This is the tensorflow implementation of Reinforcement Learning with Perturbed Rewards as described in the following AAAI 2020 paper (Spotlight):

@inproceedings{wang2020rlnoisy,
  title={Reinforcement Learning with Perturbed Rewards},
  author={Wang, Jingkang and Liu, Yang and Li, Bo},
  booktitle={AAAI},
  year={2020}
}

The implementation is based on keras-rl and OpenAI baselines frameworks. Thanks to the original authors!

• gym-control: Classic control games
• gym-atari: Atari-2600 games

• python 3.5
• tensorflow 1.10.0, keras 2.1.0
• gym, scipy, scipy, joblib, keras
• progressbar2, mpi4py, cloudpickle, opencv-python, h5py, pandas

Note: make sure that you have successfully installed the baseline package and other packages following (using virtualenvwrapper to create virtual environment):

mkvirtualenv rl-noisy --python==/usr/bin/python3
pip install -r requirements.txt
cd gym-atari/baselines
pip install -e .

• Classic control (DQN on Cartpole)

cd gym-control
python cem_cartpole.py                                           # true reward
python dqn_cartpole.py --error_positive 0.1 --reward noisy       # perturbed reward
python dqn_cartpole.py --error_positive 0.1 --reward surrogate   # surrogate reward (estimated)

• Atari-2600 (PPO on Phoenix)

cd gym-atari/baselines
python -m baselines.run --alg=ppo2 --env=PhoenixNoFrameskip-v4 \  # true reward
       --num_timesteps=5e7 --normal=True                          
python -m baselines.run --alg=ppo2 --env=PhoenixNoFrameskip-v4 \  # noisy reward
       --num_timesteps=5e7 --save_path=logs-phoenix/phoenix/ppo2_50M_noisy_0.2 \
       --weight=0.2 --normal=False --surrogate=False --noise_type=anti_iden
python -m baselines.run --alg=ppo2 --env=PhoenixNoFrameskip-v4 \  # surrogate reward (estimated)
       --num_timesteps=5e7 --save_path=logs-phoenix/phoenix/ppo2_50M_noisy_0.2 \
       --weight=0.2 --normal=False --surrogate=True --noise_type=anti_iden

To reproduce all the results reported in the paper, please refer to scripts/ folders in rl-noisy-reward-control and rl-noisy-reward-atari:

• gym-control/scripts
  • Cartpole
    • train-cem.sh (CEM)
    • train-dqn.sh (DQN)
    • train-duel-dqn.sh (Dueling-DQN)
    • train-qlearn.sh (Q-Learning)
    • train-sarsa.sh (Deep SARSA)
  • Pendulum
    • train-ddpg.sh (DDPG)
    • train-naf.sh (NAF)
• gym-atari/scripts
  • train-alien.sh (Alien)
  • train-carnival.sh (Carnival)
  • train-mspacman.sh (MsPacman)
  • train-phoenix.sh (Phoenix)
  • train-pong.sh (Pong)
  • train-seaquest.sh (Seaquest)
  • train-normal.sh (Training with true rewards)

If you have eight available GPUs (Memory > 8GB), you can directly run the *.sh scripts one at a time. Otherwise, you can follow the instructions in the scripts and run the experiments. It ususally takes one or two days (GTX-1080 Ti) to train the policy.

cd rl-noisy-reward-atari/baselines
sh scripts/train-alien.sh

The logs and models will be saved automatically. We provide results_single.py for getting the averaged scores:

python -m baselines.results_single --log_dir logs-alien

Please cite our paper if you use this code in your research work.

Please submit a Github issue or contact [email protected] if you have any questions or find any bugs.