The decision
A choice over the legal moves each turn. The state is the board, described as text: positions, the score and what just happened.
How to build it
- Write the game loop, or use an existing Snake, Tetris or Connect Four implementation.
- Each turn, describe the board and list the legal moves as options.
- Take the most probable move, and draw the probabilities on screen: it makes a great demo.
- Record games, and count wins and survival time.
Make it better
Decider's authors show it playing Tetris, Breakout, Pong, Snake and Connect Four, and NanoJev was trained on game environments. Compare the two, then have a heuristic shortlist the candidate moves so the model only chooses among good ones.