Macro-action-based multi-agent deep reinforcement learning in cooperative tasks
Xingyu Lu · 2021
This research is oriented towards the deep multi-agent reinforcement learning to solve thefully cooperative problems in robotic domains. We have developed a cluster of macro-action-based deep multi-agent reinforcement learning frameworks based on which the agent is able to learn the macro-action-based policy for asynchronous sequential decision making. The algorithms to be used are classified into two groups : 1). the decentralized learning with decentralized execution ; 2). the centralized learning with centralized execution. To evaluate the learning methods in a tactical collaborative-task-based robotic environment, we design and implement a multi-quadcopter cooperative simulation environment called "Airborne Balloon Battle Royale"(ABBR) which is also beneficial and applicable in standard macro-action reinforcement learning research and robotic application development.--Author's abstract