Reinforcement Learning applied to Tetris

A school project with a classmate, and one of the more purely enjoyable things I have built.

Tetris is a good teaching problem for reinforcement learning because the reward is so easy to get wrong. Score per line cleared sounds obvious, and produces an agent that stacks recklessly and dies. The learning is in the shaping: how much to punish height, how much to punish holes, and how much to leave alone so the agent can discover something you did not think of.

Beat the bot

The agent below is the real one. Its network is tiny — four inputs, two hidden layers of 32, one output — so the trained weights are exported straight from the Keras checkpoint to plain JSON and run in your browser with no ML runtime.

It never chooses individual key presses. For each piece it tries every rotation and column, measures the board that would result (lines cleared, holes, bumpiness and total height) and asks the network how good that board is. The dashed outline is the placement it picked, and V̂ is the value it expects from it. It sees only the current piece and does no lookahead.

You both get the same pieces in the same order. The speed setting only changes how fast it moves; its choices are the same at every speed.

Playable

Beat the bot

Same pieces, two minutes, most lines wins.

Show the agent's chosen placement

You: 0; lines · 0 pts
Clock: 2:00; Level
Agent: 0; lines · 0 pts

Younext

AgentV̂ —next

Loading the agent…

← → move · ↑ / X rotate · Z rotate back · ↓ soft drop · Space drop · P pause

seed -----

Play it full screen →

Have something hard that has to work?

Contract work, a full-time role, or a problem you want a second opinion on — I read every message and reply to the ones that aren't a template.

Usually replies within a day or two · St. Louis, Missouri