Value-Decomposition Multi-Agent Actor-Critics

The exploitation of extra state information has been an active research area in multi-agent reinforcement learning (MARL). QMIX represents the joint action-value using a non-negative function approximator and achieves best performance on StarCraft II micromanagement testbed, common MARL benchmark. However, our experiments demonstrate that, some cases, performs sub-optimally with A2C framework, training paradigm that promotes algorithm efficiency. To obtain reasonable trade-off between efficiency performance, we extend value-decomposition to actor-critic methods are compatible propose novel (VDAC). We evaluate VDAC task proposed framework improves median over other methods. Furthermore, use set ablation identify key factors contribute VDAC.