m-rr-j.github.io

Source for m-rr-j.github.io — Maxime Jousset's notes on reinforcement learning, self-supervised representation learning, and transformers, built from scratch one experiment at a time.

The blog is the main event: tutorial-style write-ups of an RL-from-scratch series that trains a Super Mario Bros agent one algorithm at a time — DQN → PPO → ICM → RND → kNN novelty — each model earning its place by fixing a limitation of the last.

The code the posts describe lives in two companion repos:

• rl-factory — the environment-agnostic training engine (adding an algorithm is registering a learner, not rewriting a training loop).
• rl-ablations — the five models, on Mario, that plug into it.

Build locally

bundle install
bundle exec jekyll serve

Built with the al-folio Jekyll theme (MIT); see LICENSE.