alessiodm/drl-zh

Deep Reinforcement Learning: Zero to Hero!

What it solves

Reading about reinforcement learning (RL) is easy; actually getting algorithms to work is hard. This project turns that gap into a hands-on path: you learn by building the algorithms yourself, not just by following textbook explanations. It gives you a structured course where every key idea comes with an exercise notebook and a full solution to check against when you're stuck.

How it works

The course is a series of Jupyter notebooks numbered from 00 to 18, grouped into tracks like Foundations, Breaking assumptions, Planning, Modern AI stack, Production, World models, and Meta + wrap-up. The root notebooks contain guided TODO sections where code is intentionally removed; the solution/ folder holds complete, runnable versions so you can unblock yourself anytime. A Docker workspace bundles everything you need (code-server, Python 3.13, Jupyter, dependencies, and the optional DRL-ZH AI Companion, a VS Code extension that gives Socratic hints instead of spoilers, with text or voice input). You can also set up natively using the manual instructions.

Who it's for

This is built for people who are already comfortable with Python, know PyTorch basics, and have the usual ML math background (probability, statistics, linear algebra, derivatives). The notebooks teach RL, but they assume you can read and modify real training code, so it's for learners who want to write the code themselves rather than consume pre-built demos.

Highlights

  • Build over 18 notebooks worth of RL algorithms, from MDPs and tabular RL to DQN, PPO, SAC, AlphaZero, RLHF, Decision Transformers, and Dreamer.
  • Exercises with guided TODOs and full solutions in separate notebooks for self-pacing.
  • Optional AI Companion (VS Code extension) offers Socratic hints with text or voice, supporting Gemini (default), OpenAI, Anthropic, or Groq.
  • Docker-based setup ensures a reproducible environment with GPU and CPU variants, plus a native setup path.
  • Covers production practices like TensorBoard, checkpointing, debugging, multiple seeds, Ray, and Optuna.

Related

  • Project
    galilai-group/stable-worldmodel

    A platform for reproducible world model research and evaluation that provides a unified interface for data collection, training, and model-predictive control evaluation.

  • Project
    Vespa314/chan.py

    chan.py is a Python 3.11 library that implements the Chan‑Theory technical‑analysis framework, offering calculations of strokes, segments, central zones, and buy‑sell points, plus feature engineering, ML model integration, back‑testing, and live‑trading connectors (Futu, data APIs, MySQL/SQLite). It includes plotting utilities for static and animated visualisation.

  • Project
    feature-engine/feature_engine

    A Python library providing Scikit-learn compatible transformers for comprehensive feature engineering and selection to improve machine learning model performance.

  • Project
    dstackai/dstack

    dstack is a unified control plane that orchestrates GPU resources across cloud providers, Kubernetes, and on-premise clusters to streamline AI development, training, and inference.