Neural Network Approximations of Solution Concepts for Multiagent Coalitions

Coalition formation is an important aspect of multiagent systems because it enables agents to achieve goals more efficiently or goals they cannot accomplish individually. In this paper we consider an approximate method based on neural networks to estimate two important values used for dividing the payoff of a coalition, namely the Shapley value and the nucleolus. We try several neural network topologies and different training algorithms and evaluate the behavior of an especially designed multiagent system when the payoff values are computed by exact and approximate methods.