E3-Equivariant Actor-Critic Methods for Cooperative Multi-Agent Reinforcement Learning
- Train E3AC on MPE:
cd E3AC/E3-MPE/maddpg
python main_vec_e3.py --scenario [simple_spread_n3,simple_spread_n6,simple_coop_push_n3,simple_coop_push_n6,simple_tag_n3,simple_tag_n6] --continuous --actor_type [MLP,gcn_max,segnn] --critic_type [MLP,segnn] --actor_lr [Please refer to the appendix] --critic_lr [Please refer to the appendix] --fixed_lr --batch_size [Please refer to the appendix] --actor_clip_grad_norm 0.5 --cuda
- Train E3AC on MuJoCo Tasks (single- and multi-agent reacher):
cd E3AC/E3-MuJoCo/Reacher or cd E3AC/E3-MuJoCo/MA_Reacher
python train.py actor_type=[mlp,segnn] critic_type=[mlp,segnn] pixel_obs=false action_repeat=1 frame_stack=1 task=reacher_hard agent=ddpg_e3 lr=[Please refer to the appendix]
- Train E3AC on MuJoCo (multi-agent hopper 3D):
cd E3AC/E3-MuJoCo/MA_Hopper_3D/src
python main.py --env_name hopper --morphologies hopper --exp_path ../results --config_path configs/3d.py --gpu 0 --custom_xml environments/3d_hoppers/3d_hopper_3_shin.xml --actor_type [mlp_v,segnn] --critic_type [mlp_v,segnn]
The E3-MPE code is based on https://github.com/IouJenLiu/PIC.
The E3-MuJoCo is based on https://github.com/sahandrez/homomorphic_policy_gradient and https://github.com/alpc91/SGRL.
The SEGNN is based on https://github.com/RobDHess/Steerable-E3-GNN.