gkswamy98/adversarial_rl

CS 294-131 Project, reimplementing https://arxiv.org/pdf/1702.02284.pdf.

Adversarial Attacks on Deep Reinforcement Learning

Installation

1. Run git clone --recurse-submodules https://github.com/gkswamy98/adversarial_rl.git to clone with submodules and cd adversarial_rl.
2. Run pip install -e . to install the main dependencies.
3. Run mv baselines ../baselines, cd ../baselines, then pip install -e . --user to install our baselines fork.

Running Experiments

• To train models, run bash scripts/train.sh.
  • To parallelize training runs, modify the training script to echo commands instead of executing them and run bash ../scripts/train.sh | xargs -PN -ICMD /bin/bash -exc CMD, replacing N with the number of cores available.
• To attack trained models, run bash scripts/attack.sh.
  • Use the same trick as above to parallelize.

Results

PPO, TRPO, A2C and DQN models trained on CartPole and Acrobot are in the models/ folder. Results of attacks are available in accompanying writeup.

Authors

• @gkswamy98
• @noahgolmant