Error-Bounded Approximations for Infinite-Horizon Discounted Decentralized POMDPs

摘要

We address decentralized stochastic control problems represented as decentralized partially observable Markov decision processes (Dec-POMDPs). This formalism provides a general model for decision-making under uncertainty in cooperative, decentralized settings, but the worst-case complexity makes it difficult to solve optimally (NEXP-complete). Recent advances suggest recasting Dec-POMDPs into continuous-state and deterministic MDPs. In this form, however, states and actions are embedded into high-dimensional spaces, making accurate estimate of states and greedy selection of actions intractable for all but trivial-sized problems. The primary contribution of this paper is the first framework for error-monitoring during approximate estimation of states and selection of actions. Such a framework permits us to convert state-of-the-art exact methods into error-bounded algorithms, which results in a scalability increase as demonstrated by experiments over problems of unprecedented sizes.