Assessing Multi-Agent Reinforcement Learning Algorithms for Autonomous Sensor Resource Management

Abstract

Unmanned aerial vehicles (UAVs) have applications in search and rescue operations and such operations could be more efficient by using appropriate artificial intelligence (AI) to enable a UAV agent to operate autonomously. Sensor resource management (SRM), which leverages capabilities across location intelligence, facilitates the efficient and effective use of UAVs and their sensors to complete a set of tasks. Furthermore, multiple UAVs, each with different sensor configurations, must be considered when maximizing mission effects. Instantiating operational autonomy for such teams requires considerable coordination. One AI approach relevant to this task is multi-agent reinforcement learning (MARL). However, MARL has seen limited prior use in SRM. This work evaluates the trade-space of MARL algorithms with respect to performing heterogeneous sensor resource management (SRM) tasks, considers the concept of evaluating MARL in a test and evaluation framework and compares a suit of algorithms with random and Bayesian hyperparameter optimization methods.