We perform experiments on 4 new tasks in SMACv2. For calarity and computation limit, we do not include the results of the linear baseline. The results show that our method SPMARL consistently outperforms the baselines.
We compare the std. on two more tasks from BenchMARL. Consistent to previous experiments in SMACv2, the results show that our method SPMARL has lower estimation variance.
We perform ablation study on the $V_{LB}$.