Reinforcement Learning for Collaborative Search and Rescue Using Unmanned Aircraft System Swarms

This work presents a novel approach to search and rescue, and other wide-area search problems, by combining the effectiveness of deep reinforcement learning (DRL) with the power of Subsumption architecture in a collaborative multi-agent system. We present a multi-agent parameter-sharing deep reinforcement learning paradigm with an action space consisting of the activation of an array of premade “roles.” Each role is a collection of individual behaviors organized in a Subsumption-inspired hierarchy. We describe a low-cost implementation of our approach and its results using a basic DRL algorithm, DQN, and a simple reward signal. Using only a small amount of training time and power with our minimal implementation, we were able to see improvement in swarm performance over baseline statistical methods. These results indicate that our technique can be extended to achieve even greater performance results over a wide variety of problem sets. Additionally, this technique can be extended to many problems within the swarm and collaborative multi-agent system space.