History-Based Controller Design and Optimization for Partially Observable MDPs

Abstract

Partially observable MDPs provide an elegant framework forsequential decision making. Finite-state controllers (FSCs) are often used to represent policies for infinite-horizon problems as they offer a compact representation, simple-to-execute plans, and adjustable tradeoff between computational complexityand policy size. We develop novel connections between optimizing FSCs for POMDPs and the dual linear programfor MDPs. Building on that, we present a dual mixed integer linear program (MIP) for optimizing FSCs. To assign well-defined meaning to FSC nodes as well as aid in policy search, we show how to associate history-based features with each FSC node. Using this representation, we address another challenging problem, that of iteratively deciding which nodes to add to FSC to get a better policy. Using an efficient off-the-shelf MIP solver, we show that this new approach can find compact near-optimal FSCs for severallarge benchmark domains, and is competitive with previous best approaches.