This repository provides the source code for the O'Reilly Japan book "Deep Learning from Scratch 4: Reinforcement Learning". It offers practical implementations of reinforcement learning algorithms for researchers and practitioners looking to build a foundational understanding and apply these techniques.

How It Works

The project is structured by chapter, with Python scripts demonstrating core concepts like bandit problems, dynamic programming, Monte Carlo methods, TD learning, Q-learning, DQN, and policy gradients. It utilizes the custom DeZero framework developed in previous series books, with optional PyTorch implementations also provided. This approach allows for a deep dive into the mechanics of RL algorithms without abstracting away essential details.

Quick Start & Requirements

• Install DeZero: pip install dezero
• Prerequisites: Python 3.x, NumPy, Matplotlib, OpenAI Gym.
• Code execution: Run Python scripts from respective chapter directories (e.g., python ch01/avg.py).
• Jupyter Notebooks are available for cloud execution (Google Colab, Kaggle Notebooks).

Highlighted Details

• Companion "100 Reinforcement Learning Problems" for verification.
• Jupyter Notebooks available for cloud environments.
• PyTorch implementations provided in the pytorch folder.
• Official errata available on GitHub wiki.

Maintenance & Community

The repository is maintained by oreilly-japan. Community interaction and error reporting are encouraged via email to japan@oreilly.co.jp.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive for commercial and non-commercial use.

Limitations & Caveats

The code is tied to the specific DeZero framework from the book's series, which may require understanding its internal workings. While PyTorch versions are available, the primary focus is on DeZero.