de73998802680548b916f1947ffbad76-Supplemental.pdf

Now we present a detailed proof for this proposition. In each update, the value of both the two objectives start from the respective lower bounds and areupdated conservativelyduring theoptimization epochs. Similar to Appendix A.5, the decentralized policies can be viewed independently, thus The optimization ofboth the actors and critics is conducted using RMSprop with the learning rate of5 10 4 and α of 0.99. C.2 SMAC The same actor-critic network architecture are utilized for all maps we have evaluated on. The optimization of both the actors and critics is conducted using Adam withthelearning rateof5 10 4 andoptimizer epsilon of1 10 5. Noweight decay isusedin the optimizers.