A Novel Exploration Technique For Multi-Agent Reinforcement Learning
Devarani Devi Ningombam · 2022 IEEE 3rd Global Conference for Advancement in Technology (GCAT) · 2022
Multi-agent reinforcement learning (MARL) involves interacting with their environment as well as among themselves, as opposed to single-agent reinforcement learning (RL) which involves one agent interacting with the environment and operated by itself. In most studies on MARL, the calculation of global action-reward value is used to enhance exploration. However, agents' contributions cannot be quantified. By combining curiosity-based intrinsic reward with hierarchical prioritized experience replay, in this paper we propose a technique that quantifies individual agents' contributions to the global achievement. A comparison was made between our proposed algorithm and the existing state-of-the-art algorithms using the StarCraft 2 learning environment. Numerical results demonstrate that our proposed algorithm achieves a higher test win rate and outperforms the existing algorithms under all scenarios tested.