Efficient, guaranteed search with multi-agent teams

Abstract

Here we present an anytime algorithm for clearing an environment using multiple searchers. Prior methods in the literature treat multi-agent search as either a worst-case problem (i.e., clear an environment of an adversarial evader with potentially infinite speed), or an average-case problem (i.e., minimize average capture time given a model of the targetas motion). We introduce an algorithm that combines finite-horizon planning with spanning tree traversal methods to generate plans that clear the environment of a worst-case adversarial target and have good average-case performance considering a target motion model. Our algorithm is scalable to large teams of searchers and yields theoretically bounded average-case performance. We have tested our proposed algorithm through a large number of experiments in simulation and with a team of robot and human searchers in an office building. Our combined search algorithm both clears the environment and reduces average capture times by up to 75% when compared to a purely worst-case approach. I. INTRODUCTION Imagine you are the leader of a team of agents (humans, robots, and/or virtual agents), and you enter a building looking for a person, moving object, or contaminant. You wish either to locate a target in the environment or authoritatively say that no target exists. Such a scenario may occur in urban search and rescue [1], military operations, network decontamination [2], or even aged care [3]. In some special cases, you may have a perfect model of how the target is moving; however, in most cases you will only have an approximate model or even no model at all. To complicate the situation further, the target may be adversarial and actively avoiding being found. Known algorithms would force you, the leader, to make a choice in this situation. Do you make the worst-case assumption and choose to treat the target as adversarial? This would allow you to utilize graph search algorithms to guarantee finding the target (if one exists), but it would not allow you to take advantage of any model of the targetas motion. As a result your search might take a very long time. Or do you decide to trust your motion model of the target and assume that the target is non-adversarial? This assumption would allow the use of efficient (average-case) search methods from the optimization literature, but it would eliminate any guarantees if the model is inaccurate. In this case, your target may avoid you entirely. It is necessary to make one of these choices because no existing method provides fast search times and also guarantees finding a target if the model is wrong. Fig.