• Stars
    star
    1,497
  • Rank 31,357 (Top 0.7 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created over 5 years ago
  • Updated 7 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

bsuite is a collection of carefully-designed experiments that investigate core capabilities of a reinforcement learning (RL) agent

Behaviour Suite for Reinforcement Learning (bsuite)

PyPI Python version PyPI version pytest

radar plot

Introduction

bsuite is a collection of carefully-designed experiments that investigate core capabilities of a reinforcement learning (RL) agent with two main objectives.

  1. To collect clear, informative and scalable problems that capture key issues in the design of efficient and general learning algorithms.
  2. To study agent behavior through their performance on these shared benchmarks.

This library automates evaluation and analysis of any agent on these benchmarks. It serves to facilitate reproducible, and accessible, research on the core issues in RL, and ultimately the design of superior learning algorithms.

Going forward, we hope to incorporate more excellent experiments from the research community, and commit to a periodic review of the experiments from a committee of prominent researchers.

For a more comprehensive overview, see the accompanying paper.

Technical overview

bsuite is a collection of experiments, defined in the experiments subdirectory. Each subdirectory corresponds to one experiment and contains:

  • A file defining an RL environment, which may be configurable to provide different levels of difficulty or different random seeds (for example).
  • A sequence of keyword arguments for this environment, defined in the SETTINGS variable found in the experiment's sweep.py file.
  • A file analysis.py defining plots used in the provided Jupyter notebook.

bsuite works by logging results from "within" each environment, when loading environment via a load_and_record* function. This means any experiment will automatically output data in the correct format for analysis using the notebook, without any constraints on the structure of agents or algorithms.

We collate all of the results and analysis in a pre-made jupyter notebook bit.ly/bsuite-colab.

Getting started

If you are new to bsuite you can get started in our colab tutorial. This Jupyter notebook is hosted with a free cloud server, so you can start coding right away without installing anything on your machine. After this, you can follow the instructions below to get bsuite running on your local machine.

Installation

We have tested bsuite on Python 3.6 & 3.7. To install the dependencies:

  1. Optional: We recommend using a Python virtual environment to manage your dependencies, so as not to clobber your system installation:

    python3 -m venv bsuite
    source bsuite/bin/activate
    pip install --upgrade pip setuptools
  2. Install bsuite directly from PyPI:

    pip install bsuite
  3. Optional: To also install dependencies for the baselines examples (excluding OpenAI and Dopamine examples), run:

    pip install bsuite[baselines]

Environments

Complete descriptions of each environment and their corresponding experiments are found in the analysis/results.ipynb Jupyter notebook.

These environments all have small observation sizes, allowing for reasonable performance with a small network on a CPU.

Loading an environment

Environments are specified by a bsuite_id string, for example "deep_sea/7". This string denotes the experiment and the (index of the) environment settings to use, as described in the technical overview section.

For a full description of each environment and its corresponding experiment settings, see the paper.

import bsuite

env = bsuite.load_from_id('catch/0')

The sequence of bsuite_ids required to run all experiments can be accessed programmatically via:

from bsuite import sweep

sweep.SWEEP

This module also contains bsuite_ids for each experiment individually via uppercase constants corresponding to the experiment name, for example:

sweep.DEEP_SEA
sweep.DISCOUNTING_CHAIN

In addition, sequences of bsuite_ids with the same tag can be loaded via:

from bsuite import sweep

sweep.TAGS

The TAGS variable groups bsuite environments together by their underlying tag, so all the basic tasks or scale tasks can be loaded with:

sweep.TAGS['basic']
sweep.TAGS['scale']

Loading an environment with logging included

We include two implementations of automatic logging, available via:

We also include a terminal logger in logging/terminal_logging.py, exposed via bsuite.load_and_record_to_terminal.

It is easy to write your own logging mechanism, if you need to save results to a different storage system. See the CSV implementation for the simplest reference.

Interacting with an environment

Our environments implement the Python interface defined in dm_env.

More specifically, all our environments accept a discrete, zero-based integer action (or equivalently, a scalar numpy array with shape ()).

To determine the number of actions for a specific environment, use

num_actions = env.action_spec().num_values

Each environment returns observations in the form of a numpy array.

We also expose a bsuite_num_episodes property for each environment in bsuite. This allows users to run exactly the number of episodes required for bsuite's analysis, which may vary between environments used in different experiments.

Example run loop for a hypothetical agent with a step() method.

for _ in range(env.bsuite_num_episodes):
  timestep = env.reset()
  while not timestep.last():
    action = agent.step(timestep)
    timestep = env.step(action)
  agent.step(timestep)

Using bsuite in 'OpenAI Gym' format

To use bsuite with a codebase that uses the OpenAI Gym interface, use the GymFromDMEnv class in utils/gym_wrapper.py:

import bsuite
from bsuite.utils import gym_wrapper

env = bsuite.load_and_record_to_csv('catch/0', results_dir='/path/to/results')
gym_env = gym_wrapper.GymFromDMEnv(env)

Note that bsuite does not include Gym in its default dependencies, so you may need to pip install it separately.

Baseline agents

We include implementations of several common agents in the [baselines/] subdirectory, along with a minimal run-loop.

See the installation section for how to include the required dependencies at install time. These dependencies are not installed by default, since bsuite does not require users to use any specific machine learning library.

Running the entire suite of experiments

Each of the agents in the baselines folder contains a run script which serves as an example which can run against a single environment or against the entire suite of experiments, by passing the --bsuite_id=SWEEP flags; this will start a pool of processes with which to run as many experiments in parallel as the host machine allows. On a 12 core machine, this will complete overnight for most agents. Alternatively, it is possible to run on Google Compute Platform using run_on_gcp.sh, steps of which are outlined below.

Running experiments on Google Cloud Platform

run_on_gcp.sh does the following in order:

  1. Create an instance with specified specs (by default 64-core CPU optimized).
  2. git clones bsuite and installs it together with other dependencies.
  3. Runs the specified agent (currently limited to /baselines) on a specified environment.
  4. Copies the resulting SQLite file to /tmp/bsuite.db from the remote instance to you local machine.
  5. Shuts down the created instance.

In order to run the script, you first need to create a billing account. Then follow the instructions here to setup and initialize Cloud SDK. After completing gcloud init, you are ready to run bsuite on Google Cloud.

For this make run_on_gcp.sh executable and run it:

chmod +x run_on_gcp.sh
./run_on_gcp.sh

After the instance is created, the instance name will be printed. Then you can ssh into the instance by selecting Compute Engine -> Instances and clicking SSH. Note that this is not necessary, as the result will be copied to your local machine once it is ready. However, sshing might be convenient if you want to make local changes to agent and environments. In this case, after sshing, do

~/bsuite_env/bin/activate

to activate the virtual environment. Then you can run agents via

python ~/bsuite/bsuite/baselines/dqn/run.py --bsuite_id=SWEEP

for instance.

Analysis

bsuite comes with a ready-made analysis Jupyter notebook included in analysis/results.ipynb. This notebook loads and processes logged data, and produces the scores and plots for each experiment. We recommend using this notebook in conjunction with Colaboratory.

We provide an example of a such bsuite report here.

bsuite Report

You can use bsuite to generate an automated 1-page appendix, that summarizes the core capabilities of your RL algorithm. This appendix is compatible with most major ML conference formats. For example output run,

pdflatex bsuite/reports/neurips_2019/neurips_2019.tex

More examples of bsuite reports can be found in the reports/ subdirectory.

Citing

If you use bsuite in your work, please cite the accompanying paper:

@inproceedings{osband2020bsuite,
    title={Behaviour Suite for Reinforcement Learning},
    author={Osband, Ian and
            Doron, Yotam and
            Hessel, Matteo and
            Aslanides, John and
            Sezener, Eren and
            Saraiva, Andre and
            McKinney, Katrina and
            Lattimore, Tor and
            {Sz}epesv{\'a}ri, Csaba and
            Singh, Satinder and
            Van Roy, Benjamin and
            Sutton, Richard and
            Silver, David and
            van Hasselt, Hado},
    booktitle={International Conference on Learning Representations},
    year={2020},
    url={https://openreview.net/forum?id=rygf-kSYwH}
}

More Repositories

1

deepmind-research

This repository contains implementations and illustrative code to accompany DeepMind publications
Jupyter Notebook
13,132
star
2

alphafold

Open source code for AlphaFold.
Python
12,602
star
3

sonnet

TensorFlow-based neural network library
Python
9,769
star
4

mujoco

Multi-Joint dynamics with Contact. A general purpose physics simulator.
Jupyter Notebook
8,113
star
5

pysc2

StarCraft II Learning Environment
Python
8,001
star
6

lab

A customisable 3D platform for agent-based AI research
C
7,101
star
7

graph_nets

Build Graph Nets in Tensorflow
Python
5,352
star
8

graphcast

Python
4,517
star
9

open_spiel

OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games.
C++
4,231
star
10

alphageometry

Python
4,079
star
11

learning-to-learn

Learning to Learn in TensorFlow
Python
4,064
star
12

dm_control

Google DeepMind's software stack for physics-based simulation and Reinforcement Learning environments, using MuJoCo.
Python
3,793
star
13

acme

A library of reinforcement learning components and agents
Python
3,466
star
14

trfl

TensorFlow Reinforcement Learning
Python
3,136
star
15

dm-haiku

JAX-based neural network library
Python
2,848
star
16

alphatensor

Python
2,670
star
17

dnc

A TensorFlow implementation of the Differentiable Neural Computer.
Python
2,478
star
18

gemma

Open weights LLM from Google DeepMind.
Python
2,421
star
19

mctx

Monte Carlo tree search in JAX
Python
2,313
star
20

code_contests

C++
2,064
star
21

optax

Optax is a gradient processing and optimization library for JAX.
Python
1,670
star
22

kinetics-i3d

Convolutional neural network model for video classification trained on the Kinetics dataset.
Python
1,639
star
23

penzai

A JAX research toolkit for building, editing, and visualizing neural networks.
Python
1,639
star
24

mathematics_dataset

This dataset code generates mathematical question and answer pairs, from a range of question types at roughly school-level difficulty.
Python
1,621
star
25

educational

Jupyter Notebook
1,398
star
26

jraph

A Graph Neural Network Library in Jax
Python
1,349
star
27

rc-data

Question answering dataset featured in "Teaching Machines to Read and Comprehend
Python
1,285
star
28

mujoco_menagerie

A collection of high-quality models for the MuJoCo physics engine, curated by Google DeepMind.
Jupyter Notebook
1,278
star
29

tapnet

Tracking Any Point (TAP)
Jupyter Notebook
1,266
star
30

rlax

Python
1,223
star
31

scalable_agent

A TensorFlow implementation of Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures.
Python
981
star
32

android_env

RL research on Android devices.
Python
977
star
33

neural-processes

This repository contains notebook implementations of the following Neural Process variants: Conditional Neural Processes (CNPs), Neural Processes (NPs), Attentive Neural Processes (ANPs).
Jupyter Notebook
969
star
34

mujoco_mpc

Real-time behaviour synthesis with MuJoCo, using Predictive Control
C++
959
star
35

dramatron

Dramatron uses large language models to generate coherent scripts and screenplays.
Jupyter Notebook
947
star
36

tree

tree is a library for working with nested data structures
Python
925
star
37

materials_discovery

Jupyter Notebook
866
star
38

xmanager

A platform for managing machine learning experiments
Python
815
star
39

open_x_embodiment

Jupyter Notebook
785
star
40

chex

Python
751
star
41

ferminet

An implementation of the Fermionic Neural Network for ab-initio electronic structure calculations
Python
707
star
42

reverb

Reverb is an efficient and easy-to-use data storage and transport system designed for machine learning research
C++
700
star
43

funsearch

Jupyter Notebook
699
star
44

alphadev

Python
688
star
45

pycolab

A highly-customisable gridworld game engine with some batteries included. Make your own gridworld games to test reinforcement learning agents!
Python
659
star
46

concordia

A library for generative social simulation
Python
634
star
47

hanabi-learning-environment

hanabi_learning_environment is a research platform for Hanabi experiments.
Python
614
star
48

recurrentgemma

Open weights language model from Google DeepMind, based on Griffin.
Python
603
star
49

ai-safety-gridworlds

This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents.
Python
577
star
50

meltingpot

A suite of test scenarios for multi-agent reinforcement learning.
Python
576
star
51

ithaca

Restoring and attributing ancient texts using deep neural networks
Jupyter Notebook
547
star
52

dqn

Lua/Torch implementation of DQN (Nature, 2015)
Lua
546
star
53

uncertain_ground_truth

Dermatology ddx dataset, Jax implementations of Monte Carlo conformal prediction, plausibility regions and statistical annotation aggregation from our recent work on uncertain ground truth (TMLR'23 and ArXiv pre-print).
Python
534
star
54

distrax

Python
527
star
55

long-form-factuality

Benchmarking long-form factuality in large language models. Original code for our paper "Long-form factuality in large language models".
Python
526
star
56

surface-distance

Library to compute surface distance based performance metrics for segmentation tasks.
Python
526
star
57

tracr

Python
496
star
58

alphamissense

Python
494
star
59

dsprites-dataset

Dataset to assess the disentanglement properties of unsupervised learning methods
Jupyter Notebook
476
star
60

narrativeqa

This repository contains the NarrativeQA dataset. It includes the list of documents with Wikipedia summaries, links to full stories, and questions and answers.
Shell
452
star
61

clrs

Jupyter Notebook
444
star
62

lab2d

A customisable 2D platform for agent-based AI research
C++
420
star
63

dqn_zoo

DQN Zoo is a collection of reference implementations of reinforcement learning agents developed at DeepMind based on the Deep Q-Network (DQN) agent.
Python
406
star
64

alphastar

Python
403
star
65

dm_pix

PIX is an image processing library in JAX, for JAX.
Python
386
star
66

opro

official code for "Large Language Models as Optimizers"
Python
383
star
67

mathematics_conjectures

Jupyter Notebook
367
star
68

spriteworld

Spriteworld: a flexible, configurable python-based reinforcement learning environment
Python
367
star
69

torax

TORAX: Tokamak transport simulation in JAX
Python
361
star
70

dm_env

A Python interface for reinforcement learning environments
Python
343
star
71

dm_robotics

Libraries, tools and tasks created and used at DeepMind Robotics.
Python
341
star
72

spiral

We provide a pre-trained model for unconditional 19-step generation of CelebA-HQ images
C++
327
star
73

launchpad

Python
310
star
74

leo

Implementation of Meta-Learning with Latent Embedding Optimization
Python
302
star
75

enn

Python
291
star
76

streetlearn

A C++/Python implementation of the StreetLearn environment based on images from Street View, as well as a TensorFlow implementation of goal-driven navigation agents solving the task published in “Learning to Navigate in Cities Without a Map”, NeurIPS 2018
C++
285
star
77

gqn-datasets

Datasets used to train Generative Query Networks (GQNs) in the ‘Neural Scene Representation and Rendering’ paper.
Python
269
star
78

treescope

An interactive HTML pretty-printer for machine learning research in IPython notebooks.
Python
256
star
79

multi_object_datasets

Multi-object image datasets with ground-truth segmentation masks and generative factors.
Python
254
star
80

AQuA

A algebraic word problem dataset, with multiple choice questions annotated with rationales.
238
star
81

synjax

Python
238
star
82

grid-cells

Implementation of the supervised learning experiments in Vector-based navigation using grid-like representations in artificial agents, as published at https://www.nature.com/articles/s41586-018-0102-6
Python
236
star
83

card2code

A code generation dataset for generating the code that implements Hearthstone and Magic The Gathering card effects.
236
star
84

arnheim

Jupyter Notebook
235
star
85

torch-hdf5

Torch interface to HDF5 library
Lua
234
star
86

kfac-jax

Second Order Optimization and Curvature Estimation with K-FAC in JAX.
Python
231
star
87

dm_memorytasks

A set of 13 diverse machine-learning tasks that require memory to solve.
Python
221
star
88

Temporal-3D-Pose-Kinetics

Exploiting temporal context for 3D human pose estimation in the wild: 3D poses for the Kinetics dataset
Python
218
star
89

dm_alchemy

DeepMind Alchemy task environment: a meta-reinforcement learning benchmark
Python
197
star
90

neural_testbed

Jupyter Notebook
191
star
91

perception_test

Jupyter Notebook
184
star
92

jmp

JMP is a Mixed Precision library for JAX.
Python
183
star
93

neural_networks_chomsky_hierarchy

Neural Networks and the Chomsky Hierarchy
Python
183
star
94

xquad

180
star
95

nanodo

Python
180
star
96

pg19

179
star
97

spectral_inference_networks

Implementation of Spectral Inference Networks, ICLR 2019
Python
165
star
98

barkour_robot

Barkour Robot: Agile Quadruped Robots by Google DeepMind
C++
165
star
99

onetwo

Python
164
star
100

abstract-reasoning-matrices

Progressive matrices dataset, as described in: Measuring abstract reasoning in neural networks (Barrett*, Hill*, Santoro*, Morcos, Lillicrap), ICML2018
162
star