This article presents a music generation system based on a modified interactive genetic algorithm (IGA) with dynamic user engagement tracking. The developed approach allows the creation of monophonic MIDI compositions based on user feedback. The proposed model automatically adjusts the probabilities of mutations and injections in the population depending on user engagement level, there-by reducing user fatigue and accelerating convergence toward acceptable results. Conducted experiments with five volunteers showed that the system is able to generate compositions rated by users as attractive in an average of 4.6 iterations, while the generation time of a new generation is less than 1 ms. The system has a simple interface based on .NET MAUI and allows exporting results in MIDI format. The proposed solution combines high speed with the possibility of fine-tuning the generation parameters, which makes it promising for creating interactive music products in real time.
Dynamic music generation has been known and used in various digital products since the end of the
are widely used in the field of music generation, where the subjective attractiveness of the same track
depends on the aesthetic preferences of a particular user or group of users [
Several existing systems demonstrate the potential of IGAs for music generation. For example,
GenJam [
This paper examines an improved adaptive interactive genetic algorithm designed to generate music in the MIDI format. The developed model is characterized by simplicity and high speed but differs from a number of other models. In particular, the developed system monitors user feedback in different generations and evaluates user engagement based on statistical indicators calculated from the flow of user ratings. The calculated indicator dynamically adjusts the probability of mutations and injections in the population in order to reduce user fatigue, accelerate convergence, and maintain interest throughout the entire process of using the system.
The article [
Learning mode. Random compositions are generated; the user evaluates them. The evolutionary process has not started.
Demo mode. The best of the previously generated compositions is played.
Evolution mode. Genetic operations are used for the dynamic generation of a new population of musical compositions.
GenJam uses a two-level genetic coding scheme to represent compositions. Musical content is hierarchically structured into phrases and measures, each encoded as binary chromosomes of fixed length. This presentation combines rhythm and pitch. Each bar corresponds to a 32-bit chromosome representing eight consecutive eighth-note positions in a 4/4-time signature. Each position is encoded by a 4-bit event: Rest, Hold, New Note.
The system is quite limited: the compositions do not differ in speed (BPM value), they use predetermined sequences of chords.
The GP-Music system described in the article [
Another approach is described in the article [
Musical Digital Interface (MIDI) is a standard developed in 1983 as a result of a collaboration between leading manufacturers of electronic instruments. The standard provides interoperability between equipment from different manufacturers and pro-vides a digital abstraction layer over analog and digital sound generation processes.
The Standard MIDI File (SMF) format allows compact storage of data about musical compositions. Unlike classic audio files that store the data about sound directly, MIDI files store sequences of discrete musical events that are then converted into music using software or hardware solutions. Each MIDI file consists of a header and one or more tracks. The header defines global parameters such as file type, number of tracks, and tempo. Each track consists of a sequence of time-ordered events, each of which has a timestamp relative to the previous event. This allows for precise placement of notes and other musical events throughout the track.
The most common are the Note On and Note Off event types, which together determine when a
note starts and ends. Each note playback event is characterized by a MIDI channel (0 15), note pitch
(0 127), note velocity (0 127). These parameters describe both the horizontal (time-based) and
vertical (pitch-based) structure of a musical composition [
The simplicity and widespread nature of the MIDI format justify its choice as the basis for the system under development. Considering the need to create simple monophonic compositions, as well as the absence of the need to change the velocity of notes, it is possible to formally describe a musical composition as follows: where P is the set of allowed pitch values; D is the set of allowed values of the length of the note.
Then, formally, the task of developing a music generation system is reduced to finding such a function M: = (
, { 1, 2, . . . , }), = ( , ) or = (∅, ),
∈ , ∈ , ( +1) = ( , ), = { 1, 2, . . . , }, = { 1, 2, . . . , }, ∈ , (1) (2) (3) (4) (5) (6) (7) where T is the set of all possible musical compositions, considering the limitations of the system.
The most difficult challenge during system development is finding such a function M, which is able to quickly generate compositions that will receive a high rating from the user.
A system was developed based on a modified interactive genetic algorithm capable of generating and playing small musical compositions using the MIDI technology described above. After the generation of the next generation Sn the system offers the user to rate each of the compositions on a 10-point scale. Obtained values Yn are used as a fitness function of compositions. 4.1. The system encodes each composition using a genome consisting of discrete, quantitative, and sequential genes. Discrete genes encode the scale and root note. Quantitative genes encode the number of beats per minute. Sequential genes encode harmonic and rhythmic sequences and arpeggio types in bars. Such data is enough to encode compositions that are not limited to one genre of music, as in the GenJam system. The system is flexible and customizable, because it allows the configuration of limit values for quantitative genes and possible values for discrete genes. 4.2.
Workflow
Generate the initial population S1 random musical compositions (individuals). Value x 2. Show the user the current population. Get the rating value yi for each individual si from the Generate a new population of x individuals. Each individual in the population Sn will be the offspring of two individuals from the previous population S(n-1) with probability 13. 4. 5.
user.
6. Return to step 2.
( ) = −1 ,
∑ =0
During reproduction, the offspring receives a random combination of the genes of its parents. Each of the genes is independently generated on the basis of parental genes. Genes encoding quantitative parameters are chosen randomly between parental values. Genes encoding discrete parameters randomly take one of the parental values. Genes encoding sequence parameters are generated based on the random combination of the sequences of the parameters of the parents.
where P(i) the probability of choosing the i-th individual as a parent, and yi is the i-th rating
received from the user, yi ∈ [
A feature of the developed modified system is a change in the probability of injections and mutations based on the parameter user interest. This technique is important for a system performing a task such as music generation, where formal evaluation of the quality of the generated population is not possible. Using this parameter allows dynamic change of the rate of mutations and injections in the population of music compositions, decreasing the diversity proportionally to user interest. = + ( min − = = 0.5 ∗
, ( 1, 2, … ), where ri the i-th rating received from the user, in the range 0 of the values r.
, + ( min = − the minimum injection probabilities specified as system (12) (13) (14) (15) (16) where max value of gene k for generated individual. , ( 1) and ( , 2) values of gene k for the first and second parent individuals respectively, U(x,y) is a random variable with a uniform distribution on the interval [x, y]. Lchild is the length of the generated sequenced gene, L1, L2 ⬚ lengths of this gene for the first and second parent, respectively. , is the i-th element of the k-th sequential gene.
This crossover implementation correctly handles different genes according to their nature. For example, the BPM (speed of composition) value for the generated individual will lie between the values of this gene in the parent individuals. And the chord sequence will be a combination of the chords used in the parent individuals.
A simple and concise user interface based on the popular cross-platform .NET MAUI library was
developed for the system. This allows using the developed system on Windows and MacOS platforms
[
The system allows users to listen and evaluate each of the compositions generated in the current generation many times in an arbitrary order. Each of the compositions can be exported as a file in form of a graphic indicator at the bottom of the window. Additional functionality is also provided in the form of statistics ex-port containing user ratings for each of the compositions in each generation. ulated value of the user's interest in the
volunteer.
The developed system was tested using the help of five volunteers. The volunteers were instructed to use the system until one of the generated short compositions was subjectively pleasing to the which the user rated with the minimum possible rating. The composition sounds like a random set of sounds and is musically unappealing. Fig. 2 shows the composition that was generated by the system at generation 5 and evaluated for the maximum number of points. The first composition is characterized by a sharp transition to the use of short sixteenth notes and pauses, multiple repetition of individual notes in the middle of the composition, lack of harmonic movement, but the second composition does not have such problems. In the second composition, there is a harmonious movement that alternates ups and downs several times, there is no sharp and unnatural use of short notes and pauses. There is no repetition of the same notes several times in a row, instead simple but effective techniques of creating tension for the listener and relieving it are used. In general, the second composition is perceived as more aesthetically pleasing due to greater musicality and structural coherence.
The results of the interaction of 5 volunteers with the system are shown in the graph in Figure 3. The graph shows the value of the maximum rating for compositions from each generation. The test results indicate the following properties of the system. Its disadvantages are the unpredictability of the results of the system. But the system also demonstrated the ability to generate an acceptable result quite quickly: the highest speed of achieving an acceptable result demonstrated by 5 volunteers is 9 iterations of the system. The average value is 4.6 iterations.
Testing of the developed software system was carried out on a computer with the Windows 11 operating system equipped with an Intel I7-12650H central processor. Table 1 and Figure 4 show the results of testing over 10 generations. Based on the results of the tests, it was established that the average time for the computing of the next generation on this hardware is less than 1 millisecond, generating the next generation is imperceptible to real users of the system.
Table 1 System performance test results
Generation Time spent, milliseconds 1 2,3868 2 0,3068 3 0,1194 4 0,119 5 0,0923 6 0,0747 7 0,0646 8 0,1081 9 0,0931 10 0,1471
While the developed system, like most models based on genetic encoding, is still restricted to generating only monophonic compositions, this limitation should be viewed in the broader context of its performance characteristics. For many practical applications: background music for games, simple generative soundscapes, educational tools, or interactive art installations, monophonic sound can be sufficient. More importantly, the system demonstrates several notable advantages that distinguish it from analogues.
First, its dynamic tracking of user engagement directly addresses one of the key drawbacks of interactive genetic algorithms: user fatigue. By continuously adapting mutation and injection probabilities, the system reduces repetitive or unproductive iterations, enabling users to reach satisfying results in fewer cycles. This approach can be seen as a lightweight alternative to surrogate fitness functions, which are computationally demanding and often difficult to generalize across users.
Second, the computational efficiency of the model makes it practical for real-time applications. Tests confirm that new generations can be produced in less than one millisecond, even on low-end hardware. Thus, the system can be used with software where performance is critical without requiring cloud-based processing or expensive hardware. For example, video games or live interactive performances.
Third, the system allows users to configure stylistic parameters such as tempo, scale, and chord sets, thereby ensuring that the generated music aligns more closely with the intended genre or style. This adaptability makes the system more adaptive than solutions like GenJam or DarwinTunes, which are narrowly specialized in genre or structural constraints.
In summary, while the lack of polyphonic capability represents a structural limitation of the reducing user fatigue make it a promising and practical tool. These strengths collectively suggest that the proposed solution may serve not only as a research prototype but also as a basis for a realworld interactive music generation system.
Methods and existing systems of music generation based on genetic methods were reviewed. In particular, systems that use interactive genetic algorithms and also use surrogate fitness functions in the process of evolution are considered.
A system is proposed that generates music using a modified interactive genetic algorithm and
also uses dynamic user engagement tracking. The proposed system also differs from analogues in
the ability to generate music in different styles, as well as more subtle settings available to users.
In the future, it is possible to improve the system by integrating a surrogate fitness function similar
to the one used in the GP-Music system [
The article was prepared as part of the research work Information Technologies of Intelligent Computing .