Our model:

Trained without any extrinsic reward

Architecture:

Please check the report for details of this work.

Requirements

• python3
• gym
• gym-super-mario-bros
• OpenCV
• PyTorch
• tensorboardX

Train

Train network with a separate controller(Original model but with LSTM as the forward network):

python3 train.py

Train network with controller with shared features with ICM (Our model):

python3 train.py --shared_features

Evaluate

Evaluate network with a separate controller (Original model but with LSTM as the forward network):

python3 eval.py --name eta-0.2_stack-1_sparse_extrinsic_run1 --number 5734400

Evaluate network with controller with shared features with ICM (Our model):

python3 eval.py --name eta-0.2_rnn_forward_both_shared_features_stack-1_only_intrinsic_gradients_feat_run3 --number 5734400

References

Code has been heavily borrowed from the first two. Thanks a lot!

https://github.com/ctallec/world-models

https://github.com/jcwleo/curiosity-driven-exploration-pytorch

https://github.com/pathak22/noreward-rl