Stick Fight Arena
Two stick figures brawl on a procedurally generated street, and every move either one makes comes from a small neural network bred by evolutionary self-play.
A neural-network proof of concept dressed up as a fist fight. Each fighter is driven by a feed-forward network — 40 perception inputs, a 40-24-16-12 tanh topology — with a tactical state machine and personality weights on top, and a "comedy director" nudging things toward slapstick. The networks are bred rather than scripted: evolutionary self-play, fitness curves tuned against degenerate strategies, recorded lineage across generations, round-robin tournaments with Elo ratings, and versioned weights that can be rolled back. The bodies are a fixed-timestep sim of 14-segment articulated ragdolls with spring-damper joints and impulse resolution, and you can shove them around with the mouse mid-fight. TypeScript and Canvas 2D over a PHP 8.2 / MariaDB back end; deterministic replays, weather, a spectator crowd and a matrix-mode filter on top. This link opens the arena with the network activity and a live Sankey view of each fighter's decisions drawn alongside the fight, an untrained scribbler put up against an evolved one.