Inspired by the advancements of board game AI using deep reinforcement learning, e.g. AlphaGo, I applied the technique to a different game close to my heart — Tetris. In particular, tackling the piece-efficiency challenge in “dig” modes.
This is a fun personal project with a very experimental technology that is S4TF (Swift for TensorFlow). In a big Notebook, I implemented most of the basic components, including game mechanics, MCTS (Monte-Carlo Tree Search) algorithm, game record saving, and training the convolutional network.
At this point, the AI still doesn’t play very well yet. I plan to revisit this project after S4TF sees better support for saving and loading models. The next step would be to refactor the code out of the Notebook format, improve performance, and set up for more intensive self-play training.
You can find the Colab Notebook here.