Article: Polyphonic music melody generation model based on inverse reinforcement learning algorithm Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.35 pp.37 - 56 Abstract: Automated polyphonic music generation remains a challenging task due to the difficulty in designing reward functions and modelling long-range dependencies. To address this, this paper proposes a polyphonic music generation model based on generative adversarial imitation learning. Our model employs a Gated Transformer-XL as its core to effectively capture intricate contrapuntal relationships. Experimental results demonstrate that the model achieves superior performance across multiple metrics: it reduces the Earth mover's distance to 0.23; increases voice separation mutual information to 0.45, and achieves a 91.2% harmonic rule compliance rate. In subjective evaluations, the model attained average opinion scores of 8.7 for melodic fluency and 8.9 for harmonic richness, significantly outperforming all baseline models. These results validate the effectiveness of our approach in generating high-quality polyphonic music with both technical proficiency and artistic merit. Inderscience Publishers - linking academia, business and industry through research

Title: Polyphonic music melody generation model based on inverse reinforcement learning algorithm

Authors: Haoxun Yan; Xiaojing Feng; Zheng Liu

Addresses: School of Arts and Communication Department, Beijing Normal University, Beijing 100875, China ' School of Arts and Communication Department, Beijing Normal University, Beijing 100875, China ' Department of Computer Science and Technology, Taiyuan University, Shanxi 030032, China

Abstract: Automated polyphonic music generation remains a challenging task due to the difficulty in designing reward functions and modelling long-range dependencies. To address this, this paper proposes a polyphonic music generation model based on generative adversarial imitation learning. Our model employs a Gated Transformer-XL as its core to effectively capture intricate contrapuntal relationships. Experimental results demonstrate that the model achieves superior performance across multiple metrics: it reduces the Earth mover's distance to 0.23; increases voice separation mutual information to 0.45, and achieves a 91.2% harmonic rule compliance rate. In subjective evaluations, the model attained average opinion scores of 8.7 for melodic fluency and 8.9 for harmonic richness, significantly outperforming all baseline models. These results validate the effectiveness of our approach in generating high-quality polyphonic music with both technical proficiency and artistic merit.

Keywords: generative adversarial imitation learning; polyphonic music generation; music AI; inverse reinforcement learning.

DOI: 10.1504/IJICT.2026.153005

International Journal of Information and Communication Technology, 2026 Vol.27 No.35, pp.37 - 56

Received: 18 Nov 2025
Accepted: 19 Dec 2025
Published online: 17 Apr 2026 *

Title: Polyphonic music melody generation model based on inverse reinforcement learning algorithm

Keep up-to-date