遗传算法在旋律创作中的应用研究

发布时间：2018-05-07 17:00

本文选题：算法作曲 + 遗传算法　；参考：《河南师范大学》2014年硕士论文

【摘要】：随着计算机技术的不断发展，自动化技术已被广泛应用于旋律创作领域。算法作曲的出现，使得旋律创作变得容易、方便。作为算法作曲的方法之一，遗传算法能够灵活地产生各种不同风格的旋律。但是，应用遗传算法进行旋律创作时，适应度函数的构建仍是一大挑战。以往的适应度函数大多数很主观而且容易受设计者的偏见影响，评估过程耗费时间长且效率低下。本文着重研究将遗传算法应用于旋律创作时适应度函数的设计及其他相关问题。首先，分析了将遗传算法应用于旋律创作时生成初始群体常用的几种人工智能方法，包括数学模型的方法、人工神经网络的方法、音乐文法的方法等。为了实时产生旋律，本文中采用基于数学模型的方法来生成初始群体。并且为了避免生成的群体过度随机，在生成初始群体时对相关旋律参数进行了限制，如调号、拍号等。其次，设计了一种新的旋律编码方式，即以旋律中的每个音符为单位，基于该音符的音级、时值和符号对旋律进行编码。该编码方法可以直观形象地表现旋律的信息，而且能够保证旋律信息的正确性。再次，，构建了旋律特征提取器，定义了用于自动评价的多目标适应度函数。特征提取器用于提取旋律乐谱的特征，如旋律中各音符的自相似性，旋律的图形，音符的线性度，所使用的音域，等等。将这些特征进行线性组合作为多目标适应度函数，对所有生成的旋律个体进行适应度计算，根据计算得到的适应度函数值决定个体是否需要进化。然后，对进化过程中的交叉和变异操作作了研究，以使旋律的进化能够顺利进行。交叉操作时着重考虑了音符的时值修正问题，确保交叉前后每一小节中音符的时值之和保持不变。执行变异操作时，对于音级的变异，保证相邻音符的音程不超过8度。最后，搭建了基于多目标适应度函数的遗传算法旋律创作实验系统平台。实验结果表明，该系统在确保生成具有较高质量的旋律的前提下，缩短了评估所用的时间，大大提高了旋律创作的效率。
[Abstract]:With the development of computer technology, automation technology has been widely used in melody creation field. The emergence of algorithm composition makes melody creation easy and convenient. As one of the algorithms to compose music, genetic algorithm can produce various melodies flexibly. However, the construction of fitness function is still a challenge when using genetic algorithm to create melody. Most of the previous fitness functions are subjective and susceptible to designer bias, and the evaluation process is time-consuming and inefficient. This paper focuses on the application of genetic algorithm to the design of fitness function in melody creation and other related problems. Firstly, several artificial intelligence methods which are commonly used to generate initial population when genetic algorithm is applied to melody creation are analyzed, including mathematical model method, artificial neural network method, music grammar method and so on. In order to generate melody in real time, the method based on mathematical model is used to generate initial group. In order to avoid the excessive random of the generated group, the related melody parameters, such as tone, racquet, etc, are restricted when the initial group is generated. Secondly, a new melodic coding method is designed, in which each note in the melody is taken as the unit, and the melody is encoded based on the tone level, time value and symbol of the note. This coding method can express melody information intuitively and guarantee the correctness of melody information. Thirdly, a melody feature extractor is constructed and a multi-objective fitness function for automatic evaluation is defined. The feature extractor is used to extract the features of melodic music, such as the self-similarity of the notes in the melody, the graph of the melody, the linearity of the notes, the range used, and so on. The linear combination of these features is used as a multi-objective fitness function to calculate the fitness of all the generated melody individuals and determine whether the individual needs to evolve according to the calculated fitness function. Then, the crossover and mutation operations in the evolution process are studied to make the melodic evolution proceed smoothly. In order to ensure that the sum of the time values of the notes in each section before and after the crossover is kept unchanged, the time value correction problem of the notes is considered in the crossover operation. When performing mutation operation, ensure that the interval of adjacent notes is not more than 8 degrees for the variation of tone level. Finally, a multi-objective fitness function based on genetic algorithm melody creation experimental system platform is built. The experimental results show that the system can shorten the time of evaluation and greatly improve the efficiency of melody creation on the premise of creating melody with high quality.
【学位授予单位】：河南师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP18;J614.6

【参考文献】