were also surprised that they perform poorly, so we changed to optimizing individual rewards (easier settings) to

In addition, similar conclusions can be found in Ref.[24], "learning nearly decomposable value StarCraft II (cooperative game) and believed that this is because it cannot deal with the issue of reward assignment. Predators do not need to learn sophisticated strategy to encircle multiple moving preys. We will benchmark I2C alone and add it up for a more thorough comparison in the final version. If agents do not have any visibility radius, it means the environment is fully observable including all other agents. In addition, too much redundant information could impair the learning.