(A) study on deep network-based multi-objective optimization and automatic evolutionary music composition

다국어 초록 (Multilingual Abstract)

In recent decades, artificial intelligence has achieved remarkable growth with rising of deep learning. It has been applied in various fields of engineering, and significantly contributed to industrial development. Nowadays, many artificial intelligence researchers have been a lot of attention to implementing creative ability of human such as music composition. Evolutionary Computing has shown excellent performance in producing music since 1990. It has been established as a research field called “Evolutionary music composition”, and it plays a central role in automatic music composition.
Music composition process can be divided into various sub-tasks, such as creating melody, harmony, and rhythm composition. Although all the composition elements are important, melody composition is the most creative part, and it is the most difficult part to automate. In the field of evolutionary music composition, research has been carried out to program the computer to automatically generate melodies. However, existing evolutionary approaches could have generated only a few melodies in a certain style that is specified by pre-defined fitness functions. They are not able to deal with the multi-dimensional nature of music. There are many perspectives in music assessment. Some measurements are conflicting with each other because of a human listener’s preference. An existing method does not consider the trade-off in composing music. Thus, the composition results could not satisfy the various musical preferences. This paper presents a novel multi-objective evolutionary approach to automatic melody composition to produce a variety of melodies at once. Two conflicting fitness measures are investigated to evaluate the melody’s fitness; stability and tension. The fitness functions are elaborately designed based on music harmony theory. The proposed method can provide all well-composed set of melodies between the two conflicting perspectives to the user at a single run.
Moreover, the music composition process is a high-dimensional problem; the number of combinations can be infinitely high, depending on composition length and the number of available tones. As the size of the problem grows, the optimization process of conventional evolutionary algorithms can become very inefficient. Therefore, we propose a new Multi-Objective Generative Deep network-based Estimation of Distribution Algorithm (MODEDA) to solve the problem with a high-dimensional solution by modeling and utilization of linkage information of music.

목차 (Table of Contents)

• 1. Introduction 1
• 2. Background 5
• 2.1 Multi-objective Optimization 5
• 2.2 Evolutionary Music Composition 7
• 2.3 Estimation of Distribution Algorithm 9
• 2.4 Variational Autoencoder 10
• 2.5 Music Theory 12
• 3. Generative Deep Networks-based Estimation of Distribution Algorithm for Solving Multi-objective Optimization Problems 16
• 3.1 Algorithm Framework 17
• 3.2 Modeling and Sampling Using VAE 19
• 4. Multi-objective Approach to Evolutionary Melody Composition 22
• 4.1 Representation 23
• 4.2 Multi-objective fitness functions 24
• 4.2.1 Fitness function on stability 25
• 4.2.2 Fitness function on tension 27
• 4.2.3 Fitness penalty function 29
• 5. Experimental Results 36
• 5.1 Multi-objective Knapsack Problem Experiment Results 36
• 5.2 Discussion 47
• 5.3 Experiment and Discussion: Melody Composition 49
• 6. Conclusion 56
• References 59
• Appendix 64
• Korean Abstract 77