Cody M. Lewis

Attempting to Machine Learn 2048

So I figured that 2048 would be a reasonably simple game to apply machine learning techniques to win, it seems I was quite wrong.

I first tried a genetic algorithm, the genome will just be the loop of movements and the fitness function is the points acquired. Although, from training I found that the algorithm would always get stuck in a local minima of 4 moves. Interestingly those 4 moves would always be a direction then the two directions perpendicular to the first. This strategy would only achieve a 512 block on the board, albeit being most of the time.

Since the genetic algorithm did not work so well, I decided to try an online learning setup with a neural network. The network would repeatedly play games and be taught what moves are possible for the current board state. The intention was that given any game state it would be able to figure out at least a move that works. This technique failed as the network would always get stuck at new game states it finds, and would only ever achieve 512 block on the board.

RSS ⬆ Back to top