Emma Brunskill, Bethany R. Leffler, Lihong Li, Michael L. Littman, Nicholas Roy.
Year: 2009, Volume: 10, Issue: 68, Pages: 1955−1988
To quickly achieve good performance, reinforcement-learning algorithms for acting in large continuous-valued domains must use a representation that is both sufficiently powerful to capture important domain characteristics, and yet simultaneously allows generalization, or sharing, among experiences. Our algorithm balances this tradeoff by using a stochastic, switching, parametric dynamics representation. We argue that this model characterizes a number of significant, real-world domains, such as robot navigati on across varying terrain. We prove that this representational assumption allows our algorithm to be probably approximately correct with a sample complexity that scales polynomially with all problem-specific quantities including the state-space dimension. We also explicitly incorporate the error introduced by approximate planning in our sample complexity bounds, in contrast to prior Probably Approximately Correct (PAC) Markov Decision Processes (MDP) approaches, which typically assume the estimated MDP can be solved exactly. Our experimental results on constructing plans for driving to work using real car trajectory data, as well as a small robot experiment on navigating varying terrain, demonstrate that our dynamics representation enables us to capture real-world dynamics in a sufficient manner to produce good performance.