American Checkers

Also, known as English Draughts, this game features an 8x8 board, 24 total board pieces, and 4 types of pieces. Along with the rules, this means that there are 500,995,484,682,338,672,639 legal game states.

Rules:

• Regular pieces can only move diagonally forward
• To capture, a piece jumps over an adjacent enemy piece
• If a capture can be made, it must be done
• Multiple captures can be performed in a single turn, but only with one piece
• King pieces can move both forward and backward
• Regular pieces are crowned Kings, if they reach the other side of the board
• A player loses by having all of their pieces captured or not moving during their turn

This AI game agent is a multi-layer perceptron (deep neural network) model. It was initially trained by a weakly supervised learning program, that generated batches of game states for the model to study. Afterwards the model was given to a reinforcement learning program, in which the model improved by playing against itself 100,000 times.

The AI demo above and reinforcement learning program do not use any sort of look-ahead algorithm, like MinMax. Otherwise, I fear my laptop would explode. The AI demo above looks-ahead 2 turns, using the MinMax decision algorithm. No longer uses the simple, but slow, MinMax calculation. Instead, the MLP was retrained.