Code for experiments in Let’s Play Again: Variability of Deep Reinforcement Learning Agents in Atari Environments
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
baselines @ f912048
data
models
plots
.gitignore
.gitmodules
README.md
extract_learning_curves.py
plot_learning_curves.R
plot_scores.R
requirements.txt
run.sh
run_scores.sh
start_python

README.md

variability-RL

Accompanying code for "Let’s Play Again: Variability of Deep Reinforcement Learning Agents in Atari Environments"

To cite this work, please use:

@inproceedings{clary2018variability,
  title={{Let's Play Again: Variability of Deep Reinforcement Learning Agents in Atari Environments}},
  author={Clary, Kaleigh and Tosch, Emma and Foley, John and Jensen, David},
  booktitle={{Critiquing and Correcting Trends in Machine Learning Workshop at Neural Information Processing Systems}},
  year={2018}
  }

Installation and Dependencies

  1. Clone this repo, and install the Python dependencies listed in requirements.txt. We used Python 3.6.
  2. Clone the OpenAI Baselines fork listed as a submodule and follow the installation instructions.

Experiment Replication

  1. Execute run.sh to train agents 10M steps on several Atari environments. For information about the cluster used for these experiments, please see the UMass gypsum documentation.
  2. To plot learning curves, use extract_learning_curves.py to pull all logs into a single file, and use plot_learning_curves.R to generate learning curve plots in ggplot.
  3. Once models have trained, execute run_scores.sh to play 100 games with each model and collect final game scores.
  4. Use plot_scores.R to replicate the score distribution plots shown in the paper.