Learning Cube Strategy in Backgammon with Neural Networks

摘要

While neural networks have been used to study checker play in backgammon, the decision of giving or taking the doubling cube, especially in relation to the match score, has been largely considered an orthogonal problem. We integrate the doubling cube and match scoring into the network. In doing so, our network learns how differences in match scores and cube circumstances influence not just further cube decisions, but checker play as well. We find that by incorporating the doubling cube situation into the input of the neural network, we were able to statistically outperform networks without this feature.