pymarl
pymarl copied to clipboard
Reward function design
Hi
I am currently wondering about the reward design. The implemented function resembles a global reward function and these should be normalized by the amount of agents. This seems to happen here with respect to the total reward obtainable by kills and round win. But what about the battle reward? It includes rewards from kills, deaths, damage taken and dealt. Am I missing it or how are damage taken, damage dealt and deaths normalized over the amount of considered agents?