Reproducing paper Prioritized Experience Replay.
Prioritized experience replay (PER) develops a framework for prioritizing experience, so as to replay important transitions more frequently. There are two variants of prioritizing the transitions, rank-based and proportional-based. Our implementation is the proportional variant, which has a better performance, as reported in the original paper.
Results have been reproduced with Double DQN on following three environments:
- paddlepaddle>=1.6.1
- parl
- gym[atari]==0.17.2
- atari-py==0.2.6
- tqdm
- ale_python_interface
Train on BattleZone game:
python train.py --rom ./rom_files/battle_zone.bin
To train on more games, you can install more rom files from here.