论文: QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning

背景

同VDN

4. QMIX

假设 Qtot 与 Qi 有如(4)式的关系.

(4)式可以通过(5)式实现.

(5)式可以通过如下Fig2的QMIX网络架构实现

1. agent networks: 进行local决策
   • 即DRQN, Qi
2. mixing network: 实现线性并保证单调(式(5))
   • 保证单调的方法
     1. 网络权重W均为正 (bias无所谓)
     2. (非线性的)激活函数单调递增
3. hypernetworks: 融入state信息
   1. state为何不直接与Qi 一同作为输入?
      1. 因为 Qtot 和 st 之间没有单调关系
   2. 实现灵活输入以方便 mix net 评估 joint action-value
      1. 由state经过NN生成mix net的weight 和 bias
   3. 为什么最后的bias用了两层Linear?

7.2. Ablation Results

both central state information and non-linear value function factorisation is required to achieve good performance.

A.1. Representational Complexity

1. Three Keys:
   The value function class representable with QMIX includes any value function that can be factored into a non-linear monotonic combination of the agents’ individual value functions in the fully observable setting.
   1. non-linear: 比VDN更expressive
   2. monotonic (对于单个agent来说) : 单个agent的最优action与其他agent无关
   3. fully observable: observartion != state
      In a Dec-POMDP, QMIX cannot necessarily represent the value function. For example, if