Deep Dive on Checkers Endgame Data

For games such as checkers and chess, large endgame databases/tablebases have been constructed to capture the perfect win/loss/draw value for positions near the end of the game. Such databases/tablebases can be used to enhance game-playing performance. However, this approach quickly runs into computational and storage resource limitations. An enticing alternative is to learn from such data and apply the learned evaluation to even larger data sets through transfer learning. This paper reports on research that uses deep learning to a) correctly learn a high percentage of checkers endgame positions; b) learn patterns that can be used for transfer learning; c) demonstrates that learning from a small sample of a large data set is an efficient way to compute a neural net evaluation that achieves most of the benefits; and d) shows that dynamically choosing between neural network prediction and using it in a one-ply search yields about 96% prediction accuracy.