bsuite is a benchmark suite for evaluating core reinforcement learning (RL) agent capabilities. It provides a standardized set of carefully designed experiments to facilitate reproducible research and the development of more general and efficient RL algorithms. The target audience includes RL researchers and practitioners seeking to rigorously assess and compare agent performance across fundamental challenges.

How It Works

bsuite comprises a collection of RL environments, each defined with specific configurations and analysis plots. Experiments are instrumented to log results automatically using load_and_record* functions, outputting data in a format compatible with a provided Jupyter notebook for analysis. This approach ensures that any agent can be evaluated without structural constraints, simplifying the process of benchmarking and comparison.

Quick Start & Requirements

• Install via pip: pip install bsuite
• For baseline examples: pip install bsuite[baselines]
• Tested on Python 3.6 & 3.7.
• Colab tutorial available: bit.ly/bsuite-colab

Highlighted Details

• Environments feature small observation sizes, enabling reasonable performance on CPUs with small networks.
• Supports logging to CSV, SQLite, or terminal, with custom logging easily implementable.
• Includes a utility to wrap environments for OpenAI Gym compatibility.
• Provides baseline agent implementations and scripts for running experiments, including parallel execution and Google Cloud Platform integration.

Maintenance & Community

• Developed by Google DeepMind.
• Accompanying paper available: https://openreview.net/forum?id=rygf-kSYwH

Licensing & Compatibility

• Apache 2.0 License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The project specifies compatibility with Python 3.6 & 3.7, suggesting potential issues with newer Python versions. Baseline dependencies are not installed by default, requiring explicit installation for users who wish to utilize them.