一种基于神经网络和Bradley-Terry模型的围棋等级分模型

A GO RATING MODEL BASED ON NEURAL NETWORK AND BRADLEY-TERRY MODEL

竞技类体育赛事和游戏等一般都依赖于等级分系统进行评价,现有等级分系统存在对局信息未得到充分利用问题。针对围棋对局数据的时间跨度大及包含让子棋局特点,基于经典成对数据比较模型(Bradley-Terry模型)构建神经网络等级分模型(NN-Rating)。通过历史衰减方法提高模型时效性,同时借鉴主场优势特性扩展模型处理让子棋局。在真实围棋比赛数据上的实验结果及其分析表明,NN-Rating模型有良好的客观性和稳定性,相比较Elo、Trueskill和Whole-History Rating(WHR)算法具有更高的预测准确性。

Competitive sports events and games generally rely on rating system,and existing rating system has problems such as insufficient use of the game information.In this paper,the neural network rating model(NN-Rating)is constructed based on the Bradley-Terry model of the classic paired data comparison model for the time span of Go game and the characteristics of Go that including handicap game data.We improved the timeliness of the model through historical decay method,and extended the model to make it possible to process handicap games.The experimental results and analysis on the real Go game data show that the NN-Rating model has good objectivity and stability,and has higher prediction accuracy than Elo,Trueskill and Whole-History Rating(WHR)algorithms.