Extending evolutionary approximation of polyphonic music via multi-objective optimisation

Schörken, Malte; Dettmer, Justin; Vatolkin, Igor; Rudolph, Günter; Hoos, Holger Hendrik

doi:10.18154/RWTH-2026-08900

Extending evolutionary approximation of polyphonic music via multi-objective optimisation

Schörken, Malte^RWTH*

2026

VerantwortlichkeitsangabeMalte Schörken

ImpressumAachen : RWTH Aachen University 2026

Umfang1 Online-Ressource : Illustrationen

Masterarbeit, RWTH Aachen University, 2026

Veröffentlicht auf dem Publikationsserver der RWTH Aachen University

Genehmigende Fakultät
Fak09

Hauptberichter/Gutachter
Hoos, Holger Hendrik (Thesis advisor)^RWTH* ; Rudolph, Günter (Thesis advisor) ; Vatolkin, Igor (Consultant)^RWTH* ; Dettmer, Justin (Consultant)^RWTH*

Tag der mündlichen Prüfung/Habilitation
2026-05-12

Online
DOI: 10.18154/RWTH-2026-08900
URL: https://publications.rwth-aachen.de/record/1041789/files/1041789.pdf

Einrichtungen

Lehrstuhl für Methodik der Künstlichen Intelligenz (Informatik 14) (125710)

Thematische Einordnung (Klassifikation)
DDC: 004

Kurzfassung
Instrument and pitch detection in polyphonic audio recordings is a fundamental yet challenging problem in the field of musical information retrieval. State-of-the-art deep learning approaches require large annotated datasets and extensive training, whereas evolutionary algorithms (EAs) offer a data-independent alternative through direct signal approximation. Previous studies demonstrated that EAs are capable of approximating target recordings to detect pitches and instruments simultaneously. This work aims to improve the detection accuracy by introducing a multi-objective evolutionary algorithm (MOEA) that jointly optimises complementary fitness functions. A fitness function evaluates the similarity between the target signal and its synthesised approximation by combining an audio feature with a corresponding distance metric. Candidate fitness measures were first evaluated in the single-objective setting and groups of complementary objectives were identified using Shapley value analysis. The MOEA was evaluated on single chords and synthetic polyphonic pieces. It was compared to single-objective baselines using Jaccard distance, success rate and perceptual metrics. The resulting populations can further support applications ranging from music transcription and rearrangement to content-based music search and genre classification.

OpenAccess:
PDF
(additional files)