University of Ghana http://ugspace.ug.edu.gh
UNIVERSITY OF GHANA 
COLLEGE OF BASIC AND APPLIED SCIENCE 
 
 
 COMPOSITION AND EVALUATION OF MUSIC USING ARTIFICIAL 
INTELLIGENCE: A GENETIC ALGORITHM APPROACH 
 
BY 
ABIGAIL WIAFE 
(10703638)  
 
A THESIS SUBMITTED TO THE SCHOOL OF GRADUATE STUDIES IN PARTIAL 
FULFILMENT OF THE AWARD OF DEGREE OF MASTER OF PHILOSOPHY IN 
COMPUTER SCIENCE 
 
DEPARTMENT OF COMPUTER SCIENCE 
 
October 2020 
University of Ghana http://ugspace.ug.edu.gh
Declaration 
I hereby declare that I am the sole author of this thesis and all materials used from other 
sources or in collaboration with other researchers have been properly and fully 
acknowledged. 
 
Abigail Wiafe (student)    Dr Ebenezer Owusu (Supervisor) 
___________Date: October 8, 2020   ___ ____Date: October 8, 2020 
 
Dr Ferdinand Kastriku (Co-supervisor) 
_____________Date: October 8, 2020 
 
 
 
  
ii 
University of Ghana http://ugspace.ug.edu.gh
 
  
iii 
University of Ghana http://ugspace.ug.edu.gh
Abstract 
Advancement in technology has contributed enormously to music composition, distribution 
and consumption. Algorithmic music is the use of technologies and techniques to aid music 
composition. Currently, there are various algorithms including rule-based, memetic, deep 
convolutional generative adversarial networks (GAN), particle swarm optimization (PSO) 
and stochastic composition algorithm for composing music. However, music obtained from 
these algorithms may not be appealing although the algorithm used may produce the optimal 
music based on the criteria defined by the composer. Furthermore, the optimal music 
produced shows that the aesthetic quality of the music is limited. This study, therefore, seeks 
to compose monophonic music using genetic algorithm and assess how it is appealing to 
humans. Nine (9) melodies were composed. The compositions were three versions each for a 
GA terminated at ten thousand, twenty thousand and sixty thousand generations. The changes 
in total fitness scores over the various generations were plotted and analysed. It was concluded 
that the algorithm performed best when terminated at ten thousand generations by producing 
the optimal music around 750th generation. However, a survey conducted to evaluate the 
aesthetic performance of the algorithm through human evaluation demonstrated that the music 
obtained from the termination at 60000th generation was the most pleasant composition. This 
therefore suggests that although a GA algorithm may perform better in terms of composition 
time, it does not guarantee an aesthetic music. 
  
iv 
University of Ghana http://ugspace.ug.edu.gh
   Dedication 
This work is dedicated to my husband, Dr. Isaac Wiafe for his wonderful encouragement, 
endless support and contribution to my education and also to my children Kofi Osei Wiafe, 
Kwabena Asante Wiafe and Kweku Obe Wiafe. 
 
 
 
 
  
v 
University of Ghana http://ugspace.ug.edu.gh
Acknowledgement 
I am thankful to the Almighty God for his mercies, divine favour, guidance and protection.  
I wish to express my heartfelt gratitude to my supervisors, Dr. Ebenezer Owusu and Dr. 
Ferdinand Kastriku for their direction and supervision throughout my research project. I also 
appreciate their intuitive reviews and evaluations that shaped and enhanced the quality of this 
study. A profound thanks to my husband, Dr. Isaac Wiafe and Dr. Jamal-Deen Abdulai for 
their immense support. God bless you all.  
A special appreciation goes to my children, Kofi Osei Wiafe, Kwabena Asante Wiafe, Kweku 
Obe Wiafe and my mother Rosina Serwaa. Last but not the least, I would like to thank Ms 
Gifty Budu, Ms Adelaide Oduro- Asante, Rev Michael Owusu Tabiri, my coursemates Father 
William Abaiku Apprey, Samuel Abedu, Melody Nunya Kakrabah, Fredrick Boafo, Akon 
Obu Ekpezu, Jacqueline Asor Kumi and Mr Charles Nutrokpor for his technical assistance. 
  
  
vi 
University of Ghana http://ugspace.ug.edu.gh
Table of Contents 
Abstract ............................................................................................................................... iv 
Dedication ............................................................................................................................ v 
Acknowledgement .............................................................................................................. vi 
Table of Contents ............................................................................................................... vii 
List of Figures .................................................................................................................... xii 
List of Tables .................................................................................................................... xiv 
Related Publications .......................................................................................................... xv 
Chapter 1 INTRODUCTION .................................................................................................. 1 
1.1 Research Background .................................................................................................... 1 
1.2 Research Problem and Relevance .................................................................................. 3 
1.3 Aims and Objectives ...................................................................................................... 3 
1.4 Expected Contributions ................................................................................................. 4 
1.4.1 Theoretical Contribution ......................................................................................... 4 
1.4.2 Practical Contribution ............................................................................................. 5 
1.5 Outline of Thesis ............................................................................................................ 5 
Chapter 2 BACKGROUND LITERATURE .......................................................................... 8 
2.1 Overview ........................................................................................................................ 8 
2.2 Basics of Music .............................................................................................................. 8 
2.3 Music Composition Methods ....................................................................................... 10 
2.3.1 Handwriting, Listening and Instruments ............................................................... 10 
2.3.2 Software and Computer-Generated Music ............................................................ 11 
2.4 Generating Music with Algorithms (Algorithmic Music) ........................................... 12 
vii 
University of Ghana http://ugspace.ug.edu.gh
2.5 Algorithms for Generating Music ................................................................................ 12 
2.5.1 Mathematical Models ............................................................................................ 13 
2.5.2 Grammars .............................................................................................................. 13 
2.5.3 Knowledge-Based Systems and Learning Systems .............................................. 14 
2.5.4 Evolutionary Methods ........................................................................................... 15 
2.5.5 Hybrid System ....................................................................................................... 15 
2.6 Application of Algorithmic Music ............................................................................... 16 
2.6.1 Video Games ......................................................................................................... 16 
2.6.2 Movies ................................................................................................................... 16 
2.6.3 Therapeutic Properties .......................................................................................... 17 
2.6.4 Affective Algorithmic Music Composition .......................................................... 17 
2.7 Chapter Summary ........................................................................................................ 19 
Chapter 3 EXISTING STUDIES IN ALGORITHMIC COMPOSITIONS .......................... 20 
3.1 Overview ...................................................................................................................... 20 
3.2 Existing Literature ....................................................................................................... 20 
3.3 Review Approach ........................................................................................................ 23 
3.3.1 Planning, Conducting, Study Selection and Documentation ................................ 23 
3.4 Review Questions ........................................................................................................ 24 
3.4.1 Selection Strategy ................................................................................................. 27 
3.4.2 Data Extraction ..................................................................................................... 28 
3.5 Review Findings and Results ....................................................................................... 28 
3.5.1 Publication Trends ................................................................................................ 29 
AI Methods for Algorithmic Composition ..................................................................... 31 
viii 
University of Ghana http://ugspace.ug.edu.gh
3.5.2 Evaluating Techniques in Algorithmic Composition ............................................ 33 
3.5.3 Music File Used for Composition ......................................................................... 36 
3.6 Musical Instrument Digital Interface (MIDI) .............................................................. 37 
3.6.1 MIDI File Format .................................................................................................. 39 
3.6.2 Aftertouch, Pitch bend and Channels .................................................................... 40 
3.7 Genetic Algorithms ...................................................................................................... 41 
3.7.1 Selection, Cross Over, Mutation and Termination Process .................................. 42 
3.7.2 Areas of Application ............................................................................................. 43 
3.7.3 Advantages of Genetic Algorithms ....................................................................... 44 
3.7.4 Limitations of Genetic Algorithms ....................................................................... 44 
3.8 Chapter Summary ........................................................................................................ 45 
Chapter 4 METHODOLOGY ............................................................................................... 46 
4.1 Overview ...................................................................................................................... 46 
4.2 Paradigms and Beliefs in Algorithmic Music Composition ........................................ 46 
4.3 The Design Science Method ........................................................................................ 47 
4.4 Limitation of design science method ........................................................................... 48 
4.5 Research Approach for the Study ................................................................................ 49 
4.5.1 Awareness of the problem ..................................................................................... 49 
4.5.2 Suggestions ........................................................................................................... 50 
4.5.3 Development ......................................................................................................... 51 
4.5.4 Evaluation ............................................................................................................. 51 
4.5.5 Conclusion ............................................................................................................ 51 
4.6 Chapter Summary ........................................................................................................ 52 
ix 
University of Ghana http://ugspace.ug.edu.gh
Chapter 5 COMPOSITION MODELLING AND FITNESS ANALYSIS ........................... 53 
5.1 Overview ...................................................................................................................... 53 
5.2 Genetic Algorithm Implementation ............................................................................. 53 
5.2.1 Composition Modelling ........................................................................................ 53 
5.3 Resultant Melodies ...................................................................................................... 56 
5.3.1 Compositions at 10K Generations ........................................................................ 57 
5.3.2 Compositions after 20K generations ..................................................................... 60 
5.3.3 Compositions after 60K generations ..................................................................... 64 
5.4 Chapter Summary ........................................................................................................ 67 
Chapter 6 EVALUATION AND DISCUSSIONS ................................................................ 68 
6.1 Overview ...................................................................................................................... 68 
6.2 Human Evaluation ....................................................................................................... 68 
6.2.1 Data Gathering ...................................................................................................... 69 
6.2.2 Evaluation Results ................................................................................................. 71 
6.3 Implications of Study ................................................................................................... 72 
6.3.1 Fitness Functions and Human Evaluation ............................................................. 72 
6.4 Variations in Convergence Timing .............................................................................. 74 
6.5 Chapter Summery ........................................................................................................ 74 
Chapter 7 STUDY LIMITATION AND CONCLUSION .................................................... 76 
7.1 Overview ...................................................................................................................... 76 
7.2 Evaluation of Research Objectives .............................................................................. 76 
7.3 Limitations and Future Work ....................................................................................... 77 
7.3.1 Monophonic Composition ..................................................................................... 77 
x 
University of Ghana http://ugspace.ug.edu.gh
7.3.2 MIDI Files ............................................................................................................. 78 
7.3.3 Sequence of Listening by Subjects ....................................................................... 79 
7.3.4 Expert Evaluation .................................................................................................. 79 
References ............................................................................................................................. 80 
 
  
xi 
University of Ghana http://ugspace.ug.edu.gh
List of Figures 
Figure 1-1  Thesis outline ........................................................................................................ 6 
Figure 3-1 Steps in the Review Process ................................................................................ 23 
Figure 3-2 Number of Studies Identified for the Review ...................................................... 26 
Figure 3-3 Distribution of primary studies by year ............................................................... 29 
Figure 3-4 Distribution of publication venue ........................................................................ 30 
Figure 3-5 Geographical distribution of primary studies ...................................................... 31 
Figure 3-6 Types of file Used for Composition .................................................................... 37 
Figure 3-7 Connected keyboard, PC and one generator for MIDI ........................................ 38 
Figure 3-8 A Byte structure of a MIDI file ........................................................................... 39 
Figure 3-9 Cross-over and mutation ...................................................................................... 42 
Figure 3-10 Sample Pseudo code for GA .............................................................................. 43 
Figure 4-1 Relevance of the study ......................................................................................... 50 
Figure 5-1 GA music composition process used for the study .............................................. 56 
Figure 5-2 Resultant composition after 10 K generations (v1) ............................................. 57 
Figure 5-3 Total fitness scores for 10k generations (v1) ....................................................... 58 
Figure 5-4 Resultant composition after 10K generations (v2) .............................................. 58 
Figure 5-5 Total fitness scores for 10K generations (v2) ...................................................... 59 
Figure 5-6 Resultant composition after 10K generations (v3) .............................................. 59 
Figure 5-7 Total fitness Scores for 10K generations (v3) ..................................................... 60 
Figure 5-8 Resultant composition after 20K generations (v1) .............................................. 61 
Figure 5-9 Total fitness Scores for 20K generations (v1) ..................................................... 61 
Figure 5-10 Resultant composition after 20K generations (v2) ............................................ 62 
xii 
University of Ghana http://ugspace.ug.edu.gh
Figure 5-11 Total fitness Scores for 20K generations (v2) ................................................... 62 
Figure 5-12 Resultant composition after 20K generations (v3) ............................................ 63 
Figure 5-13 Total fitness Scores for 20K generations (v3) ................................................... 63 
Figure 5-14 Resultant composition after 60K generations (v1) ............................................ 64 
Figure 5-15 Total fitness Scores for 60K generations (v1) ................................................... 64 
Figure 5-16 Resultant composition after 60K generations (v2) ............................................ 65 
Figure 5-17 Total fitness Scores for 60K generations (v2) ................................................... 65 
Figure 5-18 Resultant composition after 60K generations (v3) ............................................ 66 
Figure 5-19 Total fitness Scores for 60K generations (v3) ................................................... 66 
 
  
xiii 
University of Ghana http://ugspace.ug.edu.gh
List of Tables 
Table 2-1 List of music terms and their meanings .................................................................. 9 
Table 3-1 Goal-Question-Metric adopted from (Basili et al., 1994) ..................................... 24 
Table 3-2 Review Questions and their Objective .................................................................. 25 
Table 3-3 Databases, Websites and Keywords ..................................................................... 27 
Table 3-4 AI methods for composing music ......................................................................... 32 
Table 3-5 Methods for Evaluating Algorithmic music ......................................................... 34 
Table 6-1 Description of Participants (n=12) ........................................................................ 70 
Table 6-2 Responses of subjects (n=12) ................................................................................ 72 
Table 6-3 Model Performance vs Human Evaluation (n=12) ............................................... 73 
 
  
xiv 
University of Ghana http://ugspace.ug.edu.gh
Related Publications 
i. Wiafe A., Nutrokpor C., Owusu E., Kastriku F.A., Wiafe I., (submitted). Design and 
Evaluation of Algorithmic Music Using Genetic Algorithms in the 2020 Science and 
Development Platform, Accra, Ghana 
ii. Wiafe A., Owusu E., Armah Z. A., Wiafe I., (accepted) A Systematic Review of 
Artificial Intelligence Methods in Automatic Music Composition in 3rd International 
Telecommunications Society Conference, Cape Town, SA 
iii. Wiafe A and Fränti P. (2020) Security and Ethical Concerns of Affective Algorithmic 
Music Composition in Smart Spaces in Modern Theories and Practices for Cyber 
Ethics and Security Compliance. eds. Yaokumah, W., Rajarajan, M., Abdulai, J., 
Wiafe, I., & Katsriku, F. A. Hershey, PA: IGI Global. doi: 10.4018/978-1-7998-3149-
5
xv 
University of Ghana http://ugspace.ug.edu.gh
Chapter 1 
INTRODUCTION 
1.1 Research Background 
The success of machine learning (ML) as applied to music composition enables researchers, 
artists and novice musicians to use computers and software tools to compose music. 
Development of computer or algorithmic music is related to the different technologies and 
techniques that aid in music composition. Many individuals and artists compose music using 
their knowledge, however, some of these pieces of music lack aesthetic and creativity 
because a formal approach used. Music composition requires in-depth knowledge of 
numerous processes and thus the introduction of machine learning and artificial intelligence 
for composition presents a number of benefits including reducing financial cost, time and 
stress (Lopez-Rincon et al., 2018). 
Artificial Intelligence (AI) aids in the composition process by analysing data from various 
musical pieces to produce new music. This approach is referred to as automatic music 
composition. In most cases minimal human intervention is required in generating the music 
(Alpern, 1995). The adaptation of formal instructions and processes to create music dates 
back to the ancient Greek era where Pythagoras (the Greek philosopher) laid the foundation 
or the domain of study (Simoni, 2003). He discovered that numbers and mathematical 
equations can be used for music composition (Fauvel et al., 2006;  Papadopoulos, 2002; 
Pitkänen, 2014) since the study of music is based on creating sound patterns which can be 
applied in mathematics. This has been emphasised in recent times (Beynon 2012). 
1 
University of Ghana http://ugspace.ug.edu.gh
Researchers and practitioners of algorithmic music composition adopt different approaches 
of composing music (Roads & Strawn, 1996). Mathematical modelling is perhaps one of the 
earliest formulations of automatic music composition. Methods including stochastic 
processes (Jones, 1981) and Markov Chains (Ramanto et al., 2017) have been demonstrated 
to be effective in algorithmic music composition. Other researchers and practitioners have 
adopted different approaches including grammars. Grammars have been identified to be 
effective in the analysis and composition of algorithmic music. In particular, relevant studies 
identify six types of grammars used for music composition (Roads & Wieneke, 1979). Also, 
methods such as knowledge-based systems (Delgado et al., 2009), evolutionary methods 
(Matić, 2010), learning systems (Rodrigues et al., 2016) and hybrids of these methods (Fox 
& Crawford, 2016) have been used to compose music. 
Amidst the development and advancement in algorithmic music composition, the ability to 
compose music that is pleasing to listen to remains. This study, therefore, seeks to use a 
machine learning algorithm to compose monophonic music and evaluate its aesthetics. 
Specifically, genetic algorithms are used to compose different music files at different 
generations and the aesthetic quality of the resulting music at the various generations is 
assessed by human listeners to evaluate it. Aesthetic quality describes the expression of 
personal “taste” in music (Aldrich, 1963; Scruton, 1982) relating to how beautiful, great, 
dynamic, noisy or terrible a music is composed. 
 
 
 
2 
University of Ghana http://ugspace.ug.edu.gh
1.2 Research Problem and Relevance 
The adoption of machine learning techniques for music composition seeks to ease 
composition process. Currently, there are various techniques for composing music. 
However, overtly some experience challenges because well-defined fundamental theories 
for compositions are not integrated in the field of music. This therefore suggests that music 
obtained from these algorithms may not be appealing although the algorithm used may 
produce the optimal music based on the criteria defined by the composer. Yet, studies that 
seek to assess the relationship between the optimal music produced versus the aesthetic 
quality of the music is lacking. 
1.3 Aims and Objectives 
This research therefore seeks to compose a simple music and assess how it appeals to humans 
through evaluation. The motivation to produce a simple composition is based on the fact 
that, most existing ML compositions are complex and thus make it difficult for humans to 
assess its basic aesthetics quality. The study adopts a genetic algorithm (GA) to compose the 
music. GAs have been largely adopted by a number of researchers for composing music 
(Anderling et al., 2014) and it has also been demonstrated to produce “good” music of high 
aesthetic quality. 
To achieve this aim, the following objectives are identified: 
i. To review existing machine learning methods for music composition. 
ii. To compose a simple music using a genetic algorithm. 
iii. To evaluate the performance of the algorithm using fitness scores. 
iv. To evaluate the aesthetic performance of the algorithm through human evaluation. 
3 
University of Ghana http://ugspace.ug.edu.gh
Although the above-listed objectives seek to address the main aim of the study, it is also 
expected to contribute to the knowledge gap in ML algorithms for music composition as 
well as music composition in general. It is emphasised that the aim of this study is not to 
produce an appealing music, but to compose a monophonic music using a simple population 
of musical tones, and also to evaluate the relationship between the optimal music and human 
quality assessment. The next section discusses the expected research contributions to the 
study. 
1.4 Expected Contributions 
The result of this research is expected to contribute to machine learning and artificial 
intelligent research with specific emphasis on music composition algorithms. Particularly, it 
is hoped to contribute to theory and practice regarding the aesthetics quality of music 
composed using machine learning methods (genetic algorithms). 
1.4.1 Theoretical Contribution 
This research shall provide theoretical insights on the relationship between aesthetic quality 
of music generated using genetic algorithms. It seeks to explore and explain the music 
quality of an optimal music generated using a genetic algorithm that is pleasing to the human 
listener. As argued earlier, research that seeks to assess the aesthetic quality of simple ML 
generated music is limited. Yet, this information provides relevant insight that may guide 
the development of complex ML music models that will meet the aesthetic quality of 
listeners. It is therefore expected that the findings of this study shall extend existing literature 
on algorithmic music composition. 
4 
University of Ghana http://ugspace.ug.edu.gh
1.4.2 Practical Contribution 
In terms of practice, the findings are expected to facilitate the design of music composition 
systems. In effect, it will provide both novice and experienced composers basic but relevant 
information that will guide the use of ML methods for music composition. In addition, it is 
expected that, this study will provide developers of automatic music composer information 
regarding GA fitness functions that will guide the composition of better-quality music that 
are pleasing to the human ear. 
1.5 Outline of Thesis 
This thesis is in seven chapters. Chapter One presents the introduction and background of 
the research by providing a brief highlight of the problem. The chapter also states the 
research aim and objectives and explains briefly the expected contribution of the research. 
Highlights of the structure of the thesis are also presented in this chapter. 
Chapter Two discusses the related literature. It explains the basic concepts of music and 
discusses the various categories in the field. The chapter further discusses the various 
methods used in the composition of music and the advancement of algorithmic compositions 
in the composition of affective music. 
Chapter Three presents a summary of algorithmic music composition. It explains a 
systematic review conducted and reported findings of existing studies in algorithmic music 
composition. It discusses existing reports that perform literature reviews by conducting a 
systematic review of the literature. Furthermore, to ensure that the review performed is 
auditable and verifiable, the approach and search strategy used for conducting the systematic 
5 
University of Ghana http://ugspace.ug.edu.gh
mapping are explained. The findings and results from the survey are presented to guide the 
direction of the study. 
Chapter Four provides a guide on how the research was conducted. It presents design 
science research paradigm as appropriate for this study and thus demonstrated the systematic 
and scientific approach used in the study. 
 
Figure 1-1  Thesis outline 
Chapter Five presents a discussion on the development and evaluation of algorithmic music. 
It discusses how the genetic algorithm was implemented, the experimental setup, and the 
resultant composition after 10,000 generation, 20,000 generation and 60,000 generation. The 
chapter also presents an evaluation of the composed music. It further presents results of a 
survey conducted as part of the validation process of the composed music. 
Chapter Six presents implementation of the study including the theoretical implication, 
practical implication, and related security implications of ML music models. 
6 
University of Ghana http://ugspace.ug.edu.gh
Chapter Seven concludes the study. It provides the limitations associated with the study 
and future research directions. It also evaluates the research objectives to assess how well 
the study addressed its objectives. 
 
Figure 1-1 shows how the various chapters are related to guide the reader. The chapters are 
categorised into four main types (i.e. motivation, approach, outcome and assessment). 
Chapters 1, 2 and 3 present discussions on the motivation for the study while chapter 4 is the 
research approach. Chapters 5 and 6, present the outcome of the study with parts of chapter 
6 and all of chapter 7 discussing the assessment of outcome. 
  
7 
University of Ghana http://ugspace.ug.edu.gh
Chapter 2 
BACKGROUND LITERATURE 
2.1 Overview 
According to Merriam & Merriam (1964), human history has witnessed the creation and 
pleasure of music by various cultures across the globe. Music entails unique sounds that 
humans attribute to creativity and aesthetic in the composition process. It is the organisation 
of sound with the potential to convey a feeling either in a power-driven or natural sense 
(Hermitary & Meng-hu 2010). “Pure” music is made of instrumentals without any 
accompaniment (Kania 2017). 
This chapter seeks to partially address the second objective of this study. It provides the 
basic concepts of music. It starts with a presentation on the basic terminologies of music, 
followed by existing methods that are used for composing music. Methods such as 
grammars, learning systems, evolutionary systems and the handwriting approach to music 
composition are discussed. The chapter demonstrates how advancement in algorithmic 
compositions advances the composition of affective music. 
2.2 Basics of Music 
Music composition involves creativity in order to generate new and unique melodies. 
Compositions can be done by writing the music with the hand (Charles, 1958), using an 
instrument (Chew, 2011; Magnusson & Mendieta, 2007), software (Lumsden & Ludlow, 
2007; Magnusson & Mendieta, 2007) or by training the ear to listen and compose (Woody 
8 
University of Ghana http://ugspace.ug.edu.gh
& Lehmann, 2010).The use of various terminologies facilitates the understanding, writing 
and reading of music. Table 2-1 is a list of some terminologies in music and their meanings. 
Table 2-1 List of music terms and their meanings 
Term Meaning 
Note A small bit of sound or is a symbol denoting a musical sound. 
Pitch The highness or lowness of a note 
Semitone The smallest practically usable space between two tones 
Octave The interval between one musical pitch and another with half or double 
its frequency 
Rhythm The timing of how long or short a musical note will last 
Volume The loudness or intensity of a tone 
Melody Represented by pitches arranged in a horizontal sequence, one 
 sounding after another 
Tonality The notes in which specific hierarchical pitch relationships are based 
on a key "centre" or tonic 
Interval The distance between the two notes, either when they sound 
simultaneously or one after the other 
Tempo The speed of a given musical piece. It is measured in beat per minute  
Time signature   Notation used to measure how music is counted  
Key signature A set of symbols sharps [#] and flats [b] natural (♮) on a musical staff 
to indicate which note will be played. 
Chord	 A harmonic set of pitches that consist of notes  
9 
University of Ghana http://ugspace.ug.edu.gh
Term Meaning 
Chord	 How the series of chords are played in sequence, also known as 
Progression	 harmonic progression 
Diatonic	scale	 A scale that with five whole steps or tone and two half steps in each 
octave 
Voice	leading	 The linear progression of individual melodic lines and the interaction 
between that creates harmonies  
Smoothness	 The connection between notes in music 
Step	 The difference in pitch between 2 consecutive notes in a musical scale 
Rest	 A musical notation sign that indicates the absence of a sound 
2.3 Music Composition Methods 
There are different methods for generating music and below is a brief description of the 
various methods. 
2.3.1 Handwriting, Listening and Instruments 
The concept of paper and pencil was considered as the best tools for the music composition 
in the 1940s. This is because it was considered to be handy (Day & Jarrett, 2008). Composers 
with knowledge in the solfege (do, re, mi, fa, so, la, ti) used them to represent a composition. 
The sequential flow of the notes was known and thus made it easier for others to play or 
translate to music notation. 
The effect of music listening is to aid people in composing music. Hedden (1981) researches 
how music educators develop one’s ability to hear and compose music by using various 
10 
University of Ghana http://ugspace.ug.edu.gh
strategies. A listener, listening without any distortion, and a repeat hearing have positive 
influence on the individual such that the individual’s attention is absorbed in the music. 
Furthermore, with music classified as a universal language (Campbell, 1997) when good 
listening skills are developed with critical thinking phrases, it can be understood and used in 
the composition of meaningful music. 
Piano, keyboard, flute, harp, saxophone, xylophone, violin and double bass are some 
examples of instruments that are used in the composition of music. The piano is a polyphonic 
instrument that can play two melodies at the same time, unlike flute. Both the keyboard and 
piano are easy to learn in the composition of music. 
2.3.2 Software and Computer-Generated Music 
Nowadays, the use of computers and software programs play a vital role in music 
composition. Novice, researchers and artist are able to compose music that fit their ideas, 
reduce their financial cost and provide them with a stress-free environment for the 
composition process. Software such as MuseScore, Cubase, Anvil studio, Finale, 
Famitracker, Anvil studio, Pro Tools are all examples of composing tools used for creating 
music. 
Additionally, some instruments work together with software packages to aid music 
generation. Music sequencer software and hardware are connected to computers to facilitate 
composition, recording, editing or playing back of the music. MIDI (discussed later in this 
chapter) or other plug-ins are used to control notes and generate music. It is however limited 
because sequencers have less music notation capabilities and thus can play only simple 
melodies when compared to other plug-ins. Plug-ins such as Capella allows keyboard or 
11 
University of Ghana http://ugspace.ug.edu.gh
MIDI instruments to be plugged into its hardware to facilitate music creation by interpreting 
what is played and produces a written music notation. 
2.4 Generating Music with Algorithms (Algorithmic Music) 
There are two approaches used in generating algorithmic music; music composed by the 
computer (when the algorithm can make choices of its own during the creation process) and 
music composed with the aid of computer. Although algorithmic composition was 
recognised in the mid 1950s, many composers use different techniques in composing music 
(McAlpine et al., 1999). 
AI technologies are useful in music composition since intelligent agents can “think” and 
solve problems like humans. Algorithmic music composition follows a set of rules to 
compose music. It is the process of using a structured methods to create music with minimal 
human intervention (Alpern 1995). In some cases it uses a trial and error  approach (Simoni, 
2003) and this may lead to some “useless” melody. Hence, some researchers have advocated 
for the use of humans to listen and revise melodies to help determine its quality (Jacob, 1995, 
1996). Human may provide a better result as compared to the automatic composition process 
which adheres to rules and constraints in the algorithms and by its propositions is not able 
to determine the quality of music. However, although this assertion is intuitive, it has not 
been empirically demonstrated. 
2.5 Algorithms for Generating Music 
Currently, researchers and practitioners of algorithmic music composition adopt different AI 
approaches to facilitate music generation. These AI approaches are classified based on their 
12 
University of Ghana http://ugspace.ug.edu.gh
development and features. Papadopoulos & Wiggins (1999) grouped these algorithms into 
six namely, mathematical models, grammars, knowledge-based systems, evolutionary 
methods, systems which learn and hybrid systems. Below is a brief discussion on these 
models. 
2.5.1 Mathematical Models 
Developing mathematical models and implementing it effectively and efficiently enhance 
algorithmic composed music. Stochastic process for music composition (Jones, 1981) and 
Markov Chains (Ala & Wadi, 2012; Ames, 1989; Cambouropoulos, 1994; Ramanto et al., 
2017; Volchenkov et al., 2012) are similar. Markov chains are used for composition by 
generating different states of sequences and each sequence state is evaluated using 
probability theories. Random note selection are used to compose high-quality music (Moorer 
1972). Markov chains are mostly used for melody generation but exhibit some challenges. 
One of such challenges identified is how it assesses probabilities of the state diagrams that 
are generated randomly. This is because they have to be derived from the existing composed 
music. Thus, they most seek to generate outputs similar to the inputs. They also produce 
unconnected pitch values (Järveläinen, 2000). 
2.5.2 Grammars 
Grammars are methods which display special features in the analysis and composition 
process. They exert some degree of hierarchical structure in music composition with rules 
that connect notes to form phrases, segments and pieces of music. Linguistic models which 
lead to the discovery of grammars that can be used for music composition was developed by 
Chomsky (1957). Roads & Wieneke (1979) argued that there are six types of grammars. This 
13 
University of Ghana http://ugspace.ug.edu.gh
includes the four basic types introduced earlier by Chomsky (1957). The first type is free 
grammar. Free grammar has no constraints on the rules used in the composition process as 
compared to context-sensitive (Swyter, 2009). 
Context-free grammar aids in the generation of multi-level hierarchical trees and finite -state 
requires at least one non terminal token on either side to enable rewriting of rules since they 
are not able to compose music generated by the context-free grammars. Transformation 
grammar covers three sections with the first creating the abstract sentences, the next converts 
it into sentences and map the sentence to a phoneme. The sixth type is the regulated grammar 
which uses control languages to decide when to combine production rules with a context-
free grammar. 
2.5.3 Knowledge-Based Systems and Learning Systems 
The knowledge-based system is a form of Artificial Intelligence (AI) which uses human 
expert knowledge to aid in decision making. Algorithmic music incorporates knowledge-
based systems in composition by modelling computers to compose music. The process is 
facilitated by using the knowledge-based to select notes based on precise reasoning. 
Also, some applications are designed to learn and compose music. These systems use current 
information available but not previous knowledge for composition. They are grouped into 
artificial neural network (ANN) and ML based on how they perform their task and store 
information. ANN uses a set of input patterns to train the network in order to generate similar 
patterns thus compositions are achieved by representing music notation in the form of 
temporal patterns as inputs trained to generate outputs in a network. They are also capable 
of using classification algorithms to classify features for music composition. 
14 
University of Ghana http://ugspace.ug.edu.gh
2.5.4 Evolutionary Methods 
An evolutionary algorithm is an algorithm that follows the laws of evolution. It is an AI 
technique and can be used for music composition. Evolutionary algorithms (EAs) are driven 
by optimization process such as the Darwinian evolution. Genetic algorithm (discussed in 
the next chapter) is an example of an evolutionary algorithm. Genetic algorithm, Evolution 
Strategies (Rechenberg, 1989) and Evolutionary Programming(Jones, 1998) are three basic 
application of evolutionary algorithm. Although they demonstrate similarities such as the 
use of fixed-length strings, mutation and selection operators they behave differently (Fraser, 
1957; Goldberg & Holland, 1988; Holland, 1975; Sastry et al., 2005). 
2.5.5 Hybrid System 
Hybrid systems consist of a mishmash of AI techniques used in the composition process. 
Gutknecht (1992) suggested that the “postmodern mind” identifies various AI potencies 
combines and adopts them. HARMONET a hybrid system (Hild et al., 1992) combines back 
propagation methods and symbolic algorithms in a hierarchical system to compose music. 
In some cases a combination of stochastic processes and genetic algorithms are used (Fox & 
Crawford, 2016). Although, they have been demonstrated to be effective in composing  
music (Chuan & Chew, 2007). 
Some researchers (Papadopoulos & Wiggins, 1999) have argued that hybrid systems 
demonstrate complications when used with a fully integrated mode. This is because the 
implementation of the algorithm is time-consuming when checking for verification and 
validation. 
15 
University of Ghana http://ugspace.ug.edu.gh
2.6 Application of Algorithmic Music 
The ability to generate music using algorithms and modern advancement in parallel 
computing have provided both researchers and practitioners the opportunity to advance the 
use of music in various domains. Below is a discussion of three main areas of application. 
2.6.1 Video Games 
Background music in video games plays an important role such that it creates a lifelike 
environment with natural sound and an add-on for better entertainment. Creation of music 
in video games enhances user’s interaction experience. Algorithmic music can be used to 
generate music to influence the events of the game, the characters, as well as provide a clue 
about the position or distance of an enemy, or an event about to happen (Prechtl, 2016; 
Prechtl et al., 2014). AMEETM by Hoeberechts et al., (2007) is a music composition engine 
used for generating adaptive music for video games. Other video games that uses automatic 
composed music includes Avatar (Rice et al., 2013), the Thief (Kennerly et al., 2009) and 
Splinter cell (Thorsen, 2018). 
2.6.2 Movies  
Similar to video games, music generated for movies provide appropriate emotional support 
for the various states and functions. Livingstone et al., (2007) explained how the altering of 
music can produce emotional variation based on structural and performance rules. 
Algorithmic compositions are used in movies to help create and shape emotional response 
such that the movie captures the emotions of the audience. They are used as background 
music to provide audience feelings of being scared, anxious or hopeful. They also provide 
16 
University of Ghana http://ugspace.ug.edu.gh
rhythms to scenes and interpret the next activity about occur. Metascore is an example 
application designed to provide and improve movie music (Callum et al., 2008). 
2.6.3 Therapeutic Properties 
Music has gained recognition for its therapeutic purposes (Rafieyan & Ries, 2007). It is safe 
and does not involve the introduction of medical instrument. It is also less expensive and can 
be delivered to reduce postoperative pain and anxiety needs (Hole et al., 2015). Kemper & 
Danhauer (2005) explained how selected music can reduce stress, surgical, acute, procedural 
and chronic pain among patients and also provide comfort and relaxation. Algorithmically 
composed music therapy focuses on all aspects of humans: from the brain, through the body 
to the mind. It provides emotional healing for people. An example of computer-aided music 
therapy is Carnatic music used for curling ailment (Vasudha et al., 2011). 
2.6.4 Affective Algorithmic Music Composition 
Advancement in Algorithmic music composition has made it possible to compose music 
with improved affective response. Over the years, different algorithmic music composition 
models and applications have been developed based on artificial intelligent (AI) techniques 
such as neural networks, deep learning, stochastic and heuristic composition models (Lopez-
Rincon et al., 2018). Lopez-Rincon et al., (2018) further acknowledged that recent 
advancements have improved automation of algorithmic music composition. Therefore, use 
of computers for automatic composition presents an opportunity in which computer-aided 
composition and emotional assessment are combined to produce affective algorithmic 
composition (AAC). 
17 
University of Ghana http://ugspace.ug.edu.gh
AACs are emotionally intelligent automatic music generation systems that explore current 
emotions or mood of listeners to compose affective music (Kirke et al ., 2013, Williams et 
al., 2017). These compositions aim at altering a person’s mood to a predetermined one. 
Specifically, they seek to target an individual’s affective descriptor (emotional response) in 
order to alter his or her mood (Williams et al., 2015). 
Williams et al., (2015) explained that AAC presents several opportunities and benefits to 
almost all aspects of human wellbeing. For example, AAC is used to compose music that 
reduces the emotional distress of patients based on their mood (Kemper & Danhauer, 2005). 
Again, they have the ability to produce larger music samples within a short time: a feature 
that provides a larger sample of music that can be selected by users for reducing stress levels. 
Similar to arguments made by Boyes (2003), AAC is challenged in terms of security. These 
are threats from malicious outsiders; malicious insiders; non-malicious insiders; and nature. 
These threats may come in different forms including, impersonation and identity spoofing, 
eavesdropping, user profiled tampering, malicious codes and denial-of-service.  
The security impact of impersonation in this context may lead to two major issues: such as 
data theft and unauthorized composition or unauthentic (i.e. non-affective music) music for 
legitimate users. When AAC is enabled in smart space, the impersonator may have access to 
sensitive data about a user and use it for malicious purposes. Particularly, the impersonator 
can compose music. In such instances, the authenticity of the composed music becomes 
questionable. Considering the possible impact of impersonation, it is recommended that 
biometric authentications are used in these systems.  
18 
University of Ghana http://ugspace.ug.edu.gh
Eavesdropping is another common network attack that needs to be considered in AAC 
systems. It is an unethical act in which an individual illegitimately intercept a private 
conversation, communication, digital transmission or activities of others without their 
consent. It poses privacy issues in AAC because it may lead to the disclosure of listeners 
profiles which can later be used for impersonating. There is no ambiguity about the emotional 
contagion ability of music (Egermann & McAdams, 2013) and hence measure that seek to 
prevent eavesdropping cannot be overemphasised.   
Ethical considerations are vital in all systems that seek to alter cognition, emotions or mood. 
AAC systems have functionalities that include compositional processes and affective 
performance structures. They use generative or transformative algorithms to induced or 
perceived to induce affective states of users (Williams et al., 2015). They adopted emotional 
models (i.e. categorical or dimensional models) to profile its users. Since users interact with 
systems that have intentions, any unethical practice may result in an unwanted effect.  
2.7 Chapter Summary 
Understanding the concepts of music and the use of computers to compose music is relevant 
since they facilitate the composition process. Although there are various techniques, a 
comprehensive analysis of the techniques is important. Accordingly, discussions in this 
chapter have demonstrated that research in algorithmic music composition is advanced. 
However, literature that summarises existing studies in the domain is scare. Hence, in the 
next chapter, a summary and discussion of existing studies that seek to compose music using 
algorithms is presented. 
  
19 
University of Ghana http://ugspace.ug.edu.gh
Chapter 3 
EXISTING STUDIES IN ALGORITHMIC COMPOSITIONS 
3.1 Overview 
As mentioned in the previous chapter, advancement in algorithmic composition of music has 
made it possible for it to be used in various domains and there is no ambiguity that its 
application will continue to be diverse. Yet, studies that seek to provide summaries in the 
domain is lacking. To ensure a comprehensive review, this chapter addresses the first 
objective of the research (partially) by reporting findings of a systematic review of existing 
studies in algorithmic music composition. It deviates from existing reports that perform 
classic literature reviews by conducting a systematic review of literature. To ensure that the 
review is auditable and verifiable, the approach and search strategy used for conducting 
systematic mapping are explained. The findings and results from the survey are presented to 
guide the direction of the study. 
3.2 Existing Literature 
Advancement in algorithmic music compositions have made it possible to produce lyrics and 
melodies from real-world data such as academic papers (Scirea, Barros, Shaker, & Togelius 
2015), self-generating composition (Papadopoulos, Roy, & Pachet 2016)and also enhance 
voice and word representation in music composition (Makris et al., 2015; Shin et al., 2017). 
Yet, the complexities and the multi-disciplinary nature of the domain makes it a challenge 
for researchers and practitioners in the domain to realise its full potential. 
20 
University of Ghana http://ugspace.ug.edu.gh
Radhakrishnan et al., (2007) explained that algorithm music is tied with the field of music 
composition since it involves a finite set of precise instructions that performs computation 
or solves a music problem. Nonetheless, it is significant to distinguish algorithm as important 
in music composition since it enables one to know the impact it has on the various music 
styles and genres. The pressing interest in computer-generated music is motivated by the fact 
that contemporary artists, novice and researchers can use computers to compose music 
reduced financial cost, stress and composition time (Chamberlain et al.,2018). 
Although there is success regarding algorithmic composition, literature in the domain is 
scarce and this poses a challenge to researchers regarding the successes, challenges and 
future directions. To address this knowledge gap, a well-established method (Kitchenham, 
2007; Kitchenham & Charters, 2007; Petersen et al., 2015) for conducting systematic 
reviews studies was adopted. The method involves a systematic approach to answering 
research questions that seek to provide relevant outcomes within a specific research domain 
(Da Mota Silveira Neto et al., 2011; James et al., 2016; Petersen et al., 2015).  
Some existing studies have reviewed works on algorithmic music composition. Roads 
(1985) reviewed how computers use AI techniques to analyse music focusing on the 
composition, performance, music theory and digital sound processing. Although the study 
provided a meaningful understanding of how systems can learn and compose music using 
the various AI techniques, it failed to further explain the various types of techniques that are 
applied. Papadopoulos & Wiggins (1999) provided a summary of the early history of the 
field. The review discussed the various AI techniques used in algorithmic composition and 
concluded that systems that apply only one method for composition are not efficient. 
Although Papadopoulos & Wiggins, (1999) review outlined the various techniques, the study 
21 
University of Ghana http://ugspace.ug.edu.gh
was arguably not systematic, since the methods used to select was undefined open to 
biasness. The study also failed to provide discussions concerning patterns and trends of the 
various AI techniques used. 
Fernandez & Vico (2013) provided summaries on AI techniques in algorithmic composition. 
Their study emphasised the techniques, languages and tools used to automatically perform a 
creative compositional task with minimal or no human intervention. Again, their study 
cannot be a systematic review as it failed to discuss the methods used in the selection of the 
literature. Furthermore, the study failed to discuss concerns about the patterns and trends of 
various techniques and languages. 
Similarly, other studies (Lopez-Rincon et al., 2018) have provided summaries on techniques 
in automatic music composition and identified that the main challenge in algorithmic 
composition is performance measurement. Although the review identified the trends, it 
cannot claim to be systematic when compared with the systematic review guidelines as 
defined by Kitchenham & Charters (2007) and  Petersen et al. (2015). They failed to outline 
the method adopted for the paper selection (i.e. inclusion and exclusion criteria, search terms 
and phrases), the databases which were queried, and the data extraction method used. Also, 
even though they reported on known existing AI techniques, trends and performance 
measure, they fail to discuss the approach. Accordingly, a systematic review was conducted 
to validate existing claims by empirically evaluating existing studies to draw conclusions to 
either support or refute such claims. The next section provides the step used for conducting 
the systematic study. 
22 
University of Ghana http://ugspace.ug.edu.gh
3.3 Review Approach 
Figure 3-1 provides an overview of the process followed in the systematic study. The process 
was divided into three phases, i.e. planning, conducting, and documentation. Each phase has 
inputs and output artefacts that were followed to minimize potential bias. The artefacts were 
assessed and by external experts for independent views and their comments and 
recommendations were adopted. This led to a review protocol for the study. 
 
Figure 3-1 Steps in the Review Process 
3.3.1 Planning, Conducting, Study Selection and Documentation 
To perform a systematic mapping study on algorithmic composition, the research questions 
were identified with a well-defined review protocol evaluated and accepted to further 
describe the various steps the study needs to follow. The Conducting phase involved a search 
that combines the various techniques applied in identifying a comprehensive set of candidate 
studies on algorithmic music composition. An outline consisting of a set of parameters for 
23 
University of Ghana http://ugspace.ug.edu.gh
analyzing the primary studies was also defined. The final phase involved reasoning to obtain 
findings to evaluate the quality of the study. Some activities performed include a systematic 
explanation of the data extracted in the previous phase to achieve results that relate to the 
context, analysing the evaluating techniques and writing a detail report describing the 
systematic mapping study conducted. 
3.4 Review Questions 
The Goal-Question-Metric perspectives were adopted to guide the review process (see Table 
3-1). The purpose, issue, objective, view point were formulated before identifying and 
outlining the research questions as suggested by Basili (1992). This approach ensures a 
precise and concise review questions tailored to fit the knowledge gap (Brereton et al., 2007). 
Based on the goal, the review questions in Table 3-2 were crafted. 
Table 3-1 Goal-Question-Metric adopted from (Basili et al., 1994) 
Purpose Analyze the Review Questions 
Issue Publication trends, AI techniques, type of music files used and validation 
strategies 
Object of existing algorithmic music composition 
Viewpoint from a researcher´s point of view 
A search strategy was adopted to detect as many relevant studies as possible. Therefore the 
Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) approach 
(Stovold et al., 2014) was used. 
24 
University of Ghana http://ugspace.ug.edu.gh
Table 3-2 Review Questions and their Objective 
 Question Objective 
RQ1 What are the publication trends To classify primary studies to assess interest, 
of studies in algorithmic music relevant venues, and contribution types; 
composition between 2010 and depending on the number of primary studies, 
2019? trends can be assessed over the years 
RQ2 What AI techniques are to analyse and classify all the existing 
dominating in algorithmic approaches for algorithmic music composition 
composition? concerning the specific concerns they want to 
address. 
RQ3 What are the evaluating to analyse and classify all the existing evaluating 
techniques for validating the approaches for algorithmic music composition to 
composition algorithms? the strategies used for assessing their validity 
RQ4 What type of music files are to analyse and classify all the existing music files 
mostly used for automatic used for the algorithmic composition 
compositions? 
 
An automatic search was performed. This included the execution of a search string on a set 
of electronic databases (see Table 3-3 for list of databases used). This approach is mostly 
used due to its potential to identify relevant studies. The keywords Algorithmic music, 
Composition, Machine Learning, Artificial Intelligence and Music Composition were used 
to form different phrases and searched in the databases. 
25 
University of Ghana http://ugspace.ug.edu.gh
Figure 3-2 presents the number of studies realised at each stage of the search process. The 
after-effect of the automatic search indicated that all irrelevant sets that are not stated in the 
inclusion criteria are removed. This include workshops, tables of contents, maps, lists of 
program committee members, keynotes, surveys, tutorial or invited talks, and messages from 
chairs and co-chairs and book chapters are all not included (see Figure 3-2). Only studies 
conducted between 2010 and 2019 were considered. The search resulted in 4413 potentially 
relevant studies. Furthermore, a manual search is also performed on the relevant 
publications. The search performed considers title, keywords and abstract of each study. 
 
Figure 3-2 Number of Studies Identified for the Review 
 
The obtain optimal search strings Zhang et al., (2011) quasi-gold standard (QGS) was used.  
It argues that the search process is valid and acceptable when the results obtained the meet 
the search criteria. 
26 
University of Ghana http://ugspace.ug.edu.gh
Table 3-3 Databases, Websites and Keywords 
Library Website Keywords 
ACM Digital https://dl.acm.org Algorithmic music Composition, 
Library Machine Learning, Artificial 
Intelligent, Music Composition 
IEEE Explore https://ieeexplore.ieee.org Algorithmic music Composition, 
Machine Learning, Artificial 
Intelligent, Music Composition 
SpringerLink https://link.springer.com Algorithmic music Composition, 
Machine Learning, Artificial 
Intelligent, Music Composition 
3.4.1 Selection Strategy 
As illustrated in Figure 3-2, after the search activity, all identified studies were filtered using 
the exclusion criteria specified in the review protocol and stated below. The following were 
considered as exclusion criteria: 
i. Studies not subject to peer review 
ii. Studies not in English 
iii. Studies focusing on methods or techniques not specific to algorithmic music 
composition system (e.g. studies focusing on either the composition components of 
the systems) 
iv. Secondary or tertiary studies (e.g., systematic literature reviews and surveys) 
v. Studies that were not retrievable from the selected databases  
27 
University of Ghana http://ugspace.ug.edu.gh
vi. Studies in the form of tutorial papers, short papers, poster papers and editorials 
3.4.2 Data Extraction 
In analysing each primary study, an outline for classifying research studies on algorithmic 
music composition was defined (Jonnalagadda et al., 2015). Also, parameters for classifying 
AI methods or techniques used in algorithmic music composition were defined. A data 
extraction form that comprised a list of attributes was designed to collect information from 
each primary study. This included the name of authors, authors country, year of publication, 
title, and publication details of the study, methods or technique used for composing the music 
and the method of evaluating the music. The data was organized and summarized from the 
extracts to understand, analyse, and classify current research on algorithmic music 
composition. Both quantitative and qualitative synthesis methods were applied in analysing 
the 49 primary studies. 
3.5 Review Findings and Results 
To assess the publication trends concerning algorithmic music composition, a set of variables 
focusing on the publication and bibliographic data of each primary study is identified. Each 
primary study including its title, authors, authors’ countries, publication year, publication 
venue (i.e., journal, conference) were analysed. 
Error! Reference source not found. presents the distribution of the selected publications o
n algorithmic music composition from 2010 to 2019. It can be observed that there is research 
in the domain that has not been growing in a sturdy pattern. Although the number of studies 
per year have increased, the increase is not progressive. However, from 2014 there was a 
28 
University of Ghana http://ugspace.ug.edu.gh
continuous growth trend over the years, climaxing in 2015 and decreasing again in 2016. In 
2017 the highest number of publications were achieved. Publication score decreased to 9% 
in 2018 but there was an emergent increase again in 2019. Furthermore, it can be concluded 
that the sharp increase in the number of selected studies in the year 2015 and 2107; would 
be trace to innovation of algorithmic music composition techniques gaining interest and 
attention (Liu & Ting, 2016).The study covers papers published before October 2019 and 
for a partial year, it accounted for 17% of the primary studies surveyed. This result further 
confirms the growing attention and need for research on algorithmic music composition. 
10
9
8
7
6
5
4
3
2
1
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019
Years
Figure 3-3 Distribution of primary studies by year 
3.5.1 Publication Trends 
Each primary study used was either published as a journal paper or conference paper. Figure 
3-4 shows the distribution of primary studies over the study period and the publication type. 
29 
Numbers of pubication 
The most common publication were conferences. It accounted for 43 (88%) whiles journals 
were 6 (12%). The number of conference publications on algorithmic music suggests that 
the domain is still at its infancy and thus researchers are actively discussing progress and 
also consolidating ideas. 
9
8
7
6
5
4
3
2
1
0
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
conference journal
Figure 3-4 Distribution of publication venue 
Research on algorithmic music composition is pursued across the Globe. However, the study 
revealed, identified Europe, Asia, North America, South America, and Oceania. Europe has 
(58%) of the publication whiles South America and Oceania have 4% each. It was also 
observed that most of the researcher’s work on algorithmic music composition were from 
Europe and there was no publication from Africa. See Figure 3-5 is a chart for the 
geographical distribution of studies surveyed. 
30 
south america Oceania
4% 4%
North america
7%
Asia 
27% Europe
58%
 
Figure 3-5 Geographical distribution of primary studies 
AI Methods for Algorithmic Composition 
The study indicated that AI techniques for generating algorithmic music are progressively 
becoming popular. With a series of techniques identified, evolutionary algorithms were 
mostly used, and it contributed eighteen (37%) of the primary studies. Artificial neural 
networks and deep learning models followed with eleven (23%) while mathematical models 
covered eight (16%). The heuristic methods were seven (14%) of the primary studies 
whereas grammar models were the least as it contributed five (10%) of the primary studies. 
Table 3-4 shows the various techniques used and the corresponding study that used the 
technique. It was observed that genetic algorithm was the most used evolutionary algorithm. 
31 
It recorded a total of eighteen (18) out of ten (10) (Genetic Algorithms is discussed later in 
this chapter). 
Table 3-4 AI methods for composing music 
Algorithm Used Method Description Study % 
Mathematical The process of developing mathematical [31],[13],[46],[33], 16 
models models and implementing it effect. [49],[23],[29],[32] 
Grammar Grammars are powerful methods which [12],[44],[45],[20], 10 
display unique features in the analysis [41] 
and composition of algorithmic music. 
There exists a degree of hierarchical 
structure in the composition of music 
with rules that connect notes to form 
phrases, segments, and pieces of music. 
Evolutionary An evolutionary algorithm is described [1],[2] ,[3],[4],[5], 37 
methods as algorithms that follow the laws of [16],[11],[7],[35], 
evolution. Evolutionary algorithms [36], [8],[37],[9], 
(EAs) are driven by optimization [38],[24], 
process such as the Darwinian [39],[26],[10] 
evolution. 
Knowledge- The knowledge-based system is a form [43],[14],[19],[40], 14 
based systems of Artificial Intelligence (AI) which [42],[28],[48] 
32 
Algorithm Used Method Description Study % 
intends to use human expert knowledge 
to aid in decision making. 
Learning systems Applications including systems which [47],[15],[18],[34], 11 
learn can be used to compose music. [22],[30],[25],[27], 
Systems which learn use current [17], [6], [21] 
information available but not previous 
knowledge to compose music. These are 
grouped into an artificial neural network 
(ANN) and ML based on how they 
perform their task and store 
information.  
 
3.5.2 Evaluating Techniques in Algorithmic Composition 
Evaluating techniques are used to validate the formal mathematical proof used to ascertain the 
appropriateness of the model. In all a total of 9 different techniques were found to have been 
used for evaluating the use of algorithms for composing music. From Error! Reference s
ource not found. eighteen representing 37% of the studies used qualitative observation 
(listening test) for assessing the quality of the composed music. This is followed by Fitness 
function which is a method for selecting individuals to mate in a GA (discussed later in this 
chapter). Ten studies used fitness functions alone while 5 studies combined fitness function 
and listening for evaluating the resultant music. Seven of the studies failed to report the 
method used for evaluating the composed music.  
33 
Table 3-5 Methods for Evaluating Algorithmic music 
Type of Method Description Study % 
evaluation 
Fitness value The value of the fitness function for that [1], [2],[3],[4],[5], 26 
an individual (the fitness function is [6],[7],[8],[9],[10] 
how close a given solution is achieving [11] 
its aim) 
By observation Observation is an assessment method [12], [13], [14], [15], 35 
(listening) which provides information about the [16], [17], [18], [19], 
quality, strengths, weakness and interest [20], [21], [22], [23], 
of a thing. [24], [25], [26], [27], 
[28], [29], [30], [31] 
Up sampling Up sampling is the process of increasing [32] 2 
and down sample by inserting zero-valued whiles 
sampling down sample is the process of reducing 
the sampling rate by introducing 
integers. 
Markov’s It is a mathematical framework for [33], [34] 5 
process modelling decision in situations where 
(Probability) the outcomes are partly random and 
 partially regulated by a decision maker. 
Application of This is a combination of the fitness [35], [36], [37], [38], 12 
both fitness function and human judgement [39] 
34 
Type of Method Description Study % 
evaluation 
function and (listening) in order to assess the quality 
listening of a melody. 
Test run A test run on a prototype is to evaluate [40] 2 
(prototype) the design, product or creativity 
generated. 
Precision and Using precision and recall measure in [41] 2 
recall measures evaluation is to measure the success of 
 prediction when there is an imbalance 
situation. 
Benchmark A function which is used to test the [42] 2 
functions performance of any optimization 
problem. 
Not stated - [43], 14 
[44],[45],[46],[47], 
[48] 
 
Other methods such as rest measuring melodic and harmonic errors, application of frequency 
spectrum, up sampling and down sampling, harmonic evaluation function, precision and recall 
measures, continuous optimization functions, subjective evaluation and multi-attribute 
decision-making techniques were all identified. It can be observed that the most used 
technique for the evaluation of the performance of the various algorithm is qualitative 
observation (i.e. listening test). In this approach, listeners assess the quality of a composed 
35 
music by listening to and classifying it as appealing or not. Although this method is 
appropriate some researchers argue that the use of fitness function and human evaluators is 
more effective. 
3.5.3 Music File Used for Composition 
Considering the complexity exhibited in the representation of music scores by computers, 
music files are created to make composition less challenging. They create flexibilities for 
research and artist during composition.  
From Figure 3-6, thirty (30) representing 61% of the primary studies used did not specify the 
type of music file used. This is worrying considering that the domain is relatively new and 
thus researchers in the field need adequate information from existing studies to support their 
work. The most used music file was Musical Instrument Digital Interface (MIDI). In one 
instance the pdf file displaying MIDI notations (or instructions) was used. MIDI file is a 
standard protocol that governs the interaction of digital instruments using computers. They 
are more compact than digital audio files and allows for easier editing. It is therefore envisaged 
that these properties enable it to outperform other audio files in automatic compositions. 
From the review, it is evident that genetic algorithms and MIDI files are used predominantly 
for algorithmic compositions. Accordingly, the next section in this chapter discusses some 
relevant details on them. 
36 
35
30
25
20
15
10
5
0
MIDI PDF to  MIDI MUSIC XML FILE NOT STATED
 
Figure 3-6 Types of file Used for Composition 
3.6 Musical Instrument Digital Interface (MIDI) 
Before the development of Musical Instrumental Digital Interface (MIDI) digital musical 
instrument from the various manufacturers was unable to communicate with each other. 
However, the evolution of MIDI and standardization permitted interoperability and 
exchange of information with other musical instruments. Hardware including computers, 
sequencers and sound module from different manufacturers were able to communicate with 
each other. MIDI is a protocol which allows electronic instruments and other digital musical 
devices from different manufacturers to communicate with each other. It is worth mentioning 
that MIDI is not a music file, but rather a set of instructions that enables a device to play 
music. The quality of music produced by a MIDI file is relative to the output device. The 
electronic keyboard is one of the most popular devices used in generating MIDI messages.  
37 
In generate a MIDI file, MIDI INPUT port receives data that is transmitted from another 
MIDI device whiles the MIDI OUTPUT port transmits MIDI data to another MIDI device. 
Also, a MIDI THRU sends out a copy of an incoming MIDI data. The messages are 
channelled unto a computer or any MIDI instrument. Figure 3-7 is a diagrammatic 
representation of how various devices may be connected to a computer to generate a MIDI 
file (.mid). 
 
Figure 3-7 Connected keyboard, PC and one generator for MIDI 
By pressing or selecting a key on the keyboard, the note on message and the note-off 
messages are generated. Notes have two linked parameters; the note number and the velocity 
number. The note number indicates the pressed key (with the middle C note having a note 
number of 60). Note- on messages are channel voice messages, therefore, uses one of the 16 
channels. A Channel voice message uses the performance event transmitted from one 
instrument to the other using any of the sixteen (16) MIDI channels. A selected key illustrates 
the note information in bytes showing the MIDI channel, MIDI pitch “note” and the velocity. 
Pressed notes have values between 0-127, channel (1-16) and the velocity (0-127) describing 
38 
how hard the key is pressed. The velocity denotes the loudness of the sounding note which 
increases in volume with higher velocity values. Furthermore, the release of a key from the 
keyboard shows a note- off message is created which contains information about the “note” 
and also indicating the end to that specific MIDI note been played. 
 
Figure 3-8 A Byte structure of a MIDI file 
The note-off message stops the notes-on message from playing the note. However, the note-
off messages contain information such as the note -off MIDI channel, the MIDI pitch “note” 
and the note-off velocity. The note-off velocity indicates the speed or rate at which the key 
is released. A low value expresses the slow release of the key and vice versa. Once the note-
on velocity is zero, it is the same showing a note-off message and therefore a current sound 
note will have a velocity of zero. Figure 3-8 shows the byte structure of a MIDI file. 
3.6.1 MIDI File Format 
MIDI files are not sound or audio files (MP3 files) but a set of instructions or messages 
events. A MIDI performance can be recorded, stored in a data format and replay later. The 
file contains details about the data describing the various musical events including the note 
on, note off, channel pressure, key pressure, program change, pitch bend and control change. 
39 
Furthermore, since MIDI recordings do not record sound, but performance data, MIDI 
instrument (synthesizer or keyboard) interpret the data to produce sound. A MIDI recording 
which includes the performance related to a keyboard could sound like a keyboard 
instrument and that of a flute will sound like a flute because MIDI only records the messages 
and information about the notes and not the specific sound. When a MIDI instrument is 
played the entire key pressed, button pressed, the knob turned, and slider changes are 
converted into MIDI data. MIDI file presents an advantage of allowing users to edit notes 
such as changing pitch, duration or other parameters without necessary recording again. 
Also, it allows the playing of different instruments. 
3.6.2 Aftertouch, Pitch bend and Channels 
Aftertouch is the pressure sensitivity applied to a key. Similar to a note's pitch and velocity, 
it ranges from 0-127. It determines the pressed key and assigns a value. There are two types 
of aftertouch in MIDI i.e. the polyphonic aftertouch (which allows each key to respond 
individually to aftertouch) and monophonic aftertouch (all the keys played on the keyboard 
can respond equally to the aftertouch). It is noted that not all electronic keyboards have that 
functionality of producing aftertouch messages as well as not all MIDI instruments also 
provide that support. 
The pitch bend is a message control by the pitch bend wheel, and it is attached to the MIDI 
keyboard. It modifies the pitch of the notes being played either up or down. Pitch bend values 
range from 0-16,383 steps. The control range is between -8192 (lowered) and 8198 
representing fully raised pitch with 0 being an unaltered pitch of an instrument. 
40 
MIDI channels are useful in sending messages or notes to other MIDI instruments and it has 
16 channels. Each MIDI instrument can correspond to a MIDI channel when selected so that 
data can be shared. Additionally, different instruments can be set to communicate with the 
MIDI channels since they operate independently to produce a multitrack song. 
3.7 Genetic Algorithms 
Genetic algorithms (GAs) are search heuristic that uses the concept of evolutionary biology 
or mimics the natural evolution process. It uses probability search procedures designed to 
work on large spaces that can be represented as strings (Goldberg & Holland, 1988). GAs 
are used in solving optimization problems. It finds the optimal solution from all feasible 
solutions. GA uses a population of individuals where each individual represents a potential 
solution and each population is a subset of the total search space. Hence, they are population-
based models and they use iterative processes that employ selection, mutation and 
recombination operators to produce new and better individuals (it considers individuals who 
can survive and reproduce as well as move to the next generation i.e. survival of the fittest). 
A set of individuals known as the population is represented as a chromosome. The 
individuals are characterized by a set of parameters (variables) known as Genes. Genes form 
strings to form the chromosome (they are represented as a binary vector and each item in the 
vector is a gene). Each individual is represented in binary form and is assign a fitness score. 
A fitness function is used to calculate the fitness score and determine how well the problem 
can be optimized by an individual or chromosome (Mitchell, 1995). For each new population 
created, an individual fitness score is calculated, and the reproduction continues until the 
required criterion is met.  
41 
3.7.1 Selection, Cross Over, Mutation and Termination Process 
Two individual variables are selected as parents from 𝑉	 ∈ $𝑣!, 𝑣",𝑣$⋯𝑣&'. The fitness score 
of each variable 𝑣& is calculated to select the most suitable parents for mating. The Crossover 
rule combines two parents to reproduce children for the next generation. See fFigure 3-9 for 
details of the selection and cross over process. 
 
Figure 3-9 Cross-over and mutation 
Mutation applies random changes to individual parents to form children and this occurs after 
the crossover is performed. It is intended to prevent the falling of all solutions in the 
population into a local optimum. It randomly changes the children produced as a result of 
the crossover. In binary encoding, probabilities are used to switch chosen bits from 1 to 0 or 
from 0 to 1 as illustrated in fFigure 3-9. The generation process terminates when there is no 
improvement in the population for the iteration process to continue or when the required 
number of generations is achieved or when the objective function value has reaches it pre-
defined value. 
42 
 
function GeneticAlgorithm () 
   {  
Initialize population ()  
Define fitness 
Calculate fitness function; 
    While (fitness value != termination criteria) 
 {  
Selection; Crossover; Mutation; Calculate fitness function; 
Figure 3-10 Sample Pseudo code for GA 
3.7.2 Areas of Application 
GAs are used in solving complex optimization problems. Ghaheri et al., (2015)  explained 
that they are effective in the healthcare problems as they provide optimal solutions to 
complex medical analysis within a reasonable time. They are also applied in radiotherapy 
(Hole et al., 2013) for improving image quality and segmentation .They provide a good 
perception of the image during such activities as well as the detection of breast cancer 
(Karnan & Thangavel, 2007). Genetic algorithm is also useful in the selection of relevant 
molecular descriptors (Zandkarimi et al., 2014). In more recent times, they have been used 
in robot trajectory (i.e. planning the route that a robot might take from moving from one 
destination to the other). Wu et al., (2018) argued that they are effective for improving 
tracking accuracy of the singularity- robust algorithm. They support planning of robot 
motions and pre-processing of path methods for non-differentiable paths and also for 
avoidance of obstacles (Kunz & Stilman, 2013). 
43 
They have been demonstrated to effective in aircraft design (Dong et al., 2019; Marta, 2003) 
and economics (Brooks, 2000; Dawid & Kopel, 1998; Nicoară, 2015; Riechmann, 1998). In 
music composition GAs are used to determine optimal melodies (as established earlier in 
this chapter). 
3.7.3 Advantages of Genetic Algorithms 
Genetic algorithms exhibit various advantages including the provision of a possible solution 
to various search-space problems. Since each point in the problem represents a feasible 
solution, the use of robustness property enables it to generate a solution within a relative 
short time. Furthermore, it exhibits good comparability proficiencies and useful in situations 
where the search space is large and with a large number of parameters. They do not require 
derivate information which may not be available for many real-world problems but rather it 
optimizes both continuous and discrete functions as well as multi-objective problems. 
3.7.4 Limitations of Genetic Algorithms 
Similar to any other techniques, genetic algorithms suffer some limitations. During 
selections, GAs need to be designed carefully, since incorrect choice of operators affect the 
outcome of the application. In such situations, some optimal solutions are not covered. GAs 
are not suitable for simple problems which have readily available information (Bajpai & 
Kumar, 2010). Some researchers have argued that the computation of fitness scores can be 
computationally expensive for some problems (Leblanc et al., 2002) . 
44 
3.8 Chapter Summary 
This chapter provided a review of literature on the application of algorithms for music 
composition. It discussed the current state of research by highlighting on the dominant 
algorithms, techniques and methods of evaluating algorithmic music. It was concluded that 
the most dominant algorithms are evolutional algorithms. Performance evaluation continues 
to be one of the major challenges of algorithmic music composition: the ability to effectively 
evaluate the efficiency of the algorithm as well as the music. It was observed that most studies 
used human evaluation (listening test) as the preferred method. Thus, it can be inferred that 
less attention is given to model performance testing (fitness function evaluation).  
Majority of existing algorithmic compositions are achieved using genetic algorithm. Also, the 
most used music file was MIDI. Accordingly, further discussions on MIDI and genetic 
algorithms were presented. It was argued that the flexibility of MIDI (its ability to instruct 
systems to play melodies on different instruments), makes it a suitable file format for 
algorithmic composition of music. Genetic algorithm was also suitable because it is effective 
in solving complex problems that has a large number of variables. From the above, this study 
seeks to generate a monophonic music using genetic algorithm in a MIDI file format. In the 
next section, we discuss the research approach adopted for the study. 
45 
Chapter 4 
 METHODOLOGY 
4.1 Overview 
This chapter explains the research method used for the study. It presents the design science 
methodology as the appropriate approach for the study. It starts with a discussion on the basic 
beliefs and paradigms for conducting scientific studies. This is followed by a discussion on 
the design science method and how it supports this study. In line with suggestions made by 
(Hevner et al., 2004; Kuechler & Vaishnavi, 2008) the problem relevance, suggestions, 
development and evaluation of the artefact was discussed. 
4.2 Paradigms and Beliefs in Algorithmic Music Composition 
Discussions on the use of appropriate scientific methods in computer science and engineering 
continue to be a challenge, particularly, the use of appropriate methods in algorithmic music 
compositions. This can be attributed to the complexity and multi-disciplinary nature of the 
field. Algorithmic music composition combines disciplines such as mathematics, computer 
science, music, engineering and psychology. Some researchers argue that algorithmic music 
composition groupings are mostly mathematical or computer science approaches (Nierhaus, 
2009) thus it is appropriate to use methods from such disciplines. That notwithstanding, the 
appropriateness of a study depends on the beliefs and paradigms that guide the researcher. 
Beliefs are rational and they provide the philosophical assumptions for conducting research. 
According to Vaishnavi & Kuechler (2004), epistemological, ontological, methodological and 
46 
axiological are the fundamental beliefs observed in computer science research. Ontological 
beliefs are concerned with identifying research problems and examining the degree of 
improbability surrounding the occurrences. Epistemological belief probes the nature of 
knowledge by considering its firmness and how it is theorized, whilst axiological belief 
discusses what constitutes values or what is considered to be good or bad, like or dislike (Hart, 
1971). Methodological belief, on the other hand, deals with a set of guidelines and it expedites 
the understanding of viewpoints for better understanding. 
The foundation of this study is based on the premise that complexities in music production 
(i.e. the arrangements of notes at variable pitches to form a melody) can be minimised using 
mathematical algorithms. Given any set of particular notes, there is an optimal melody 
produced in which human judges the quality. Thus, the study is founded on a combination of 
ontological and methodological belief and adopts the design science method. 
4.3 The Design Science Method 
Some researchers have advocated for the use of design science method in computer science 
(Hevner et al., 2004), since it is considered to be a rigorous process used for designing artefacts 
for solving problems, evaluating the solution and communicating the outcome (Çaǧdaş & 
Stubkjær, 2011). Vaishnavi et al., (2012) explained that design science is a set of analytical 
methods that promote development of research in various sectors. It focuses on solving 
specific problems (Van Aken, 2005) in order to achieve “good” solutions. It provides results 
such as methods, models, constructs and instantiations (March & Smith, 1995). It can be 
considered as a science, a research strategy and a method (Johannesson et al., 2013). 
47 
Hevner et al., (2004) proposed guidelines for conducting research when using the design 
science approach. These are: 
i. Design as an Artefact: a suitable artefact should be produced in the form of a 
construct, technique, instantiation of or a representation 
ii. Problem relevance: to develop a technology-based solution that is applies to the 
problem 
iii. Design evaluation: the utility and quality of the design artefact must be meticulously 
and well evaluated 
iv. Research contributions: clear and demonstrable contributions must be made in the 
field where the design artefact is applied 
v. Research rigour: appropriate methods for the design and evaluation of the artefact 
vi. Design as a search process: the search for a successful artefact depends on the use 
of available methods to achieve a desirable result while satisfying the rules of the 
problem domain 
vii. Communication of research: design science research presentation must be effective 
in terms of its technology-orientation and management-orientation 
4.4 Limitation of design science method 
Although design science presents a number of benefits, it is also characterised with some 
limitations. Design research seeks to develop knowledge that is used to solve problem (Van 
Aken, 2005), yet it lacks formalized details of the process (Offermann et al., 2009). Sein et 
al., (2011) explained that it concentrates on building artefact than its evaluation. In some cases, 
it concentrates on building artefacts and values technological rigour at the expense of 
48 
organizational relevance. Some have argued that the method is not able to establish sufficient 
and convincing rigour in its approach (Purao, 2002). 
Although there are some limitations, the method was appropriate because it provides a 
thinking frame for this study. Next is a discussion about the various steps that were followed 
for conducting the research. 
4.5 Research Approach for the Study 
In order to follow an applicable method in this research, guidelines proposed by Vaishnavi & 
Kuechler (2004) is implemented. Accordingly, the following five phases were used in the 
research. 
4.5.1 Awareness of the problem 
Several issues may prompt problem awareness and it is important to find appropriate 
solutions. Recently, the concept of developing new solutions for known problems has resulted 
in rise of new technologies. Swift changes provide new ways of thinking and designing better 
solutions to solve known problems (Persson et al., 2000). 
As stated earlier, the research seeks to develop an artefact (AI model) for composing music. 
This addresses some current existing problems in composition of aesthetic music. 
Accordingly, the study is within the improvement quadrant. See Figure 4-1. 
49 
 
Figure 4-1 Relevance of the study 
4.5.2 Suggestions 
The use of rigorous methods in composition and design of artefacts is important in design 
science (Hevner et al., 2004; Kuechler & Vaishnavi, 2008). During the suggestion phase, GA 
and existing music composition theories were considered. The application of knowledge and 
theories on music composition provides an in-depth understanding of using GA techniques 
for music composition. Moreover, relevant literature relating to existing methods in 
algorithmic music composition was considered. Thus, a systematic review on algorithmic 
music composition was conducted to facilitate the identification of existing methods that can 
be used in the research. The suggestion phase also assisted the identification of appropriate 
methodology for the research. 
50 
4.5.3 Development 
At the development phase searching for the optimal solution to solve the problem is the best 
approach (Hevner et al., 2004). To search and identify an applicable technique that can be 
used for the composition of music, a GA model for music composition is identified and 
programmed. The decision to use GA was informed by the study conducted in Chapter 3 that 
identified it as the most used algorithm for developing algorithmic music. The development 
phase is reported in chapter five and addresses objective ii. 
4.5.4 Evaluation 
The evaluation of a design artefact in design science is  vital and also an essential activity  that 
confirms that the artefact produce addresses a relevant problem (Peffers et al., 2012; Venable 
et al., 2012). It provides guidance for a better development or improvement of the artefact. 
Experimentation, observation, testing, analytical, descriptive can be used to evaluate artefacts 
(Hevner et al., 2004; Kuechler & Vaishnavi, 2008). 
In this study, the design artefact was evaluated using experimentation and observation. The 
fitness score of compositions and the human evaluators were used to assess the quality of 
music produced. Fitness function guides attainment of optimal solutions in GAs. It 
demonstrates an enhanced role by accepting the best melody and decreasing the lack of 
harmony in a musical note (dissonance) in the resultant music. 
4.5.5 Conclusion 
Conforming to suggestions that results of design research must be communicated to the 
appropriate audience (Hevner et al., 2004; Kuechler & Vaishnavi, 2008). This study has been 
51 
communicated to a number of scholarly communities. Sections of these have been accepted 
for publication. 
4.6 Chapter Summary 
The need to ensure that scientific studies are auditable and can be refuted cannot be 
overemphasised. It is a basic requirement of all studies and accordingly this section provided 
foundation and steps used for conducting the study. It adopted the design science approach 
and argued that it is the most appropriate method for conducting the study. Justification for 
the awareness of the study and the evaluation method adopted were all discussed. 
The next chapter presents the findings from the study. It highlights the steps used in 
composing and evaluating the resultant music. 
  
52 
Chapter 5 
COMPOSITION MODELLING AND FITNESS ANALYSIS 
5.1 Overview 
As mentioned earlier, this study seeks to generate a monophonic music using genetic 
algorithms. Accordingly, the chapter presents the various steps used for composing the music. 
Six monophonic melodies were composed using GAs terminating at generations 10, 000, 
20,000 and 60, 000.The chapter starts with a discussion on how the population was selected, 
scored and crossed. The method for mutation is also discussed. 
Additionally, the chapter discusses how fitness scores were used to assess the performance of 
the various composition models. It is emphasised that the use of fitness scores seeks to 
evaluate the performance of the model rather than the quality of music produced. 
5.2 Genetic Algorithm Implementation 
The algorithm used two types of operators i.e. crossover and mutation to generate music. As 
explained earlier, the need to mutate after crossing ensures the use of best individuals for 
future crossing. This ensures that a good “choice” is attained to create a new and a better 
music. Figure 5-1 shows the steps used for composing melodies in this study. 
5.2.1 Composition Modelling 
Conforming to “objective iii” of the study, a set-up was made to compose a simple 4/4 bars 
compositions. Nine  compositions were made. These compositions are 3 versions of melodies 
terminated at 10000th, 20000th and 60000th generations. Refer to Figure 5-2 to Figure 5-19 for 
53 
the various sheets songs and their corresponding fitness score charts. The different music 
versions are attached to the electronic version of this thesis in a MIDI file format. The GA 
took input parameters of pitches, tones, iterations, number of generations, and the range of 
tones allowed.  
Considering music composition as an optimal search problem with a population of melodies, 
then a population of melodies is defined as 𝑀&: 𝑛 ∈ {1,2,3…100}  i.e.  limiting the population 
size to 100. The population size of 100 was used to reduce computation time and complexities. 
In genetic algorithms, large populations affect computation time because it increases the 
search space. 
For each 𝑀&, 𝐹' = 𝑓((𝑀&)  where  𝐹'	 is defined as the fitness value (score) of melody,  
𝑀& for generation  𝐺( in the genetic evolution where 𝑖 ∈ {1,2,3…ψ}	and 
𝑓((M)) = 𝛾& + 𝜇& + 𝜍& + 𝜑& + 𝜗&																																																									 
Where	
𝛾& 	= 𝑐ℎ𝑜𝑟𝑑	𝑝𝑟𝑜𝑔𝑟𝑒𝑠𝑠𝑖𝑜𝑛	𝑠𝑐𝑜𝑟𝑒 
𝜇& =	diatonic fit score 
𝜍& = voice leading score 
𝜑&*smoothness score 
𝜗& = 𝑐ℎ𝑜𝑟𝑑	𝑠𝑐𝑜𝑟𝑒 
Total fitness 𝐹(𝐺() = ∑+&*( 𝑓((𝑀&) 
54 
Refer to Table 2-1 for the meanings of the various terminologies. During the calculation of 
the fitness value of a melody, its chord progression, diatonic fit, voice leading, and smoothness 
are assessed. Defined values are assigned to each of these variables based on its 
appropriateness. 
Mutation types is defined as 𝜔, ∶ 	𝜏 ∈ {1,2,3,4,5,6,7} where 
𝜔! = ℎ𝑎𝑙𝑓	𝑠𝑡𝑒𝑝 − 𝑢𝑝, 𝜔" = ℎ𝑎𝑙𝑓	𝑠𝑡𝑒𝑝 − 𝑑𝑜𝑤𝑛, 𝜔$ = 𝑤ℎ𝑜𝑙𝑒	𝑠𝑡𝑒𝑝 − 𝑢𝑝, 𝜔- =
𝑤ℎ𝑜𝑙𝑒	𝑠𝑡𝑒𝑝 − 𝑑𝑜𝑤𝑛, 
𝜔. = 𝑜𝑐𝑡𝑎𝑣𝑒 − 𝑢𝑝, 𝜔/ = 𝑜𝑐𝑡𝑎𝑣𝑒	 − 𝑑𝑜𝑤𝑛 and 𝜔0 = 	𝑓𝑙𝑖𝑝	𝑟𝑒𝑠𝑡 
The genetic operators initiate a random population of melodies 𝑀& and for each 𝑀&, it 
calculates the fitness score 𝐹'. The number of generations is compared to the defined 
maximum number of generations. In this study, three (3) maximum generations are defined 
(i.e. 𝐺123 = 10000, 20000 or 60000). If the number of generations 𝐺(	 = 𝐺123 then the process 
is terminated else two melody parents are selected to cross. Best melodies are selected based 
on the value of 𝑓((𝑀&). 
Next, each melody is split into two for crossing.  The splitting is based on the splitting rule 
define below: 
There are three (3) types of melody splitting 𝑆!, 𝑆"	𝑎𝑛𝑑	𝑆$, for parent melodies 𝑀& and 𝑀&5( 
Where i is an arbitrary number such that 𝑛 + 𝑖 ≤ 	100, 𝑖 ≠ 	𝑛 
Then 
𝑆!𝑖𝑠	𝑎	50, 50	𝑠𝑝𝑙𝑖𝑡 (50% of 𝑀& and 50% of 𝑀&5() 
55 
𝑆"	𝑖𝑠	𝑎	25, 75	𝑠𝑝𝑙𝑖𝑡	(25% of 𝑀& and 75% of 𝑀&5() 
𝑆$	𝑖𝑠	𝑎	75, 25	𝑠𝑝𝑙𝑖𝑡 (75% of 𝑀& and 25% of 𝑀&5()  
 
Figure 5-1 GA music composition process used for the study 
The roulette wheel is used to select the splitting method. The all melodies in the population, 
𝑀& and 𝑀&5( are crossed randomly and mutated (based on the mutation probability) to 
produce a new offspring to form a new population (new 𝑀&). New offspring is produced until 
total population of offspring is 100. The fitness scores of all 𝑀& in the new population is 
calculated and the selection process is repeated until the desired generation is achieved. Refer 
to Figure 5-1 for the steps in the composition process. 
5.3 Resultant Melodies 
There were nine melodies generated in all based on the termination rule used. As mentioned 
earlier, the algorithm was terminated at three different generations (i.e. 𝐺123 = 10000, 20000 
56 
or 60000). For the purpose of evaluating the efficiency of the algorithm and the authentic 
quality of the music, three compositions were created for each termination value. The fitness 
scores of the various melodies composed are discussed in the next section. 
5.3.1 Compositions at 10K Generations 
At a composition of (𝐺123 = 10000) generations, the first version achieved an optimal 
composition with a fitness score of 10348 at a generation of 750. The second version also 
achieved a fitness score at 10347 at a generation of 7300 and lastly, the third version had a 
fitness score of 10341 at a generation of 5300. It is observed that at a composition of 10k 
generations, the first version obtained optimality earlier having achieved the optimal fitness 
score of 10348 with a generation of 750. Refer to fFigure 5-2, Figure 5-4, and Figure 5-6 for 
the various sheets music of the three compositions terminated at 10K  generations. The sheet 
music is in Bass- Clef. 
 
Figure 5-2 Resultant composition after 10 K generations (v1) 
The fitness score graph (see Figure 5-3) shows the value of fitness scores versus the 
generations (note: the generations axis for all the graphs is in a 𝑙𝑜𝑔!6 scale). It can be 
observed that fitness values increased steadily until generation 150 and drops. At generations 
57 
between 200 and 250, the fitness score also increased from 2651 to 5224, exhibiting two local 
maximum points before it converged at about 750 generations with a score of about 10300. 
10000
8000
6000
4000
2000
0
50 500 5000
 
Figure 5-3 Total fitness scores for 10k generations (v1) 
Observe also that a generation of 650, the fitness score swiftly increased before converging. 
This performance is different from that of the second composition (v2). Figure 5-5 shows a 
steady increase until 100 generations with a fitness score of 2617 but experienced a decrease 
in the fitness score from generation 150 to 950. 
 
Figure 5-4 Resultant composition after 10K generations (v2) 
 
58 
At a generation of 1000, the fitness score increased to 5220 and again became steady until it 
declined to 2653. However, with a generation of 2800, and a fitness score of 5216, the fitness 
score increased dramatically to 10345 at a generation of 4450. 
 
10000
8000
6000
4000
2000
0
50 500 5000
 
Figure 5-5 Total fitness scores for 10K generations (v2) 
Between the generation of 5100 and 5400 the fitness score slowly decreased again from 5223 
to 5229. Furthermore, at a generation of 5500 the fitness score increased steadily but decreased 
again at a generation of 7100 with a fitness score of 10350 before converging generation 7300.  
 
Figure 5-6 Resultant composition after 10K generations (v3) 
 
59 
10000
8000
6000
4000
2000
0
50 500 5000
 
Figure 5-7 Total fitness Scores for 10K generations (v3) 
As compare to Figure 5-3 and Error! Reference source not found., Figure 5-7 shows a 
gradual increment in the generation from until 350. At a generation of about 400 compositions 
the GA attained a fitness score 1364, before increasing steadily to a fitness value of 2652 at a 
generation of 450. It decreased again at a generation of 500 and increased sharply to a score 
of 5218 at a generation of 600. From generation of 650 to 2150 the fitness score is steady with 
less change in the score however, between generations of 2200 and 2300 it increased to 10341. 
At a generation between 4250 and 5050, there is a decrease in the fitness score to 5213 and 
5220 respectively. The composition algorithm converged at generation of 5300 with a fitness 
score of 10341. Evidently, the GA performed differently in all the 3 trails for the first 10k 
generations with less similarity. 
5.3.2 Compositions after 20K generations 
At the 20K generations (i.e. 𝐺123 = 20000), the first version achieved it highest fitness score 
at a generation of 1730 whereas the second version achieved at a generation of 4000. The third 
60 
version was at a generation of 1700. Refer to Figure 5-9, Figure 5-11 and Figure 5-13 for the 
evolution behaviour of the genetic algorithm.  
 
Figure 5-8 Resultant composition after 20K generations (v1) 
10000
8000
6000
4000
2000
0
100 1000 10000
 
 
Figure 5-9 Total fitness Scores for 20K generations (v1) 
It is observed that in the 20000 generation trails, the earliest optimal value was obtained at 
1700 generations as compared to the compositions for 10000 generations. The graph (see 
Figure 5-9) shows a gradual increase in the generation the first 500 generations. At generation 
600, it dropped before increasing again at a generation of 700 with a fitness score 2653. The 
GA decreased at generation 1300 with a fitness value of 2653 and steeply increased to 5220 
at generation 1500. At a generation of 1700 there is a sharp increase in the fitness score from 
5220 to 10349 at a generation of 1800 (it converged around this generation). 
61 
 
Figure 5-10 Resultant composition after 20K generations (v2) 
10000
8000
6000
4000
2000
0
100 1000 10000
 
Figure 5-11 Total fitness Scores for 20K generations (v2) 
Figure 5-11 illustrates the GA performance for the second version of the 20k composition. It 
can be observed that there was a gradual increase in the generation from 100 to 200 
demonstrating a steady fitness score between 727 and 1370. The fitness score became steady 
until a population of about 400 where it fell slightly but dramatically increased at a generation 
between 500 and 600. At a generation of about 4000 the fitness score attained its highest value. 
 
62 
 
Figure 5-12 Resultant composition after 20K generations (v3) 
10000
8000
6000
4000
2000
0
100 1000 10000
 
 
Figure 5-13 Total fitness Scores for 20K generations (v3) 
Figure 5-13 shows the behaviour of the GA over 20k generations. It can be observed that GA 
behaved differently from previous ones. At generation of 500 a fitness score of 2647 was 
achieved and fell before rising between the generation of 1400 and 1600. The optimal fitness 
value was obtained at generation of 1700. Thus, the third trail performed better than all 
previous versions in the 20k trails. 
63 
5.3.3 Compositions after 60K generations 
For the 60k composition trails (i.e. 𝐺123 = 60000), the first version achieved it highest fitness 
after 11,000 generations. However, the second was around 1800 and the third at 2400. 
Similarly, to observations made in the earlier trails (i.e. compositions for generations up to 
10k and 20k), the GA performed different in all instances.  
 
 
Figure 5-14 Resultant composition after 60K generations (v1) 
 
10300
8300
6300
4300
2300
300
300 3000 30000
 
Figure 5-15 Total fitness Scores for 60K generations (v1) 
The best performance for the 60k generations was the second version where it converged 
around 1800 generations. Figure 5-15 shows a steady increase in fitness values before 600 
64 
generations and a sharp increase at generation 10500 before attaining optimality around 
generation 110000. 
 
Figure 5-16 Resultant composition after 60K generations (v2) 
10000
8000
6000
4000
2000
0
300 3000 30000
 
Figure 5-17 Total fitness Scores for 60K generations (v2) 
In the second trail the graph (Figure 5-17) shows a steady increase in fitness at the beginning 
until generation 600 with a score of 5219. At generations ranging from 900 to 1200 the fitness 
score remained steady from 5217 to 5218, before increasing to attain the optimal fitness value. 
65 
 
Figure 5-18 Resultant composition after 60K generations (v3) 
10300
8300
6300
4300
2300
300
300 3000 30000
 
 
Figure 5-19 Total fitness Scores for 60K generations (v3) 
In the third version (Figure 5-19) the GA behaved similar to the second. Its fitness increased 
to 5208 by the 600th generation (compare to 5219 for the second version) however it 
remained steady for a longer period as compared to that of second version. It converged at a 
later generation (i.e. 2400) when compared to the second generation. Thus, although v2 and 
v3 for the 60k generation trail started behaving similarly, the performance of the GA 
differed. 
66 
5.4 Chapter Summary 
This chapter presented the steps used in the composition process. The sheet song of the 
resultant composition is displayed, and the total fitness scores are discussed. The relationship 
between total fitness and the generations are explained to assess the performance of the 
model in all the nine compositions. The findings indicated that the first version of the 10k 
composition (i.e. 𝐺123 = 10000) converged faster than the other compositions. 
  
67 
Chapter 6  
EVALUATION AND DISCUSSIONS 
6.1 Overview 
The chapter discusses the importance of the results achieved during the study. It further 
provides a new in-depth understanding of the composition of melodies using GA. A 
discussion on the evaluation performed using human judgement validated the experiment 
set-up since the quality of the music produced was pleasing to the ear. 
6.2 Human Evaluation 
It is imperative to conduct human evaluation in compositions because the existing methods 
for evaluating music quality are limited in evaluating the aesthetic quality of music 
(Loughran & O’Neill, 2016). As compared to methods such as frequency spectrum, 
Harmonic evaluation, precisions and recalls, optimization functions, melodic and harmonic 
error, and fitness functions that seek to evaluate the performance of the algorithm used for 
the composition process, human listening test seeks to assess the aesthetic quality of the 
music. The aesthetic quality of music may be assessed through its pleasantness (Friedmann, 
2015) and pleasantness is experienced when there is satisfaction of objectives. 
Currently, studies have not established the relationship between model performance and 
music quality. This is to say that although a model or an algorithm for composing music may 
be effective and efficient (i.e. can compose optimal music with the shortest possible time), 
the resultant composition may not be the most appealing music for a listener. 
68 
It has been argued that, human evaluation is the method mostly used for assessing music 
quality (Matić, 2010). Although, it does not provide a measure for assessing model 
performance, it provides indications of the model’s ability to produce aesthetic music. 
Accordingly, a human evaluation test was performed to assess the performance of the 
algorithm in terms of its ability to produce aesthetically good music. A survey was 
conducted. A convenience sample was used to select subjects. The participants were to listen 
to all different composed music (all versions composed) and complete a questionnaire. They 
were tasked to rate each of the nine compositions on a scale of 1 to 9 (based on their likeness 
for the music).  
6.2.1 Data Gathering 
Out of the twelve subjects, two were females and ten males. Whose ages range from eighteen 
and above. During the subject selection process, 15 prospective participants were selected. 
The selection process was purely arbitrary based on convenience.  Each participant was 
called (on the phone) and the procedure for the assessment was explained to them. Three of 
the participants rejected to participate in the study, thus 12 were used. All the subjects were 
graduate students: five from the University of Ghana, five from Vellore Institute of 
Technology, India and two from Ghana Technology University College. The resultant music 
after each generation (i.e. compositions 10k v1, 10k v2, 10k v3, 20k v1, 20k v2, 20k v3, 60k 
v1, 60k v2 60k v3) and a link to a Google Forms question in English were sent to each 
participant through WhatsApp in a MIDI file format. Each subject downloaded all the nine 
compositions and they were given a period of 30 minutes to listen and complete the 
questionnaire (each composition was less than a minute long). The data obtained from the 
experiment was collated and analysed. 
69 
Table 6-1 Description of Participants (n=12) 
Category   
Gender  Total 
 Male 10 
 Female 2 
Age   
 18-25 2 
 26 -30 5 
 31-35 3 
 Above 35 2 
Music Player Used   
 Windows Media Player 2 
 QuickTime 1 
 VLC 4 
 Other 5 
Device Used   
 Desktop 2 
 Laptop 4 
 Tablet 1 
 Smartphone 5 
 
70 
6.2.2 Evaluation Results 
To ensure that responses from subjects were unbiased, they were not told that the music were 
composed using a machine learning technique. Specifically, they were asked to listen to 9 
different melodies label (v1to v9) to ranking their preference for each. The number of 
generations, fitness scores for each composition, etc. was not communicated. 
The rankings of subjects were summed to calculate the most appealing music. The findings 
indicated that the most preferred music was the third trail of composition after 60k 
generation. It had a raking score of 81. Refer to Table 6-2 for details of response. It can be 
observed that apart from one participant that scored it “1”, all other participants scored a 5 
or more. However, the respondent who scored “1” for composition 60k (v3), score the same 
value for all compositions. This indicates that he or she does not find any of the composition 
appealing (similar observation can be found in one instance that the subject scored “7” for 
all compositions. The next preferred composition was that of 60k (v1). It scored a value of 
69. The least preferred composition was observed to be 10k (v1). It scored a value of 43. 
In all, it was observed that compositions that were terminated at 60,000 generations were 
considered to be more appealing by the subjects. The average score recorded for all 
compositions of 60, 000 was 71. This is followed by melodies obtained at 20,000 generations 
with an average value of 57. Melodies obtained at 10,000 generations were the least preferred 
since it scored an average of 46 for the 3 versions. Again, the least 3 preferred melodies were 
the 3 version of the 10,000th generation composition. 
 
71 
Table 6-2 Responses of subjects (n=12) 
Question Items Subjects Response ∑ ∑ 
Rank the music [VI] (10k v1) 8 4 7 2 7 2 1 1 3 3 1 4 43 
Rank the music [V2] (10k v2) 7 3 7 3 8 3 1 6 2 2 1 5 48 46 
Rank the music [V3] (10k v3) 8 3 7 1 8 1 1 9 1 1 1 6 47 
Rank the music [V4] (20k v1) 7 6 7 2 8 2 1 2 4 6 1 7 53 
Rank the music [V5] (20k v2) 7 5 7 4 8 4 1 4 5 5 1 5 56 57 
Rank the music [V6] (20k v3) 8 5 7 3 8 3 1 8 6 4 1 7 61 
Rank the music [V7] (60k v1) 9 7 7 4 7 4 1 3 9 9 9   69 
Rank the music [V8] (60k v2) 8 5 7 4 8 3 1 5 8 8 1 5 63 71 
Rank the music [V9] (60k v3) 9 8 7 5 8 6 1 7 7 7 9 7 81 
 
6.3 Implications of Study 
6.3.1 Fitness Functions and Human Evaluation 
The aim of this is to compose a monophonic music using a GA. The algorithm’s performance 
is evaluated by using a fitness function while the aesthetic quality is evaluated using human 
judgement. From the two assessments, it can be concluded that there is no evidence that 
there is a relationship between model performance and the aesthetics quality of the music 
produced. Considering the trails (9) conducted, the first version of the composition obtained 
at 10k generation (i.e.10k v1) converged earlier than all the other models. It attained a total 
fitness score of 10348 at the 750th generation. This indicates that it took a lesser time to 
72 
obtain optimality. Yet, it produced the least preferred music. Melodies composed at the 
longer generations were considered to be more appealing to respondent. 
For instance, the average number of generations for compositions that were terminated at 
60, 000 generations are 5066. This score is the highest average among the termination criteria 
(10k, 20k, 60k). Terminations at 10k produced an average generation of 2684, while that of 
20k produced an average generation of 2500 to reach optimal compositions. Refer to Table 
6-3 for details on the relationship between model performance and human preference. 
Table 6-3 Model Performance vs Human Evaluation (n=12) 
 Version Generation for Optimal Human Evaluation 
Fitness 
Preference Average 
Generations  Generation Average  
Score Preference 
1 750 43 
10,000 2 7300 2684 48 46 
3 5300 47 
1 1800 53 
20,000 2 4000 2500 56 57 
3 1700 61 
1 11000 69 
60,000 2 1800 5066 63 71 
3 2400 81 
 
73 
6.4 Variations in Convergence Timing 
The findings from this study present relevant theoretical implications that need further 
investigations. This is because similar studies conducted by Zad et al., (2006) using twenty 
(20) of Bach’s preludes attained optimality at 500 generation for the first trial and 200 
generation for the second trial. Compared to findings of this study, human evaluation of their 
composed melodies demonstrated there is no difference between the two compositions since 
50% of the evaluators found both melodies amusing. Ohlsson, (2016) also performed a 
similar study and concluded that an optimal music is attainable at 500 generations. 
Considering the findings above, the performance of the model does not determine the 
aesthetic quality of the melodies produced. Also, it can be observed that as compared to the 
other studies (Ohlsson, 2016; Zad et al., 2006), the present study did not attain optimality at 
generations around 500. 
Perhaps, this is a challenge in GAs, because the use of probabilities in the selection, 
crossover and mutation process makes it difficult to predict the time required for a model to 
converge. The initial selection of parents and the mutations plays a key role on how quickly 
the algorithm will converge to the optimal value. 
6.5 Chapter Summery 
This chapter has discussed how the study evaluated all the nine composed music as compared 
to the chapter 5 that discussed the evaluation of the model. This chapter assessed the aesthetic 
quality of the music produced. Convenience sampling technique was used to select fifteen 
subjects to rank how the music is pleasant. It was observed that the most pleasing music 
were the versions obtained for 𝐺123 = 60000. 
74 
Comparing the findings of human assessment to the model performance (i.e. the number of 
generations required to obtain the optimal composition), it was concluded that models that 
converged at later times produced more pleasant music. 
75 
Chapter 7 
STUDY LIMITATION AND CONCLUSION 
7.1 Overview 
This chapter concludes the thesis. It presents summaries of the activities performed in the 
study and details its findings and contributions. Furthermore, it presents a discussion on the 
limitation of the study, particularly, limitations regarding the design and evaluation and 
evaluation processes or methods. Finally, it suggests future research studies. 
7.2 Evaluation of Research Objectives 
As explained earlier, a design research is expected to be relevant (Hevner et al., 2004; 
Kuechler & Vaishnavi, 2008) and thus the need to demonstrate the relevance of this study is 
paramount. As such in this section, the various objectives of the study are assessed. 
Objective one was to review existing machine learning methods for music composition. 
This was addressed in chapters two and three. Chapter two started by presenting the 
background and foundation of music and its composition. In chapter three, a systematic 
review was used to assess the current state of literature in the domain. As compared to 
traditional literature reviews, systematic reviews are less bias since they follow a specified 
protocol throughout the review process (Kitchenham & Charters 2007; Petersen et al. 2015). 
Objective two was to compose a simple monophonic music using a genetic algorithm. The 
decision to use genetic algorithm was inspired by the review process that established that 
76 
genetic algorithms are the most used methods in algorithmic music composition. 
Accordingly, chapter five reported the development of a genetic algorithm. Nine (9) 
melodies were composed. The compositions were three versions each for a GA terminated 
at ten thousand, twenty thousand and sixty thousand generations.  
Objective three was to evaluate the performance of the algorithm using fitness scores. This 
was reported in chapter five. The changes in total fitness scores over the various generations 
were plotted and analysed. The chapter further explained the behaviour and performance of 
the algorithm. It was concluded that the algorithm performed best in the first trail which 
terminated at ten thousand generations. It achieves the optimal composition around 750th 
generation. 
Objective four was to evaluate the aesthetic performance of the algorithm through human 
evaluation. Accordingly, a survey was conducted to assess how the quality of the various 
composed melodies can be pleasing to humans. Subjects were tasked to rate the 9 melodies 
and it was observed that the music obtained from the termination at 60000th generation was 
the most pleasant composition. 
7.3 Limitations and Future Work 
From evaluation of the research objectives, this study has made contributions to the domain, 
yet it has limitations and drawbacks. Based on these limitations future research studies are 
recommended. Below is a discussion on its limitations and future research suggestions. 
7.3.1 Monophonic Composition 
Since the study aimed to understand the basic functionalities and features of algorithm music 
composition and its aesthetic quality, a monophonic music was used. However, the simple 
77 
nature of monophonic melodies makes it a challenge for human assessment. In particular, 
the resultant composition was in the bass-clef and this makes it a challenge for assessment. 
Although monophonic compositions are easy to process and analyse by computing systems 
(Meredith et al., 2002),because it can be represented as sequences (Hanna et al., 2007), 
polyphonic melodies are more pleasant to humans and thus reduce the evaluation stress.  
Accordingly, future research must be conducted using polyphonic compositions to assess 
aesthetic quality of GA composed melodies. 
7.3.2 MIDI Files 
Although, the study confirmed in chapter 3 that the most appropriate file for algorithmic 
composition is MIDI, it presented some challenges during the evaluation stages. This is 
because MIDI files are not sound files, rather, they instruct a device to produce. Considering 
that the melodies were sent to participants for evaluation in MIDI format, the aesthetic 
quality of the music will differ based on the instrument selected for playing the file. The 
same file may be pleasing when played on a device using piano as compared to a guitar. 
Although it would have been interesting to instruct the participants to use a specific 
instrument, it was envisaged that this may bring added stress to them and may also affect 
their responses. 
It is therefore suggested that future studies should evaluate the resultant music using a quasi-
experiment by controlling some of the parameters such as the type of instrument used and 
the duration of listening. Instrument used may also be varied to assess with instrument 
produce the most aesthetic music for a specific composition. 
78 
7.3.3 Sequence of Listening by Subjects 
Considering that the music files were sent to participants without instructions on the 
sequence of listening, it is expected that participants will listen to the versions of music in 
the chronological order that was sent to them. This may cause biasness in the evaluation of 
the melodies. Some researchers have also argued that although human rating is useful in 
assessing pleasant compositions, the process of rating music individually is tedious; this may 
affect the evaluation judge at the later stages of the assessment process. 
Again, the use of laboratory experiment will be useful in addressing these limitations. This 
is because participants could be asked to listen to the composition severally before evaluating 
them. 
7.3.4 Expert Evaluation 
Although, music is enjoyed by all and its aesthetic quality is subject to the listener, the use 
of experts for assessment may provide technical information that may be relevant. Their 
ability to confirm the performance of the algorithm by measuring the pitch, timing, tone, etc. 
shall provide the relevant information for the various compositions. Thus, it is recommended 
that future research must include music experts during the evaluation of the music. 
79 
References 
Ala, & Wadi. (2012). Analysis of Music Note Patterns Via Markov Chains (Issue 2). 
http://collected.jcu.edu/honorspapershttp://collected.jcu.edu/honorspapers/2 
Aldrich, V. C. (1963). Philosophy of Art. https://doi.org/10.2307/2105425 
Alpern, A. (1995). Techniques for Algorithmic Composition of Music. 
https://pdfs.semanticscholar.org/d701/dd6cdc82ed544422c553dab59426f759d558.pdf 
Ames, C. (1989). The Markov Process as a Compositional Model: A Survey and (Vol. 22, 
Issue 2). https://about.jstor.org/terms 
Anderling, V., Andreasson, O., Olsson, C., Pavlov, S., SVENSSON, C., & WIKNER, J. 
(2014). Generation of music through genetic algorithms (Issue June). 
Bajpai, P., & Kumar, M. (2010). Genetic algorithm–an approach to solve global 
optimization problems. Indian Journal of Computer Science and Engineering, 1(3), 
199–206. 
Basili, V. R. (1992). Software modeling and measurement: the Goal/Question/Metric 
paradigm. In Quality (p. 24). https://doi.org/10.1016/S0164-1212(99)00035-7 
Basili, V. R., Caldiera, G., & Rombach, H. D. (1994). The goal question metric approach. 
Encyclopedia of Software Engineering, 2, 528–532. https://doi.org/10.1.1.104.8626 
Beynon, M. (2012). Mathematics and Music-Models and Morals. 
http://empublic.dcs.warwick.ac.uk/projects/kaleidoscopeBeynon2005/posters/erlkonig
Poster.pdf] 
Boyes, H. (2003). Resilience and cyber security of technology in the built environment. 
80 
Institute of Engineering and Technology. 
Brereton, P., Kitchenham, B. A., Budgen, D., Turner, M., & Khalil, M. (2007). Lessons 
from applying the systematic literature review process within the software 
engineering domain. Journal of Systems and Software, 80(4), 571–583. 
https://doi.org/10.1016/j.jss.2006.07.009 
Brooks, A. C. (2000). GENETIC ALGORITHMS AND PUBLIC ECONOMICS. February, 
493–513. 
Çaǧdaş, V., & Stubkjær, E. (2011). Design research for cadastral systems. Computers, 
Environment and Urban Systems, 35(1), 77–87. 
https://doi.org/10.1016/j.compenvurbsys.2010.07.003 
Callum, L., Andrew, C., Andrew, R., Proceedings, E., Space, S., Australasian, T., 
Conference, C. M., Grove, K., Sorensen, A., & Brown, A. R. (2008). Metascore : User 
Interface Design for Generative Film Scoring Australasian CRC for Interaction 
Design. Proceedings of Sound : Space—The Australasian Computer Music 
Conference, 2008, 25–30. 
Cambouropoulos, E. (1994). Cambouropoulos. Markov Chains As an Aid to Computer 
Assisted Composition. Musical Praxis, 1 (1):41–52, 1994. - Google Search. 
https://www.google.com/search?q=Cambouropoulos.+Markov+Chains+As+an+Aid+t
o+Computer+Assisted+Composition.+Musical+Praxis%2C+1+(1)%3A41–
52%2C+1994.&rlz=1C5CHFA_enGH812GH812&oq=Cambouropoulos.+Markov+C
hains+As+an+Aid+to+Computer+Assisted+Composition.+Musica 
Campbell, P. S. (1997). Music, the universal language: Fact or fallacy? International 
81 
Journal of Music Education, 29(1), 32–39. 
https://doi.org/10.1177/025576149702900105 
Chamberlain, A., Bodker, M., Kallionpää, M., Ramchurn, R., De Roure, D., Benford, S., & 
Dix, A. (2018). The Design of Future Music Technologies. Proceedings of the Audio 
Mostly 2018 on Sound in Immersion and Emotion  - AM’18, 1–2. 
https://doi.org/10.1145/3243274.3243314 
Charles, S. (1958). PRESCRIPTIVE AND DESCRIPTIVE MUSIC-WRITING. The 
Musical Quarterly, 44(2), 184–195. 
Chew, S. K. (2011). METHOD OF COMPOSING MUSIC ON A HAND HELD DEVICE. 
2(12). 
Chomsky, N. (1957). Syntactic Structures. 
https://www.google.com/search?q=Chomsky%2C+N.+Syntactic+Structures%2C+Mo
uton%2C+The+Hague+(1957).&rlz=1C5CHFA_enGH812GH812&oq=Chomsky%2
C+N.+Syntactic+Structures%2C+Mouton%2C+The+Hague+(1957).&aqs=chrome..6
9i57&sourceid=chrome&ie=UTF-8 
Chuan, C. H., & Chew, E. (2007). A hybrid system for automatic generation of style-
specific accompaniment. Proceedings of the 4th International Joint Workshop on 
Computational Creativity, May, 57–64. 
Da Mota Silveira Neto, P. A., Carmo MacHado, I. Do, McGregor, J. D., De Almeida, E. 
S., & De Lemos Meira, S. R. (2011). A systematic mapping study of software product 
lines testing. Information and Software Technology, 53(5), 407–423. 
https://doi.org/10.1016/j.infsof.2010.12.003 
82 
Dawid, H., & Kopel, M. (1998). On economic applications of the genetic algorithm: A 
model of the cobweb type. Journal of Evolutionary Economics, 8(3), 297-
315Applying Genetic Algorithms to Optimization. 
https://doi.org/10.1007/s001910050066 
Day, H., & Jarrett, S. (2008). Music Composition for Dummies. 
Delgado, M., Fajardo, W., & Molina-Solana, M. (2009). Inmamusys: Intelligent multiagent 
music system. Expert Systems with Applications, 36(3, Part 1), 4574–4580. 
https://doi.org/https://doi.org/10.1016/j.eswa.2008.05.028 
Dong, G., Wang, X., & Liu, D. (2019). Metaheuristic approaches to solve a complex 
aircraft performance optimization problem. Applied Sciences (Switzerland), 9(15). 
https://doi.org/10.3390/app9152979 
Egermann, H., & McAdams, S. (2013). Empathy and Emotional Contagion as a Link 
Between Recognized and Felt Emotions in Music Listening. Music Perception: An 
Interdisciplinary Journal, 31(2), 139–156. 
https://doi.org/10.1017/CBO9781107415324.004 
Fauvel, J., Flood, R., & Wilson, R. (2006). Music and mathematics: from Pythagoras to 
fractals. 
https://books.google.com/books?hl=en&lr=&id=yLBykq9cjp0C&oi=fnd&pg=PA1&d
q=Music+and+mathematics:+An+overview&ots=OAf6BD-
CnE&sig=Mi5DDHmmbEZXbGYBu-tNixFI82A 
Fernandez, J. D., & Vico, F. (2013). AI Methods in Algorithmic Composition: A 
Comprehensive Survey. Journal of Artificial Intelligence Research, 48, 513–582. 
83 
https://doi.org/10.1613/jair.3908 
Fox, R., & Crawford, R. (2016). A hybrid approach to automated music composition. 
Advances in Intelligent Systems and Computing, 464, 213–223. 
https://doi.org/10.1007/978-3-319-33625-1_20 
Fraser, A. (1957). Simulation of Genetic Systems by Automatic Digital Computers I. 
Introduction. Australian Journal of Biological Sciences, 10(4), 484. 
https://doi.org/10.1071/bi9570484 
Friedmann, J. L. (2015). Music in Our Lives: Why We Listen, How It Works. McFarland & 
Company,Inc. 
https://books.google.fi/books?id=257UBQAAQBAJ&lpg=PP1&hl=fi&pg=PR5#v=on
epage&q&f=false 
Ghaheri, A., Shoar, S., Naderan, M., & Hoseini, S. S. (2015). The applications of genetic 
algorithms in medicine. Oman Medical Journal, 30(6), 406–416. 
https://doi.org/10.5001/omj.2015.82 
Goldberg, D. E., & Holland, J. H. (1988). Genetic Algorithms and Machine Learning. 
Machine Learning, 3(2), 95–99. https://doi.org/10.1023/A:1022602019183 
Gutknecht, M. (1992). The ‘Postmodern Mind’: Hybrid Models of Cognition. Connection 
Science, 4(3–4), 339–364. https://doi.org/10.1080/09540099208946623 
Hanna, P., Ferraro, P., & Robine, M. (2007). On optimizing the editing algorithms for 
evaluating similarity between monophonic musical sequences. Journal of New Music 
Research, 36(4), 267–279. https://doi.org/10.1080/09298210801927861 
Hart, S. L. (1971). Axiology--Theory of Values. Philosophy and Phenomenological 
84 
Research, 32(1), 29–41. 
Hedden, S. K. . (1981). Music Listening Skills and Music Listening Preferences. 16–26. 
https://www.jstor.org/stable/40317638 
Hermitary, & Meng-hu. (2010). HOUSE OF SOLITUDE. 
https://www.hermitary.com/solitude/cage.html 
Hevner, A. R., March, S. T., Park, J., & Ram, S. (2004). DESIGN SCIENCE IN 
INFORMATION SYSTEMS RESEARCH. MIS Quarterly, 28(1), 75–105. 
Hild, H., Feulner, J., & Wolfram Menzel. (1992). HARMONET: A Neural Net for 
Harmonizing Chorales in the Style of l. 
Hoeberechts, M., Demopoulos, R. J., & Katchabaw, M. (2007). A flexible music 
composition engine. Audio Mostly 2007 - 2nd Conference on Interaction with Sound, 
Conference Proceedings, 52–57. 
Hole, J., Hirsch, M., Ball, E., & Meads, C. (2015). Music as an aid for postoperative 
recovery in adults: A systematic review and meta-analysis. The Lancet, 386(10004), 
1659–1671. https://doi.org/10.1016/S0140-6736(15)60169-6 
Hole, K. R., Gulhane, V. S., & Shellokar, N. D. (2013). Application of Genetic Algorithm 
for Image Enhancement and Segmentation . International Journal of Advanced 
Research in Computer Engineering & Technology (IJARCET), 2(4), 1342–1346. 
Holland, J. . (1975). Adaptation in Natural and Artificial Systems. 
https://www.google.com/search?q=Holland+1975&rlz=1C5CHFA_enGH812GH812
&oq=Holland+1975&aqs=chrome..69i57&sourceid=chrome&ie=UTF-8 
Jacob, B. L. (1995). Composing with genetic algorithms. In Proceedings of the 1995 
85 
International Computer Music Conference, September, 452–455. 
Jacob, B. L. (1996). Algorithmic composition as a model of creativity. Organised Sound, 
1(3), 157. https://doi.org/10.1017/S1355771896000222 
James, K. L., Randall, N. P., & Haddaway, N. R. (2016). A methodology for systematic 
mapping in environmental sciences. Environmental Evidence, 5(1), 1–13. 
https://doi.org/10.1186/s13750-016-0059-6 
Järveläinen, H. (2000). Telecommunications software and multimedia laboratory Tik-
111.080 Seminar on content creation Spring 2000 Art@Science Algorithmic Musical 
Composition Algorithmic musical composition. 
Johannesson, P., Perjons, E., & Bider, I. (2013). What Are the Siblings of Design Science. 
The 2nd International SIG Prag Workshop on IT Artefact Design & Workpractice 
Improvement, 1–7. 
Jones, G. (1998). Genetic and Evolutionary Algorithms. 
Jones, K. (1981). Compositional Applications of Stochastic Processes. Computer Music 
Journal, 5(2), 45. https://doi.org/10.2307/3679879 
Jonnalagadda, S. R., Goyal, P., & Huffman, M. D. (2015). Automating data extraction in 
systematic reviews: A systematic review. Systematic Reviews, 4(1). 
https://doi.org/10.1186/s13643-015-0066-7 
Kania, A. (2017). The Philosophy of Music. 
https://plato.stanford.edu/archives/fall2017/entries/music/ 
Karnan, M., & Thangavel, K. (2007). Automatic detection of the breast border and nipple 
position on digital mammograms using genetic algorithm for asymmetry approach to 
86 
detection of microcalcifications. Computer Methods and Programs in Biomedicine, 
87(1), 12–20. https://doi.org/10.1016/j.cmpb.2007.04.007 
Kemper, K. J., & Danhauer, S. C. (2005). Music as therapy. Southern Medical Journal, 
98(3), 282–288. https://doi.org/10.1097/01.SMJ.0000154773.11986.39 
Kennerly, E., Witzel, A., & Zvesper, J. A. (2009). Thief belief. 169–172. 
https://doi.org/10.1109/icegic.2009.5293592 
Kitchenham, B. (2007). Source: "Guidelines for performing Systematic Literature 
Reviews in SE", Kitchenham et al Guidelines for performing Systematic 
Literature Reviews in Software Engineering. https://userpages.uni-
koblenz.de/~laemmel/esecourse/slides/slr.pdf 
Kitchenham, B., & Charters, S. (2007). Guidelines for performing Systematic Literature 
Reviews in Software Engineering. https://doi.org/10.1145/1134285.1134500 
Kuechler, W., & Vaishnavi, V. (2008). The emergence of design research in information 
systems in North America. Journal of Design Research, 7(1), 1–16. 
https://doi.org/10.1504/JDR.2008.019897 
Kunz, T., & Stilman, M. (2013). Time-optimal trajectory generation for path following 
with bounded acceleration and velocity. Robotics: Science and Systems, 8, 209–216. 
https://doi.org/10.7551/mitpress/9816.003.0032 
Leblanc, B., Lutton, E., Braunschweig, B., & Toulhoat, H. (2002). History and immortality 
in evolutionary computation. Lecture Notes in Computer Science (Including Subseries 
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2310, 
128–139. https://doi.org/10.1007/3-540-46033-0_11 
87 
Liu, C.-H., & Ting, C.-K. (2016). Computational Intelligence in Music Composition: A 
Survey. IEEE Transactions on Emerging Topics in Computational Intelligence, 1(1), 
2–15. https://doi.org/10.1109/tetci.2016.2642200 
Livingstone, S. R., Mühlberger, R., Brown, A. R., & Loch, A. (2007). Controlling musical 
emotionality: An affective computational architecture for influencing musical 
emotions. Digital Creativity, 18(1), 43–53. 
https://doi.org/10.1080/14626260701253606 
Lopez-Rincon, O., Starostenko, O., & Ayala-San Mart\’\in, G. (2018). Algoritmic music 
composition based on artificial intelligence: A survey. 2018 International Conference 
on Electronics, Communications and Computers (CONIELECOMP), 187–193. 
Loughran, R., & O’Neill, M. (2016). Generative Music Evaluation: Why do We Limit to 
“Human”? Proceedings of the 1st Conference on Computer Simulation of Musical 
Creativity, Ml, 1–16. 
Lumsden, D., & Ludlow, P. (2007). DIGITAL MUSIC COMPOSITION DEVICE, 
COMPOSITION SOFTWARE AND METHOD OF USE. 1(19). 
Magnusson, T., & Mendieta, E. H. (2007). The acoustic, the digital and the body: A survey 
on musical instruments. Proceedings of the 7th International Conference on New 
Interfaces for Musical Expression, NIME ’07, 94–99. 
https://doi.org/10.1145/1279740.1279757 
Makris, D., Kaliakatsos-Papakostas, M. A., & Cambouropoulos, E. (2015). Probabilistic 
Modular Bass Voice Leading in Melodic Harmonisation. ISMIR, 323–329. 
March, S. T., & Smith, G. F. (1995). Design and natural science research on information 
88 
technology. Decision Support Systems, 15(4), 251–266. 
Marta, A. C. (2003). Parametric Study of a Genetic Algorithm using a Aircraft Design 
Optimization Problem. Genetic Algorithms and Genetic Programming at Stanford 
2003, 133–142. http://www.genetic-programming.org/sp2003/Marta.pdf 
Matić, D. (2010). A genetic algorithm for composing music. Yugoslav Journal of 
Operations Research, 20(1), 157–177. https://doi.org/10.2298/YJOR1001157M 
McAlpine, K., Miranda, E., & Hoggar, S. (1999). Making music with algorithms: A case-
study system. Computer Music Journal, 23(2), 19–30. 
https://doi.org/10.1162/014892699559733 
Meredith, D., Lemström, K., & Wiggins, G. A. (2002). Algorithms for discovering 
repeated patterns in multidimensional representations of polyphonic music. 
International Journal of Phytoremediation, 21(1), 321–345. 
https://doi.org/10.1076/jnmr.31.4.321.14162 
Merriam, A., & Merriam, V. (1964). The anthropology of music. 
Mitchell, M. (1995). Genetic algorithms: An overview. Complexity, 1(1), 31–39. 
https://doi.org/10.1002/cplx.6130010108 
Moorer, J. A. (1972). Music and Computer Composition. Communications of the ACM, 
15(2), 104–113. https://doi.org/10.1145/361254.361265 
Nicoară, E. S. (2015). Applying Genetic Algorithms to Optimization Problems in 
Economics. Economic Insights - Trends & Challenges, 67(3), 125–132. 
http://search.ebscohost.com/login.aspx?direct=true&db=bth&AN=110030133&site=b
si-live 
89 
Nierhaus, G. (2009). Algorithmic composition: paradigms of automated music generation. 
https://books.google.com/books?hl=en&lr=&id=jaowAtnXsDQC&oi=fnd&pg=PA1&
dq=algorithmic+composition+classes&ots=GOJfcLckXn&sig=b-
fdBJdbn_c57VLIOF8ixkQ2sSo 
Offermann, P., Levina, O., Schönherr, M., & Bub, U. (2009). Outline of a design science 
research process. Proceedings of the 4th International Conference on Design Science 
Research in Information Systems and Technology, DESRIST ’09, January. 
https://doi.org/10.1145/1555619.1555629 
Ohlsson, P. (2016). Computer assisted music education. http://files/229/Computer assisted 
music education.pdf 
Papadopoulos, Alexandre, Roy, P., & Pachet, F. (2016). Assisted lead sheet composition 
using flowcomposer. International Conference on Principles and Practice of 
Constraint Programming, 769–785. 
Papadopoulos, Athanase. (2002). Mathematics and music theory: From pythagoras to 
rameau. The Mathematical Intelligencer, 24(1), 65–73. 
https://doi.org/10.1007/BF03025314 
Papadopoulos, G., & Wiggins, G. (1999). AI Methods for Algorithmic Composition : A 
Survey, a Critical View and Future Prospects. AISB Symposium on Musical 
Creativity, November, 110–117. 
http://www.doc.gold.ac.uk/~mas02gw/papers/AISB99b.pdf 
Peffers, K., Rothenberger, M., Tuunanen, T., & Vaezi, R. (2012). Design science research 
evaluation. Lecture Notes in Computer Science (Including Subseries Lecture Notes in 
90 
Artificial Intelligence and Lecture Notes in Bioinformatics), 7286 LNCS, 398–410. 
https://doi.org/10.1007/978-3-642-29863-9_29 
Persson, P., Kammerlind, P., Bergman, B., & Andersson, J. (2000). Methodology for 
multi-characteristic system improvement with active expert involvement. Quality and 
Reliability Engineering International, 16(5), 405–416. https://doi.org/10.1002/1099-
1638(200009/10)16:5<405::AID-QRE349>3.0.CO;2-5 
Petersen, K., Vakkalanka, S., Software, L. K.-I. and, & 2015, U. (2015). Guidelines for 
conducting systematic mapping studies in software engineering: An update. Elsevier. 
Pitkänen, M. (2014). Pythagoras, music, sacred geometry, and genetic code. 
http://tgdtheory.com/public_html/. 
Prechtl, A. (2016). Adaptive Music Generation for Computer Games. 140. 
Prechtl, A., Laney, R., Willis, A., & Samuels, R. (2014). Algorithmic music as intelligent 
game music. AISB 2014 - 50th Annual Convention of the AISB. 
Purao, S. (2002). Design Research in the Techonology of Information Systems: Truth or 
Dare. Unpublished Paper, School of Information Sciences and Technology, The 
Pennsylvania State University, University Park, State College, PA., June. 
Radhakrishnan, S., Kolippakkam, D., & Mathura, V. S. (2007). Introduction to algorithms. 
In Bioinformatics: A Concept-Based Introduction. https://doi.org/10.1007/978-0-387-
84870-9_3 
Rafieyan, R., & Ries, R. (2007). A description of the use of music therapy in consultation-
liaison psychiatry. Psychiatry (Edgmont (Pa. : Township)), 4(1), 47–52. 
Ramanto, A. S., Nur, U., & Maulidevi, S. T. (2017). Markov Chain Based Procedural 
91 
Music Generator with User Chosen Mood Compatibility. 
Rechenberg, I. (1989). Evolution Strategy: Nature’s Way of Optimization (pp. 106–126). 
https://doi.org/10.1007/978-3-642-83814-9_6 
Rice, M., Koh, R., Liu, Q., He, Q., Wan, M., Yeo, V., Ng, J., & Tan, W. P. (2013). 
Comparing Avatar Game Representation Preferences across Three Age Groups. 
Conference on Human Factors in Computing Systems - Proceedings. 
https://doi.org/10.1145/2468356.2468564 
Riechmann, T. (1998). Genetic Algorithms and Economic Evolution. Working Paper, 219, 
1–15. 
Roads, C. (1985). Research in Music and Artificial Intelligence. ACM Computing Surveys 
(CSUR), 17(2), 163–190. https://doi.org/10.1145/4468.4469 
Roads, C, & Wieneke, P. (1979). Grammars as Representations for Music. Computer 
Music Journal, 3(1), 48. https://doi.org/10.2307/3679756 
Roads, Curtis., & Strawn, J. (1996). The computer music tutorial. MIT Press. 
https://books.google.com/books?hl=en&lr=&id=nZ-
TetwzVcIC&oi=fnd&pg=PR9&dq=Roads,+C.+(1996).+The+Computer+Music+Tuto
rial.+Cambridge,+MA:+MIT+Press.+&ots=BdWwfjBp8v&sig=Q9-
umjKZHhEEcqKldLir-H2cDfQ 
Rodrigues, A., Costa, E., Cardoso, A., … P. M.-… I. in M., & 2016, U. (2016). Evolving l-
systems with musical notes. Springer. https://link.springer.com/chapter/10.1007/978-
3-319-31008-4_13 
Sastry, K., Goldberg, D., & Kendall, G. (2005). Genetic algorithms. In Search 
92 
Methodologies: Introductory Tutorials in Optimization and Decision Support 
Techniques (pp. 97–125). Springer US. https://doi.org/10.1007/0-387-28356-0_4 
Scirea, M., Barros, G. A. B., Shaker, N., & Togelius, J. (2015). SMUG: Scientific Music 
Generator. ICCC, 204–211. 
Scruton, R. (1982). Art and Imagination: A Study in the Philosophy of Mind. Routledge & 
Kegan Paul. 
https://www.abebooks.co.uk/servlet/BookDetailsPL?bi=22696505435&cm_sp=plp-_-
9780710090140-_-used 
Sein, M. K., Henfridsson, O., Purao, S., Rossi, M., & Lindgren, R. (2011). Action design 
research. MIS Quarterly: Management Information Systems, 35(1), 37–56. 
https://doi.org/10.2307/23043488 
Shin, A., Crestel, L., Kato, H., Saito, K., Ohnishi, K., Yamaguchi, M., Nakawaki, M., 
Ushiku, Y., & Harada, T. (2017). Melody generation for pop music via word 
representation of musical properties. ArXiv Preprint ArXiv:1710.11549. 
Simoni, M. (2003). Algorithmic Composition: A Gentle Introduction to Music 
Composition Using Common LISP and Common Music. In SPO Scholarly 
Monograph Series. Michigan Publishing, University of Michigan Library. 
https://doi.org/10.3998/spobooks.bbv9810.0001.001 
Stovold, E., Beecher, D., Foxlee, R., & Noel-Storr, A. (2014). Study flow diagrams in 
Cochrane systematic review updates: An adapted PRISMA flow diagram. Systematic 
Reviews, 3(1), 1–5. https://doi.org/10.1186/2046-4053-3-54 
Swyter, W. (2009). Composing Music with L-System Grammars. 
93 
Thorsen, T. (2018). Splinter Cell: Conviction infiltrates April 13. 
https://www.gamespot.com/articles/splinter-cell-conviction-infiltrates-april-13/1100-
6249491/ 
Vaishnavi, V. K., & Kuechler, B. (2004). Design Science Research in Information Systems 
Overview of Design Science Research. Ais, 45. https://doi.org/10.1007/978-1-4419-
5653-8 
Vaishnavi, V., Kuechler, B., & Petter, S. (2012). DESIGN SCIENCE RESEARCH IN 
INFORMATION SYSTEMS. 1, 1–66. https://doi.org/1756-0500-5-79 
[pii]\r10.1186/1756-0500-5-79 
Van Aken, J. E. (2005). Management research as a design science: Articulating the 
research products of mode 2 knowledge production in management. British Journal of 
Management, 16(1), 19–36. https://doi.org/10.1111/j.1467-8551.2005.00437.x 
Vasudha, J., Iyshwarya, G., Selvi, A. T., Iniyaa, S., & Jeyakumar, G. (2011). Application 
of Computer-Aided Music Composition in Music Therapy. IInternational Journal of 
Innovation, Management and Technology (IJIMT), Vol. 2, No. 1, Pp 55-57, February 
2011., 2(1), 2010–2012. 
Venable, J. R., Pries-heje, J., & Baskerville, R. (2012). A Comprehensive Framework for 
Evaluation in Design Science Research. 7286(May). https://doi.org/10.1007/978-3-
642-29863-9 
Volchenkov, D., René, J., & René Dawin, R. (2012). MUSICAL MARKOV CHAINS. 
International Journal of Modern Physics: Conference Series, 13, 116–135. 
https://doi.org/10.1142/S2010194512007829 
94 
Williams, D., Kirke, A., Miranda, E., Daly, I., Hwang, F., Weaver, J., & Nasuto, S. (2017). 
Affective Calibration of Musical Feature Sets in an Emotionally Intelligent Music 
Composition System. ACM Transactions on Applied Perception, 14(3). 
https://doi.org/10.1145/3059005 
Williams, D., Kirke, A., Miranda, E. R., Roesch, E., Daly, I., & Nasuto, S. (2015). 
Investigating affect in algorithmic composition systems. Psychology of Music, 43(6), 
831–854. https://doi.org/10.1177/0305735614543282 
Woody, R. H., & Lehmann, A. C. (2010). Student musicians’ ear-playing ability as a 
function of vernacular music experiences. Journal of Research in Music Education, 
58(2), 101–115. https://doi.org/10.1177/0022429410370785 
Wu, J., Bin, D., Feng, X., Wen, Z., & Zhang, Y. (2018). GA Based Adaptive Singularity-
Robust Path Planning of Space Robot for On-Orbit Detection. Complexity, 2018. 
https://doi.org/10.1155/2018/3702916 
Zad, D. D., Araabi, B. N., & Lucas, C. (2006). A novel approach to automatic music 
composing: Using genetic algorithm. International Computer Music Conference, 
ICMC 2006, January 2006, 551–555. 
Zandkarimi, M., Shafiei, M., Hadizadeh, F., Ali Darbandi, M., & Tabrizian, K. (2014). 
Prediction of pharmacokinetic parameters using a genetic algorithm combined with an 
artificial neural network for a series of alkaloid drugs. Scientia Pharmaceutica, 82(1), 
53–70. https://doi.org/10.3797/scipharm.1306-10 
Zhang, H., Babar, M. A., & Tell, P. (2011). Identifying relevant studies in software 
engineering. Information and Software Technology, 53(6), 625–637. 
95 
https://doi.org/10.1016/j.infsof.2010.12.010 
 
  
96 
Appendix One: List of Primary Studies Used for Review 
 
[S1] D. Wells and H. Elaarag, “A novel approach for automated music composition 
using memetic algorithms,” Proc. Annu. Southeast Conf., pp. 155–159, 2011, doi: 
10.1145/2016039.2016083. 
[S2] R. De Prisco, G. Zaccagnino, and R. Zaccagnino, “EvoBassComposer : a Multi-
objective Genetic Algorithm for 4-voice Compositions,” pp. 817–818, 2010. 
[S3] J. H. Jensen and P. C. Haddow, “Evolutionary Music Composition based on Zipf 
’ s Law Categories and Subject Descriptors,” pp. 41–42, 2011. 
[S4] X. Zheng, D. Li, L. Wang, Y. Zhu, L. Shen, and Y. Gao, “Chinese folk music 
composition based on genetic algorithm,” in 3rd IEEE International Conference 
on, 2017, doi: 10.1109/CIACT.2017.7977295. 
[S5] C. Sulyok, “Genetic operators in evolutionary music composition,” Proc. - 2018 
20th Int. Symp. Symb. Numer. Algorithms Sci. Comput. SYNASC 2018, pp. 253–
259, 2018, doi: 10.1109/SYNASC.2018.00047. 
[S6] R. Fox and R. Crawford, “A hybrid approach to automated music composition,” 
in Advances in Intelligent Systems and Computing, 2016, vol. 464, pp. 213–223, 
doi: 10.1007/978-3-319-33625-1_20. 
[S7] X. Zheng, W. Guo, D. Li, L. Wang, and Y. Wang, “A Hybrid Swarm Composition 
for Chinese Music,” in International Conference on Swarm Intelligence, 2017, 
vol. 1, pp. 258–265, doi: 10.1007/978-3-319-61833-3. 
[S8] R. Loughran, J. McDermott, and M. O’Neill, “Grammatical Music Composition 
with Dissimilarity Driven Hill Climbing,” 2016, pp. 110–125, doi: 10.1007/978-
3-319-31008-4. 
97 
[S9] D. M. Hofmann, “A Genetic Programming Approach to Generating Musical 
Compositions,” in Lecture Notes in Computer Science (including subseries 
Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 
2015, vol. 9027, pp. 89–100, doi: 10.1007/978-3-319-16498-4. 
[S10] M. Takano and Y. Osana, “Automatic Melody Generation Considering User’s 
Evaluation Using Interactive Genetic Algorithm,” 2014, pp. 785–797, doi: 
org/10.1007/978-3-319-13563-2_66. 
[S11] C. De Felice, R. De Prisco, D. Malandrino, G. Zaccagnino, R. Zaccagnino, and R. 
Zizza, “Chorale Music Splicing System: An Algorithmic Music Composer 
Inspired by Molecular Splicing,” vol. 9027, pp. 50–61, 2015, doi: 10.1007/978-3-
319-16498-4. 
[S12] D. Quick and P. Hudak, “Grammar-based automated music composition in 
Haskell,” Proc. ACM SIGPLAN Int. Conf. Funct. Program. ICFP, pp. 59–70, 
2013, doi: 10.1145/2505341.2505345. 
[S13] L. Mazurowski, “Computer models for algorithmic music composition,” 2012 
Fed. Conf. Comput. Sci. Inf. Syst. FedCSIS 2012, pp. 733–737, 2012. 
[S14] A. Liu et al., “Composer4Everyone: Automatic Music Generation with Audio 
Motif,” Proc. - 2nd Int. Conf. Multimed. Inf. Process. Retrieval, MIPR 2019, pp. 
502–503, 2019, doi: 10.1109/MIPR.2019.00101. 
[S15] M. Ebrahimi, B. Majidi, and M. Eshghi, “Procedural Composition of Traditional 
Persian Music Using Deep Neural Networks,” 2019 IEEE 5th Conf. Knowl. Based 
Eng. Innov. KBEI 2019, pp. 521–525, 2019, doi: 10.1109/KBEI.2019.8734959. 
98 
[S16] E. Munoz, J. M. Cadenas, Y. S. Ong, and G. Acampora, “Memetic Music 
Composition,” IEEE Trans. Evol. Comput., vol. 20, no. 1, pp. 1–15, 2016, doi: 
10.1109/TEVC.2014.2366871. 
[S17] H. Chen, Q. Xiao, and X. Yin, “Generating music algorithm with deep 
convolutional generative adversarial networks,” 2019 2nd Int. Conf. Electron. 
Technol. ICET 2019, pp. 576–580, 2019, doi: 10.1109/ELTECH.2019.8839521. 
[S18] D. Samuel and M. Pilat, “Composing Multi-Instrumental Music with Recurrent 
Neural Networks,” Proc. Int. Jt. Conf. Neural Networks, vol. 2019-July, no. July, 
pp. 1–8, 2019, doi: 10.1109/IJCNN.2019.8852430. 
[S19] A. Baratè and L. A. Ludovico, “Algorithmic generation of music tunes for an 
iPhone® game,” Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. 
Intell. Lect. Notes Bioinformatics), vol. 7990 LNCS, no. Lim, pp. 251–265, 2013, 
doi: 10.1007/978-3-642-40050-6_22. 
[S20] H. A. García Salas, A. Gelbukh, and H. Calvo, “Music composition based on 
linguistic approach,” Lect. Notes Comput. Sci. (including Subser. Lect. Notes 
Artif. Intell. Lect. Notes Bioinformatics), vol. 6437 LNAI, no. PART 1, pp. 117–
128, 2010, doi: 10.1007/978-3-642-16761-4_11. 
[S21] R. Abboud and J. Tekli, Integration of nonparametric fuzzy classification with an 
evolutionary-developmental framework to perform music sentiment-based 
analysis and composition, vol. 4. Springer Berlin Heidelberg, 2019. 
[S22] R. T. Dean and J. Forth, “Towards a Deep Improviser: a prototype deep learning 
post-tonal free music generator,” in Neural Computing and Applications, 2018, 
doi: 10.1007/s00521-018-3765-x. 
99 
[S23] C. Huang and C. Chu, “Emotion-Based Rhythmic Complexity Analysis for 
Automated Music Generation,” 2012, pp. 67–78. 
[S24] M. Scirea, J. Togelius, P. Eklund, and S. Risi, “Affective evolutionary music 
composition with MetaCompose,” Genet. Program. Evolvable Mach., vol. 18, no. 
4, pp. 433–465, 2017, doi: 10.1007/s10710-017-9307-y. 
[S25] P. Hutchings and J. Mccormack, “Using Autonomous Agents to Improvise Music 
Compositions in Real-Time,” 2017, pp. 114–127, doi: 10.1007/978-3-319-55750-
2. 
[S26] N. Otani, S. Shirakawa, and M. Numao, “Symbiotic evolution to generate chord 
progression consisting of four parts for a music composition system,” in Lecture 
Notes in Computer Science (including subseries Lecture Notes in Artificial 
Intelligence and Lecture Notes in Bioinformatics), 2014, vol. 8862, pp. 849–855, 
doi: 10.1007/978-3-319-13560-1. 
[S27] M. Modrzejewski, M. Dorobek, and P. Rokita, “Application of Deep Neural 
Networks to Music Composition Based on MIDI Datasets and Graphical 
Representation Mateusz,” 2019, pp. 143–152, doi: 10.1007/978-3-030-20912-4. 
[S28] X. Cao, L. Sun, J. Niu, R. Wu, Y. Liu, and H. Cai, “Automatic composition of 
happy melodies based on relations,” in Multimedia Tools and Applications, 2015, 
vol. 74, no. 21, doi: 10.1007/s11042-014-2057-4. 
[S29] W. Suiter, “The promise of fuzzy logic in generalised music composition,” in IFIP 
Advances in Information and Communication Technology, 2010, vol. 333, pp. 
118–127, doi: 10.1007/978-3-642-15214-6_12. 
[S30] S. Ianigro and O. Bown, Exploring transfer functions in evolved CTRNNs for 
music generation, vol. 11453 LNCS. Springer International Publishing, 2019. 
100 
[S31] W. Wainiya, “Chen ’ s Attractor in Music Composition,” 2017 21st Int. Comput. 
Sci. Eng. Conf., vol. 6, pp. 1–4, 2017. 
[S32] M. S. Sinith and K. V. V. Murthy, “Automatic music Composition based on HMM 
and identified wavelets in musical instruments,” 2011 - Int. Conf. Signal Process. 
Commun. Comput. Netw. Technol. ICSCCN-2011, no. Icsccn, pp. 163–166, 
2011, doi: 10.1109/ICSCCN.2011.6024535. 
[S33] M. Della Ventura, “Automatic Tonal Music Composition Using Functional 
Harmony,” in International Conference on Social Computing, Behavioral-
Cultural Modeling, and Prediction, 2015, pp. 290–295, doi: 10.1007/978-3-319-
16268-32. 
[S34] F. Colombo, A. Seeholzer, and W. Gerstner, “Deep Artificial Composer: A 
Creative Neural Network Model for Automated Melody Generation,” 2017, pp. 
81–96, doi: 10.1007/978-3-319-55750-2. 
[S35] O. Olseng and B. Gambäck, “Co-evolving Melodies and Harmonization in 
Evolutionary Music Composition,” in 7th International Conference, 
EvoMUSART 2018 Proceedings, 2018, pp. 239–255, doi: 10.1007/978-3-319-
77583-8. 
[S36] J. Mańdziuk, A. WoŹniczko, and M. Goss, “A Neuro-memetic System for Music 
Composing,” in IFIP Advances in Information and Communication Technology, 
2014, vol. 436, pp. 130–139, doi: 10.1007/978-3-662-44654-6_13. 
[S37] Y. W. Nam and Y. H. Kim, “Automatic jazz melody composition through a 
learning-based genetic algorithm,” in Lecture Notes in Computer Science 
(including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in 
101 
Bioinformatics), 2019, vol. 11453 LNCS, pp. 217–233, doi: 10.1007/978-3-030-
16667-0_15. 
[S38] J. H. Jeong and C. W. Ahn, “Automatic Evolutionary Music Composition Based 
on Multi-objective Genetic Algorithm,” vol. 2, pp. 479–491, 2015, doi: 
10.1007/978-3-319-13356-0. 
[S39] M. A. Kaliakatsos-Papakostas, A. Floros, and M. N. Vrahatis, “Interactive music 
composition driven by feature evolution,” Springerplus, vol. 5, no. 1, 2016, doi: 
10.1186/s40064-016-2398-8. 
[S40] Y. Nagashima, “Realtime Musical Composition System for Automatic Driving 
Vehicles,” 2018, vol. 11112, pp. 228–234, doi: 10.1007/978-3-319-99426-0. 
[S41] B. Rodríguez, R. de Piñérez, and G. M. S. M, “Using Probabilistic Parsers to 
Support Salsa Music Composition,” 2017, vol. 10527, no. June 2017, pp. 361–
372, doi: 10.1007/978-3-319-71827-9. 
[S42] R. A. Mora-Gutiérrez, J. Ramírez-Rodríguez, and E. A. Rincón-García, “An 
optimization algorithm inspired by musical composition,” Artif. Intell. Rev., vol. 
41, no. 3, pp. 301–315, 2014, doi: 10.1007/s10462-011-9309-8. 
[S43] S. Malgaonkar, Y. B. Nag, R. Devadiga, and T. Hirave, “An AI based intelligent 
music composing algorithm: Concord,” in 2013 International Conference on 
Advances in Technology and Engineering, ICATE 2013, 2013, doi: 
10.1109/ICAdTE.2013.6524723. 
[S44] M. Fridenfalk, “Algorithmic music composition for computer games based on L-
system,” in 2015 IEEE Games Entertainment Media Conference, GEM 2015, 
2016, doi: 10.1109/GEM.2015.7377230. 
102 
[S45] R. Rangarajan, “Generating Music from Natural Language Text,” 2015, pp. 85–
88. 
[S46] Y. Isowa, N. Iino, S. Okino, and Y. Iizuka, “A Musical Composition Assistant 
System Using a Stochastic Musical Model,” in Proceedings - 2017 6th IIAI 
International Congress on Advanced Applied Informatics, IIAI-AAI 2017, 2017, 
doi: 10.1109/IIAI-AAI.2017.24. 
[S47] K. Venugopal and P. Madhusudan, “Feasibility of music composition using 
artificial neural networks,” in Proceedings of the International Conference on 
Computing Methodologies and Communication, ICCMC 2017, 2017, no. Iccmc, 
doi: 10.1109/ICCMC.2017.8282520. 
[S48] Juan Manuel Escalante, “The Sound Digestive System: A Strategy for Music and 
Sound Composition,” 2015, pp. 71–77, doi: 10.1007/978-3-319-16498-4. 
[S49] A. Tulilaulu, J. Paalasmaa, M. Waris, and H. Toivonen, “Sleep musicalization: 
Automatic music composition from sleep measurements,” Lect. Notes Comput. 
Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 
7619 LNCS, pp. 392–403, 2012, doi: 10.1007/978-3-642-34156-4_36. 
 
103