Importance Prioritized Policy Distillation (IPPD)

Train the teacher policy:

Download the code for RAINBOW from github Inside the folder Rainbow-master, run:

python3 main.py —id your_runname —seed 123 —game game_name —T-max 4000000 #game_name example: bank_heist, road_runner, pong, freeway.

The hyperparameters for training teacher policy can be found in Appendix in the paper.

Then run code to collect the dataset: python data_collection.py

Policy distillation training:

## IP-KL
python PD_adaptive_importance_KL.py —game bank_heist

## IP-CE
python PD_adaptive_importance_CrossEntropy.py —game bank_heist

## IP base policy conpression
python PD_adaptive_importance_compression.py —game bank_heist

## The detail of parameter setting for policy distillation is provided in Appendix D in our paper

Policy evaluation:

python evaluation.py --game bank_heist