Tetris × ruvector: watch a machine learn Tetris

Game speed ×1

The red dashed outline is where the brain will drop the piece.

Lab notebook

Learner: cross-entropy method. Each generation, 10 brains play the same 2,000-piece game. The 3 best set the next brain.

Training speed —

ruvector features

ruvector stores the brains. Reusing the best ones is what speeds up the next run. A plain list would do the same, and the similarity-search ideas tested so far did not help.

Vector memoryStores every tested brain in a ruvector HNSW index (@ruvector/wasm). Warm startA new run begins from the best brains in memory, not from scratch. Remember across reloadsSaves the archive in this browser and rebuilds the index on load.

Fig. 1 Lines cleared by each generation’s brain (max 800)

Fig. 2 This generation, live

Fig. 3 What the brain values | = start

Fig. 4 Memory: every brain tested, mapped by similarity ⊕ current brain