Decision Making and Task Allocation in a Multi-Robot System

Abstract

We investigate several methods of explore-exploit decision making for the application of trash collection. To do this, we explore the optimization of the multi-agent Multi-Armed Bandit (MAMAB) problem using the UCB1 algorithm. We consider cases of inter-agent message passing and no message passing, as well as several re-ward distributions. We show that for fixed initial and constant rate addition reward distributions, performance improves as the degree of communication increases. We also show that beyond a certain degree of communication, there is negligible improvement in system performance. We develop a novel simulator for MAMAB problems to verify our theoretical results. We demonstrate that our adaptation of UCB1 achieves cumulative regret logarithmically with time, and is a valid approach to the trash collection problem.