This is an implementation of Random network distillation on Montezuma's Revenge using Pytorch.
paper link: https://arxiv.org/abs/1810.12894

To run the program, first install the required packages by executing:

$ pip3 install -r requirements.txt

Run the program with pretrained model and see the agent playing:

$ python3 main.py --play --path models/pretrained_model.pth

x-axis: train steps, y-axis: entropy
this diagram shows how entropy decreases. the agent starts by total random movements and learns a stochastic policy after being trained.

The pretrained model models/pretrained_model.pth is obtained by training with the following settings:

You can train from a model from scratch by using the following command. Note that if you don't specify the variables, They match the default value described in the table above. The save_int varibale describes the interval of saving a model checkpoint.

Some useful diagrams are stored in tensorboard format while training.

python3 main.py --train --num_env 64 --train_steps 12000 --predictor_update_p 0.25 --num_pre_norm_steps 10 --game_steps 128 --num_epoch 4 --mini_batch 2 --save_int 100 

Train from a checkpoint:

python3 main.py ---train --path logs/desired_checkpoint --num_env 64 --train_steps 12000 --predictor_update_p 0.25 --num_pre_norm_steps 10 --game_steps 128 --num_epoch 4 --mini_batch 2 --save_int 100