From the basics to deep reinforcement learning, this repo provides easy-to-read code examples. One file for each algorithm. Please feel free to create a Pull Request, or open an issue!

Algorithms

Grid World (1-grid-world/)

1. Policy Iteration — 1-policy_iteration.py
2. Value Iteration — 2-value_iteration.py
3. SARSA — 3-sarsa.py
4. Q-Learning — 4-q_learning.py
5. Deep SARSA — 5-deep_sarsa.py
6. REINFORCE — 6-reinforce.py

CartPole (2-cartpole/)

7. DQN — 1-dqn.py
8. A2C — 2-a2c.py
9. PPO — 3-ppo.py

Atari (3-atari/)

10. DQN — 1-dqn.py
11. PPO — 2-ppo.py

Benchmarks

Trained on a MacBook Pro 14" (Apple M3, 8 GB unified memory), macOS 26.2, Python 3.11, PyTorch 2.11 with the MPS backend. CPU / GPU figures are read from Activity Monitor on the python3.11 process after the run has stabilized (~5 min in); peak RAM is the process's real memory at its high-water mark. Final score is the mean per-game return over the last 20 episodes of training.

Atari — Breakout (10M agent steps, ALE/Breakout-v5 with sticky actions)

Algorithm	Params	Train time	Final mean (per-game)	Peak RAM	CPU%	GPU%	W&B
DQN	1.69M	~9h	93.5 ± 9.6	5.27 GB	~60	~55	report
PPO	1.69M	~3.8h	261.9 ± 6.4	1.98 GB	~62	~55	report

Single seed per row, mean ± std over the final 20 logged episodes. Params counts only trainable network weights. CPU% is the single-process value reported by Activity Monitor (sum across cores, so >100% means multi-core use); GPU% is the same column for the Apple GPU. Sticky actions (repeat_action_probability=0.25) make absolute scores lower than the deterministic *-v4 environments often cited in older papers.

Atari — Montezuma's Revenge

Mac Studio (Apple M4 Max, 64 GB), ALE/MontezumaRevenge-v5, single seed. Two protocols, not cross-comparable: sticky-action RL vs deterministic restore-based search.

Method	Protocol	Score (single seed)	Frames	Link
PPO + RND (1-ppo-rnd.py)	sticky, RL policy	~3,120	65M	report
Go-Explore — exploration (2-go-explore.py)	deterministic search	31,000 (replay-verified)	500M	run
Go-Explore — robustification (3-robustify.py)	sticky, RL policy	— (no from-reset score)	5M	—

RND (Burda et al. 2018): first key ~327k steps with 512 envs (128 never scored in 50M — parallel breadth is the lever); plateaued above the PPO baseline 2497, below RND's 8152 (~30× more experience). Exploration (Ecoffet et al. 2019/2021): best end-of-episode trajectory from a knowledge-free cell archive, no NN — a search result, not an RL score (Nature ref 24,758). Robustification (backward algorithm, Salimans & Chen 2018): bootstraps with a first-key demo + 128 envs but the curriculum plateaus ~22% on one machine — no from-reset score, a scale ceiling vs the original hundreds–thousands of envs.

Setup

Requires Python 3.11 and uv.

git clone <this repo>
cd reinforcement-learning
uv sync

Running

# Grid World
cd 1-grid-world && uv run python 3-sarsa.py

# CartPole — train
cd 2-cartpole && uv run python 1-dqn.py

# CartPole — watch training (slower)
cd 2-cartpole && uv run python 1-dqn.py --render

# CartPole — replay a trained checkpoint
cd 2-cartpole && uv run python 1-dqn.py --test

Logging to Weights & Biases (Atari only)

Both Atari scripts (1-dqn.py, 2-ppo.py) can stream training metrics to your own Weights & Biases account. One-time login, then pass --wandb:

uv run wandb login   # paste the API key from https://wandb.ai/authorize
cd 3-atari && uv run python 2-ppo.py --env breakout --wandb
cd 3-atari && uv run python 1-dqn.py --env breakout --wandb

Runs land in your rl-atari-ppo / rl-atari-dqn project — nothing is shared by default. Omit --wandb and the script runs without ever touching the network.

Updates

Modernized from the 2017 original:

• Framework: Keras + TensorFlow 1.0 → PyTorch 2.11
• Env: gym 0.8 → gymnasium 1.2
• Rendering: tkinter → pygame (cross-platform with no system Tk)
• Tooling: requirements.txt → pyproject.toml + uv
• Scope: pruned to 9 core algorithms; dropped Monte Carlo / DDQN / A3C / Atari / mountaincar; added PPO
• Layout: flat 1-grid-world/3-sarsa.py instead of nested 1-grid-world/4-sarsa/sarsa_agent.py
• Docs: each algorithm file now opens with a paper citation and the core update equation