Multiagent Planning with Trembling-Hand Perfect Equilibrium in Multiagent POMDPs

Multiagent Partially Observable Markov Decision Processes are a popular model of multiagent systems with uncertainty. Since the computational cost for finding an optimal joint policy is prohibitive, a Joint Equilibrium-based Search for Policies with Nash Equilibrium (JESP-NE) is proposed that finds a locally optimal joint policy in which each policy is a best response to other policies; i.e., the joint policy is a Nash equilibrium. One limitation of JESP-NE is that the quality of the obtained joint policy depends on the predefined default policy. More specifically, when finding a best response, if some observation have zero probabilities, JESP-NE uses this default policy. If the default policy is quite bad, JESP-NE tends to converge to a sub-optimal joint policy. In this paper, we propose a method that finds a locally optimal joint policy based on a concept called Trembling-hand Perfect Equilibrium (TPE). In finding a TPE, we assume that an agent might make a mistake in selecting its action with small probability. Thus, an observation with zero probability in JESP-NE will have non-zero probability. We no longer use the default policy. As a result, JESP-TPE can converge to a better joint policy than the JESP-NE, which we confirm this fact by experimental evaluations.