Risk-Aware Multi-Agent Path Planning for Target Detection: A Multi-Agent Reinforcement Learning Approach

摘要

In this paper, we consider the problem when a multi-agent team is tasked with detecting a mobile target under a given time period. In particular, the movement of the target and the multi-agent team are restricted to a set of graph networks. The target travels along a road network while the agents travel along their individual air track networks. The objective is to derive path planning policies for the multi-agent team that minimize the risk of not detecting the target within a given time period. To solve the problem, we propose a multi-agent reinforcement learning approach. The key idea of the proposed approach is to first build an environment model that serves as a "white box" to describe the target and agent dynamic models, and then build local actor-critic networks to train local path planning policies based on samples obtained via cueing the white box. In the proposed approach, we also design realistic reward metrics that reflect the team's probability in detecting the target when the team has not detected the target prior. The performance of the proposed approach is demonstrated by comparing the learned policies with a set of random policies in two simulation studies. The comparison shows that the proposed approach can derive individual path planning policies that dramatically outperform random policies.