Multi-Behavior Multi-Agent Reinforcement Learning for Informed Search via Offline Training

In modern informed search missions, Multi-Robot Systems (MRSs) are playing more and more important roles due to their flexibility in exploring environments. Reinforcement learning (RL) is now widely used as a decision-making method for MRS. However, existing RL-based and conventional model-based frameworks cannot deal with some challenges posed by the realworld environment. To address these challenges, a Multi-Behavior Multi-Agent Reinforcement Learning (MBMARL) framework via offline reinforcement learning method was developed. In this framework, each agent is deployed with multiple behavior policies to let the agent have choices on behaviors given a state. The proposed framework is compared with traditional reinforcement learning frameworks, including Multi-Agent Actor Critic (MAAC) and REINFORCE. The result shows that MBMARL outperforms others in both aspects of total reward and convergence time.