PyTorch implementation of paper "Attentive Update of Multi-Critic (AUMC)" which is accepted as ICME-2021 oral paper. Method is tested on MuJoCo continuous control tasks in OpenAI Gym. Agents are trained using PyTorch 1.4 and Python 3.6.
# python 3.6 (apt)
# pytorch 1.4.0 (pip)
# tensorflow 1.14.0 (pip)
# DMC Control Suite and MuJoCo
cd dockerfiles
docker build . -t aumcRLFor other dockerfiles, you can go to RL Dockefiles.
Run with the scripts batch_aumc_mujoco_4seed_cuda.sh:
# eg.
bash batch_aumc_mujoco_4seed_cuda.sh Hopper-v2 DDPG_aumc 0 0.4 # env_name: Ant-v2; algorithm: DDPG coupled with AUMC; CUDA_Num : 0; beta: 0.4.Run with the scripts render_mujoco.py / render_aumc_mujoco.py:
# eg. visulization of the environments with random actions:
python render_mujoco.py --env Ant-v2 # env_name: Ant-v2
# or visulization of the environments with trained policy:
CUDA_VISIBLE_DEVICES=0 python render_bootstrapped.py \
--policy "TD3_aumc" \
--env "Ant-v2" \
--load_model "default" \
--seed 2
# env_name: Ant-v2; load policy: policy trained with TD3_aumc with seed equaling 2Including Ant-v2, HalfCheetah-v2, Hopper-v2, Humanoid-v2, Swimmer-v2, Walker2d-v2.
