Solving Dec-MDPs with options and intention recognition

Abstract

In this thesis, we designed and implemented an algorithm to find approximate solutions to multi-agent systems. We model the problems with a Decentralized Markov Decision Process, and we make use of options and intention recognition to solve the problem. Rather than directly solving the Dec-MDP, which is NEXP-Complete, we instead solve a set of single-agent MDPs, that we can solve in P-Complete, and combine these solutions during execution time. We tested our algorithm on several instances of the Bribed Package Retrieval Problem and we were able to handle problems as large as our MDP solver would allow, which is a big improvement over what optimal Dec-MDP solvers can handle.