Research

The lab notebooks

Behind Spacewar: Orbital Duel sits a small machine-learning lab, and these are its notebooks: real training runs and real failures, written up as they happened. The two posted so far are the game's small networks, built to show the whole pipeline from first principles to a shipped artifact. The plant identifier and the river lab are not written up yet.

Every experiment is written at three depths, and you choose per chapter: the plain story anyone can read; ▸ FOR THE LEARNERS, which rebuilds the machinery from first principles for any scientifically curious reader (no machine-learning background assumed); and ▸ FOR THE DATA SCIENTISTS, the working notebook in the field's own language, with references to go deeper still.

Notebook 01 · the pilot

Half a worm, flying a spaceship

A 6,220-weight network (half a worm's brain) starts with zero knowledge and teaches itself to fly, fight, cheat, get caught, and bend lasers around the black hole. With the three-minute film.

6,220 params · reinforcement learning · self-play · sim-to-real

Notebook 02 · the eye

The net that watches the screen

The pilot got its senses served on a plate. Real systems get pixels. A 28,000-parameter network watches raw game frames and paints every ship, every laser, and the hole. Then it gets squeezed into 8-bit integers you can't tell from the original.

28k params · convolutional segmentation · int8 · pixels in, game out

The code · the pilot's lab

Train the pilot yourself

Everything in the pilot's notebook is MIT-licensed and runs on an ordinary laptop: the physics, the trainers, six graded experiments, and every era's champion ready to replay. Watching a match in your browser takes under a minute and no install.

open code · MIT · github.com/stonej-gh/ml-toys

The code · the eye's lab

Retrain the eye yourself

The spotter's whole pipeline ships in the same repo: a renderer that labels its own frames, about five minutes of training on a laptop, and the frozen golden bundle any port can prove itself against, down to the last bit.

open code · MIT · github.com/stonej-gh/ml-toys