Effective Approximations for Spatial Task Allocation Problems

Abstract

Although multi-robot systems have received substantial research attention in recent years, multi-robot coordination still remains a difficult task. Especially, when dealing with spatially distributed tasks and many robots, central control quickly becomes infeasible due to the exponential explosion in the number of joint actions and states. We propose a general algorithm that allows for distributed control, that overcomes the exponential growth in the number of joint actions by aggregating the effect of other agents in the system into a probabilistic model, called subjective approximations, and then choosing the best response. We show for a multi-robot grid-world how the algorithm can be implemented in the well studied Multiagent Markov Decision Process framework, as a sub-class called spatial task allocation problems (SPATAPs). In this framework, we show how to tackle SPATAPs using online, distributed planning by combining subjective agent approximations with restriction of attention to current tasks in the world. An empirical evaluation shows that the combination of both strategies allows to scale to very large problems, while providing near-optimal solutions.