v1.07 of an AI of Drop2048 v1.08. Original project (by @colpafu, account seems to be deactivated): https://scratch.mit.edu/projects/643484547 Just click start and let it play automatically. (You can use turbo mode to run faster) For reference, the maximum locally tested score from a single AI is 7768. This seems like that the uncontrollable behavior of the random number generation sets a limit to the algorithm. The data have been trained locally using machine learning algorithms.
v1.07: Found a bug in the training algorithm, so the data are retrained and replaced. v1.06: Fixed a bug in the surface & tsurface calculation implementation in Scratch Currently it determines the next move based on 12 parameters (feel free to suggest other useful parameters): "cell": number of tiles on the board "skyline": the total change of the height between every two consecutive columns "linkage": the number of sticked pairs in the board "diff": the sum of the difference between every two neighboring tiles "surface": the number of unique tiles that are not completely blocked "max": the largest tile "sum": the sum of all tiles "full": the number of columns which is full "logscore": ln(score + 1) after the move "score": the score after the move (replaced by "logscore" for later training) "tsurface": the number of different tiles that is connected to the bottom tile "top": the height of the current board
