| "In practice, we found that a high-entropy initial state is more likely to increase the speed of training. | |
| The entropy is calculated by: | |
| $$H=-\sum_{k= 1}^{n_k} p(k) \cdot \log p(k), p(k)=\frac{|A_k|}{|\mathcal{A}|}$$ | |
| where $H$ is the entropy, $|A_k|$ is the number of agent nodes in $k$-th cluster, $|\mathcal{A}|$ is the total number of agents. | |
| To ensure the Cooperation Graph initialization has higher entropy, | |
| we will randomly generate multiple initial states, | |
| rank by their entropy and then pick the one with maximum $H$." | |