Pursuit Policies in Dynamic Environments

Joseph L Briones, Andrea W Richa

Keywords: reinforcement learning, multi-agent reinforcement learning, deep Q-networks

TL;DR: Investigating the impact of dynamic environments on learning predator pursuit policies from partial observations with deep Q-learning.

Abstract: Cooperative pursuit is a popular multi-agent reinforcement learning (MARL) game where a team of predators target prey while avoiding obstacles. Previous literature has largely considered the impact of different predator, prey abilities on learning. Here, we investigate the impact of dynamic environments on learning predator pursuit policies from partial observations with deep Q-learning. Interestingly, we find predators are able to learn cooperative pursuit strategies that leverage moving obstacles.