基于深度强化学习的和声自动生成算法  

Harmony Autogeneration Algorithm Based on DeepReinforcement Learning

作　　者：刘至洋 刘维莎 冉黎琼 徐康镭 蒋宇河 李庆 乔少杰 LIU Zhiyang;LIU Weisha;RAN Liqiong;XU Kanglei;JIANG Yuhe;LI Qing;QIAO Shaojie(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China;Department of Applied Mathematics,University of Washington,Seattle 98105,USA)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]华盛顿大学应用数学系,华盛顿西雅图98105

出　　处：《无线电通信技术》2024年第5期985-992,共8页Radio Communications Technology

基　　金：四川省科技创新苗子工程项目(MZGC20230076);国家自然科学基金(62272066);四川省科技计划(2021JDJQ0021,2022YFG0186,2022NSFSC0511,2023YFG0027);教育部人文社会科学研究规划基金(22YJAZH088);宜宾市引进高层次人才项目(2022YG02);成都市“揭榜挂帅”科技项目(2022-JB00-00002-GX,2021-JB00-00025-GX);成都市技术创新研发项目(重点项目)(2024-YF08-00029-GX);成都市区域科技创新合作项目(2023-YF11-00020-HZ)。

摘　　要：机器自动作曲是人工智能和作曲理论结合的交叉研究领域。自动作曲算法旨在辅助使用者进行音乐创作,帮助使用者减少工作量或是提供灵感。面对当代音乐创作需求,现有的自动作曲方法无法有效表达音乐特征,生成的和声缺乏音乐结构,无法满足当代音乐创作多样性的要求。针对上述不足,提出了一种基于深度强化学习(Deep Reinforcement Learning,DRL)与和声量化的和声自动生成(Automatic Harmony Generation Algorithm Based on DRL,AHG-DRL)算法。AHG-DRL使用和声量化方法对音乐进行编码,使编码后的音乐具备更加全面的音乐特征;使用基于逆强化学习(Inverse Reinforcement Learning,IRL)与DRL的和声生成算法扩大音乐创作的搜索空间,同时使生成的音乐具备功能性。实验结果表明,所提和声自动生成算法可以生成满足创作需求和符合作曲规则的音乐,与其他和声生成算法相比,生成的和声种类更加复杂多样,在客观评价指标上更接近真实值。Machine-automated composition is a cross-field research area that combines artificial intelligence and composition theories.The automated composition algorithm aims to assist users in music creation and helps users reduce workload or provides inspiration.Faced with the demand for contemporary music creation,existing automatic composition methods fail to effectively express musical features,and the produced harmonies lack musical structures,failing to meet the diversity of contemporary music creation.To solve the aforementioned problems,a harmonic automatic generation algorithm based on Deep Reinforcement Learning(DRL)and harmonic quantization is proposed,called Automatic Harmony Generation Algorithm Based on DRL(AHG-DRL).AHG-DRL uses a harmonic quantization method to encode music,making the encoded music have more comprehensive musical features.It also uses a harmony generation algorithm based on Inverse Reinforcement Learning(IRL)and DRL to expand the search space of music creation while making the generated music have functionality.Experimental results indicate that the proposed automatic harmony generation algorithm can generate music that meets creative requirements and conforms to composition rules.When compared with other harmony generation algorithms,it can generate more complex and diverse types of harmonies,which approximates to the ground-truth values in the objective e valuation metrics.

关 键 词：和声量化 深度强化学习 自动作曲 预训练 逆强化学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]