AI to play the 2048 game. Deep learning, self-teaching.
This is a fun project, inspired by Alpha-Go, to create a 2048 AI that plays against itself, training a neural network to evaluate board positions.
I love the concept of emerging intelligence, without providing human wisdom.
There are more detailed notes in the wiki.
- Get decent performance without the AI, using a set algorithm to evaluate board positions.
- Get a reasonable degree of optimisation
- Experiment with various options to do with tree pruning, result caching
- Automate the cross-over between game and neural net. Each set of games generates a set of board positions and labels, which trains the AI, which improves for the next set of games
- Optimise AI, try different architectures etc.
Basic gameplay in place. Can play several games in parallel and compile stats. Basic un-optimised heuristic to judge board positions. No pruning in place for expectimax algorithm... maximum depth is about 4.
No deep learning added yet.
Suggest getting the AI to predict the number of moves left before death. Perhaps log(moves).
When evaluating in the expectimax tree, each combination of boards can be a new training board.