Exploiting Expert Knowledge in Factored POMDPs

摘要

Decision support in real-world applications is often challenging because one has to deal with large and only partially observable domains. In case of full observability, large deterministic domains are successfully tackled by making use of expert knowledge and employing methods like Hierarchical Task Network (HTN) planning. In this paper, we present an approach that transfers the advantages of HTN planning to partially observable domains. Experimental results for two implemented algorithms, UCT and A* search, show that our approach significantly speeds up the generation of high-quality policies: the policies generated by our approach consistently outperform policies generated by Symbolic Perseus and can be computed in less than 10% of its runtime on average.