Rules-PPO-QMIX: Multi-Agent Reinforcement Learning with Mixed Rules for Large Scene Tasks

Abstract

Under the StarCraft multi-agent challenge (SMAC), the large scenes with multiple task objectives are obstacles for directly applying the existing multi-agent reinforcement learning (MARL) algorithms. The main challenge lies in that the recent MARL algorithms have a poor performance in these situations. A novel Rules-PPO-QMIX MARL algorithm is designed to 1) determine the optimal target and paths with tools of manual rules and proximal policy optimization (PPO) and 2) perform decentralised micromanagement near the target with monotonic value function factorisation MARL algorithm (QMIX). In such way, the complete decision-making process is divided into several parts, that is, path planning, target selecting and micromanagement, which is yielded by manual rules, single-agent reinforcement learning and MARL, respectively. The effectiveness of Rules-PPO-QMIX is validated by a testing map of SMAC, with QMIX as baseline. To demonstrate the feasibility on small scenes, a classical 8m SMAC environment is also used to test the proposed algorithm.