基于多模型堆叠与特征提取的二打一叫牌算法研究
Bid recommendation model for Fighting the Landlord based on multi-model stacking and feature extraction
刘航 1丁濛 1李淑琴1
作者信息
- 1. 北京信息科技大学 计算机学院, 北京 100101;感知与计算智能联合实验室, 北京 100101
- 折叠
摘要
针对现有二打一叫牌决策研究中分类粒度低的问题,提出了一种基于多模型堆叠与关键特征提取的叫牌模型训练方案.具体来说,设计了手牌特征构建方法,即由手牌向量、牌型特征、手牌工整度、最小出牌步数以及组合丰富度共同构成玩家手牌特征;在此基础上,提出了使用堆叠法融合4类基模型的决策结果,并训练二层模型Cat-Boost给出最终决策.实验结果表明:相较于仅使用手牌向量特征,该特征构建方式可显著提升模型性能.经过堆叠法融合多模型决策后,二层模型准确率进一步提升,最终在测试集上达到84.3854％的精准度.所提方法可为其他牌类游戏的叫牌博弈决策提供参考.
Abstract
Addressing the granularity limitation observed in existing research on"Fighting the Landlord"bidding decision problem, this paper proposes an approach for training a Bid Recommendation Model.Specifically, a methodology is devised for constructing hand features, including hand vector, hand pattern features, hand tidiness, minimum step in card play, and combination richness.Based on this, we propose a stacked approach to fuse the decision results of four base models and train a meta classifier CatBoost as the final decision model for the bidding decision.Our experimental results indicate that, in comparison with relying solely on hand vector features, this feature construction method significantly enhances model performance.Following the fusion of multiple model decisions through stacking, the accuracy of the second-layer model is further improved, achieving a precision of 84.3854％ on the test set.Moreover, this method provides some references for bidding decision in other card games.
关键词
二打一/叫牌算法/计算机博弈/集成学习Key words
Fighting the Landlord/bid algorithm/game algorithm/ensemble learning引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据