Need help with adversarial-policies?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

182 Stars 30 Forks MIT License 471 Commits 3 Opened issues


Find best-response to a fixed policy in multi-agent RL

Services available


Need anything else?

Contributors list

CircleCI codecov

Codebase to train, evaluate and analyze adversarial policies: policies attacking a fixed victim agent in a multi-agent system. See paper for more information.


The easiest way to install the code is to build the Docker image in the

. This will install all necessary binary and Python dependencies. Build the image by:
  $ docker build .

You can also pull a Docker image for the latest master commit from

. Once you have built the image, run it by:
  docker run -it --env MUJOCO_KEY=URL_TO_YOUR_MUJOCO_KEY \
         humancompatibleai/adversarial_policies:latest /bin/bash  # change tag if built locally

If you want to run outside of Docker (for example, for ease of development), read on.

This codebase uses Python 3.7. The main binary dependencies are MuJoCo (version 1.3.1, for

environments, and 2.0 for the others). You may also need to install some other libraries, such as OpenMPI.

Create a virtual environment by running

. Activate it by
. ./venv/bin/activate
. Finally, run
pip install -e .
to install an editable version of this package.

Reproducing Results

Note we use Sacred for experiment configuration.

Training adversarial policies

trains a single adversarial policy. By default it will train on
for a brief period of time. You can override any of config parameters, defined in
, at the command line. For example, to replicate one of the experiments in the paper, run:
  # Train on Sumo Humans for 20M timesteps
  python -m aprl.train with env_name=multicomp/SumoHumans-v0 paper

trains multiple adversarial policies, using Ray (see below) for parallelization. To replicate the results in the paper (there may be slight differences due to randomness not captured in the seeding), run
python -m aprl.multi.train with paper
. To run the hyperparameter sweep, run
python -m aprl.multi.train with hyper

You can find results from our training run on s3://adversarial-policies-public/multitrain/paper. This includes TensorBoard logs, final model weights, checkpoints, and individual policy configs. Run `experiments/

to sync this and other data to

Evaluating adversarial policies

evaluates a pair of policies, for example an adversary and a victim. It outputs the win rate for each agent and the number of ties. It can also render to the screen or produce videos.

We similarly use

to evaluate multiple pairs of policies in parallel. To reproduce all the evaluations used in the paper, run the following bash scripts, which call
internally: -
: fixed baselines (no adversarial policies). -
. To use our pre-trained policies, use the path
after syncing against S3.

Visualizing Results

Most of the visualization code lives in the

package. To reproduce the figures in the paper, use
; for those in the appendix, use
. So:
  python -m aprl.visualize.scores with paper_config  # heatmaps in the paper
  python -m with supplementary_config  # training curves in appendix

To re-generate all the videos, use

. We would recommend running in Docker, in which case it will render using
. This avoids rendering issues with many graphics drivers.

Note you will likely need to change the default paths in the config to point at your evaluation results from the previous section, and desired output directory. For example:

  python -m aprl.visualize.scores with tb_dir= \
  python -m aprl.visualize.make_videos with adversary_path=

Additional Analysis

The density modeling can be run by

, or with custom configurations via

The t-SNE visualizations can be replicated with


Using Ray

Many of the experiments are computationally intensive. You can run them on a single machine, but it might take several weeks. We use Ray to run distributed experiments. We include example configs in

. To use
you will need to, at a minimum, edit the config to use your own AMI (anything with Docker should work) and private key. Then just run
ray up 
and it will start a cluster. SSH into the head node, start a shell in Docker, and then follow the above instructions. The script should automatically detect it is part of a Ray cluster and run on the existing Ray server, rather than starting a new one.


The codebase follows PEP8, with a 100-column maximum line width. Docstrings should be in reST.

Please run the

before committing. This runs several linting steps. These are also run as a continuous integration check.

I like to use Git commit hooks to prevent bad commits from happening in the first place:

ln -s ../../ci/ .git/hooks/pre-commit

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.