Abstraction and Refinement for Solving Continuous Markov Decision Processes.

摘要

We propose a novel approach for solving continuous and hybrid Markov Decision Processes (MDPs) based on two phases. In the first phase, an initial approximate solution is obtained by partitioning the state space based on the reward function, and solving the resulting discrete MDP. In the second phase, the initial abstraction is refined and improved. States with high variance in their value with respect to neighboring states are partitioned, and the MDP is solved locally to improve the policy. In our approach, the reward function and transition model are learned from a random exploration of the environment, and can work with both, pure continuous spaces; or hybrid, with continuous and discrete variables. We demonstrate empirically the method in several simulated robot navigation problems, with different sizes and complexities. Our results show an approximate optimal solution with an important reduction in state size and solution time compared to a fine discretization of the space. 1