Decomposition of Markov Decision Processes Using Directed Graphs
Résumé
Decomposing an MDP consists in (1) partitioning the state space into regions; (2) solving each region independently; (3) combining the local solutions to generate a global one. In our approach, domain characteristics are used to accurately define how region communicates with each others, which is very important to ensure good quality of the approximate policies obtained. We apply it to mobile robotics planning in indoor environments.