AI Tetris — evolving a strategy
Tetris played by an evolving agent. It scores the board by features — height, holes, bumpiness — and evolution tunes their weights on its own.
Tetris is about long-term strategy, not reflexes. I was curious whether evolution could handle that kind of multi-move planning — and what board representation it needs to do it.
What decides here isn’t a deep network but the board features: aggregate height, number of holes, surface bumpiness, cleared lines. Evolution tunes the weights of these heuristics — essentially rediscovering something close to Pierre Dellacherie’s classic agent.
Naively brute-forcing every rotation × position to score a move stalled the rendering. I rewrote the scoring into a single pass over the board and limited how far ahead the agent looks — and it went smooth.
Strategy, not reaction
Unlike the dino and the snake, this agent doesn’t react to a frame — it evaluates the future board: for every possible move it computes a set of features and picks the best one. Evolution tunes how harshly to penalize holes and height relative to cleared lines.
The simulation is isolated in a sandbox iframe, like the rest of the lab’s AI experiments.