Gfootball_MARL
Train a MARL algorithm in Gfootball
Design
imitation learning
- generate data
- supervised learning
It seems that it's difficult to ues imitation learning to get a original policy. Because the gfootball environment set different action-set for agent training and builtin-ai. There are two different sets, even we can't change the ai's action into agent's avaliable action.
If we want to get a nice training data (like a .dump file that contain the whole player's trajectories), it must be generated by the multi-agent game. There is only one action in the .dump from single-agent game.
rule
- write down some base rule about the football game
- simplify each scenes