Baylus / 2048I

Teaching ML to play 2048 better
MIT License
0 stars 0 forks source link

Add replay functionality #18

Open Baylus opened 1 month ago

Baylus commented 1 month ago

Currently, there are some issues with the training, where the model is not getting any better at the game. There are a ton of timeouts, which for the score really shouldn't be the case, since a low score timeout means that the model was trying to make moves that did not affect the board, which are moves we definitely want to avoid as much as possible.

Important parts for this change will be to visualize the board states, but also the moves being made, to identify when a board state did not change, thus resulting in a longer game than necessary. The reason we even have this turn limit is because the model could get stuck trying to constantly make the same move over and over again, which would mean the training comes to a grinding halt. I could attempt to disable the turn limit and just see exactly what happens if I do let it continue on.

Key functions: