A comprehensive analysis of agent factorization and learning algorithms in multiagent systems

Abstract

In multiagent systems, agent factorization denotes the process of segmenting the state-action space of the environment into distinct components, each corresponding to an individual agent, and subsequently determining the interactions among these agents. Effective agent factorization significantly influences the system performance of real-world industrial applications. In this work, we try to assess the performance impact of agent factorization when using different learning algorithms in multiagent coordination settings; and thus discover the source of performance quality of the multiagent solution derived by combining different factorizations with different learning algorithms. To this end, we evaluate twelve different agent factorization instances—or agent definitions—in the warehouse traffic management domain, comparing the training performance of (primarily) three learning algorithms suitable for learning coordinated multiagent policies: the Evolutionary Strategies (ES), the Canonical Evolutionary Strategies (CES), and a genetic algorithm (CCEA) previously used in a similar setting. Our results demonstrate that the performance of different learning algorithms is affected in different ways by alternative agent definitions. Given this, we can conclude that many important multiagent coordination problems can eventually be solved more efficiently by a suitable agent factorization combined with an appropriate choice of a learning algorithm. Moreover, our work shows that ES and CES are effective learning algorithms for the warehouse traffic management domain, while, interestingly, celebrated policy gradient methods do not fare well in this complex real-world problem setting. As such, our work offers insights into the intrinsic properties of the learning algorithms that make them well-suited for this problem domain. More broadly, our work demonstrates the need to identify appropriate agent definitions-multiagent learning algorithm pairings in order to solve specific complex problems effectively, and provides insights into the general characteristics that such pairings must possess to address broad classes of multiagent learning and coordination problems.