Lying in pursuit evasion task with multi-agent reinforcement learning

Abstract

Swarm behaviour in nature has long been an area of research, through which many algorithms have been developed and have found applications in modern problems. A particular research field of multi-agent systems, which are more general to swarms, focuses on using multi-agent reinforcement learning to develop and learn policies of high performance. Communication between agents do exist within swarms and within multi-agent systems, and have been modelled during research. However, lying during communication is an area lacking in research. This project will investigate the effects of lying on a multi-agent system in a pursuit evasion task using multi-agent reinforcement learning to learn an optimal policy, and experiment with different network configurations and techniques such as dropout and layer normalisation.