We combined Reward Propagation using Graph Convolutional Networks which was presented as a spotlight at NeurIPS 2020 with the Intrinsic Curiosity Module [Curiosity-driven Exploration by Self-supervised Prediction] (https://arxiv.org/pdf/1705.05363.pdf).
# PyTorch
conda install pytorch torchvision -c soumith
# Other requirements
pip install -r requirements.txt
pip install mujoco-py==2.0.2.2 #optional
#Installing PyGCN
python setup_gcn.py installTo launch a run on one of the Atari games, use the following command:
python control/main.py --num-frames 10000000 --algo ppo --use-gae --lr 2.5e-4 --clip-param 0.1 --value-loss-coef 0.5 --num-processes 8 --num-steps 128 --num-mini-batch 4 --gcn_alpha 0.9 --log-interval 1 --env-name ZaxxonNoFrameskip-v4 --seed 0 --entropy-coef 0.01 --use-logger --folder resultsTo launch a run on one of the delayed MuJoCo environments, use the following command:
python control/main.py --num-frames 3000000 --algo ppo --use-gae --lr 3e-4 --clip-param 0.1 --value-loss-coef 0.5 --num-processes 1 --ppo-epoch 10 --num-steps 2048 --num-mini-batch 32 --gcn_alpha 0.6 --log-interval 1 --env-name Walker2d-v2 --seed 0 --entropy-coef 0.0 --use-logger --folder results --reward_freq 20
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent