Indie Machine Learning and Video Game Dev

Neural Net Downloads
How to
Useful links
Hardware Reviews
Get Merch
16 August 2019

DeepMind Bsuite setup guide

by Mathieu Poliquin

Recently DeepMind released Bsuite (GitHub Page), a series of standard experiments to benchmark and compare core capabilities of RL algos.

Image taken from DeepMind’s Bsuite github bsuite

They already have over 20 experiments such as cartpole, mnist and memory tests

You can find the whole list here: https://github.com/deepmind/bsuite/tree/master/bsuite/experiments

There is already a setup guide on their GitHub page otherwise I made a simplified one below, targeted at testing the ppo2 algo

1. Install Bsuite

git clone https://github.com/deepmind/bsuite.git

Install bsuite with the baselines dependencies.

pip3 install -e bsuite[baselines]

Note that you will need to install openAI baselines and Gym seperatly. If you follow this guide to install it Setup OpenAI baselines and Gym Retro

2. Run test script

Now to make sure your setup work well run this test script


Note: You might need to edit the scrypt since it assumes you have python 3.6 In my case I have python 3.7 installed so I edited this line:


virtualenv -p /usr/bin/python3.6 bsuite_env


virtualenv -p /usr/bin/python3.7 bsuite_env

You should see something like this as a result: test

3. Test PPO2

First run the test to make sure the your basic setup is working. Don’t forget that you need baselines and gym python packages for that

cd bsuite/baselines/openai_ppo
python3 ./run_test.py

You should have a message saying if the test passed and how long it took

Now run the full suite of experiments:

python3 ./run.py --bsuite_id=SWEEP

Note: by default it uses MLP as neural network type as opposed to CNN for OpenAI retro You can use –network=cnn

To list all the available experiments

>>> from bsuite import sweep
>>> sweep.SWEEP

To list only a subset for example all the experiments for deep sea

>>> from bsuite import sweep
>>> sweep.DEEP_SEA
tags: DeepMind - Bsuite - setup guide - tutorial - ppo2