A simple bejeweled blitz bot written in python. Checkout the video.
For training the block image classifier,
$ python learn_board.py
This drops a file named trained.
At the beginning of the game, run the following script.
$ python bot.py
Note that the game screen has to meet the exact position the program requires.
Recognizing the board state is arguably the most difficult part. Each block image spins, changes colors, gets covered by scores etc. These changes makes it very difficult to figure out the color/shape of a block.
For this program, SVM was used for the classification of blocks. Since a block consists of 40 x 40 pixel values, this can be divided into 10 x 10 pixel boxes. Each block is divided into 4 x 4 pixel sub-block and among those 16 sub-blocks, middle 4 blocks are used to represent the entire block. The RGB value of the sub-block is averaged over to represent that sub-block. Therefore, a single block is represented by four features. Or rather, four-value tuple of a single feature.
There are 7 different kinds of base blocks (blue, green, yellow, white, red, purple, orange). Any block that is not stable (moving, being destroyed, etc,.) is classified as the 8th state.
It is worth mentioning that we need more data when the board state is unstable. Without such data, the classifier classifies invalid blocks as valid blocks, which in turn, affects how valid blocks are being classified as well.