Learning of evaluation function and prediction of winning probability of curling based on score difference and ends remaining

Abstract

In curling, strategy is as important as the technique of throwing stones to the desired spot. To evaluate this strategy, a platform called Digital Curling exists that simulates curling on a computer. It is expected that the strategy tested in the simulator will be applied to real curling, and research into curling AI is actively being conducted. In this research, we created a curling AI using this digital curling platform. The proposed model uses a neural network as a position evaluation function to predict the winning rate. In previous models, neural networks were trained with the aim of maximizing the score at each end, and thus cannot make the optimal prediction to win in some game situations. In the proposed model, the score difference and the number of remaining ends are input, and the learning data is also created taking these two factors into consideration. This is expected to make it possible to predict optimal winning rates throughout the game. As a result, in the last shot of the end, proposed model can predict winning rate in situation where the best move would be a blank end, which is a situation that previous models cannot predict. In future research, I would like to predict the winning rate over the entire game by having the model learn all shots from the end.