Hierarchical POMDP Decomposition for A Conversational Robot

POMDPs provide a useful framework for decisionmaking in the presence of uncertainty. Finding solutions to large-scale problems, however, has proven computationally infeasible. We propose a hierarchical approach to POMDPs which takes advantage of structure in the domain to decompose the problem into a collection of smaller POMDPs. These can be solved independently, allowing us to solve larger problems than were previously possible. We apply this approach to the problem of humanrobot speech dialogues, and show that appropriate decomposition can yield signi cant computational time reduction when nding a POMDP solution.