基于深度学习的MOBA类游戏胜率预测模型的研究

Research on real-time winning rate prediction model of MOBA game based on deep learning

多人在线战术竞技(MOBA)游戏是目前最为火热的竞技游戏之一,根据阵容以及比赛中的实时数据预测比赛的胜率正在成为该领域的研究热点方向。为提高MOBA游戏胜率预测的准确率,本文探索基于深度学习的组合预测方法,提出了MOBA游戏实时胜率预测模型。本文根据比赛双方所选的英雄信息以及比赛实时数据,利用双层LSTM模型并引入注意力机制的方法训练实时胜率预测模型。实验使用了近3 w场Dota2比赛数据集构建训练集和测试集。在实时胜率预测实验中,分别在比赛进行到第10、20、30、35 min时使用实时预测模型的预测胜率,准确率分别达到68.5%、71.8%、85.6%、88.4%,对比其他深度学习模型,准确率平均提高了1.5%以上。实验表明,建立的MOBA游戏实时胜率预测模型具有较高的预测准确率。

Multiplayer online tactical competitive(MOBA) game is one of the currently hottest competitive games. Predicting the winning rate of the game according to the lineup and real-time data in the game is becoming a hot research direction in this field. In order to improve the accuracy of MOBA game winning rate prediction, this paper explores the combination prediction method based on deep learning, and puts forward the real-time winning rate prediction model of MOBA game. According to the hero information selected by both sides of the game and the real-time data of the game, this paper uses the double-layer LSTM model and introduces the attention mechanism to train the real-time winning rate prediction model. The experiment used nearly 30 000 Dota2 game data sets to construct training sets and test sets. In the real-time winning rate prediction experiment, the accuracy of using the real-time prediction model at the 10, 20, 30and 35minutes of the game is 68.5%, 71.8%, 85.6% and 88.4% respectively. Compared with other deep learning models, the accuracy is improved by more than 1.5% on average. Experiments show that the established MOBA game real-time winning rate prediction model has high prediction accuracy.