single out one partic-ular variant of practical interest and provide an efficient algo-rithm and experimental results in both simulated and robotic environments.