ISBN 88-7395-155-4 © 2006 ICMPC 1848 Alma Mater Studiorum University of Bologna, August 22-26 2006 Improving algorithmic music composition with machine learning Michael Chan School of Computer Science & Engineering University of New South Wales Sydney, Australia mchan@cse.unsw.edu.au John Potter School of Computer Science & Engineering University of New South Wales Sydney, Australia Emery Schubert School of Music & Music Education University of New South Wales Sydney, Australia ABSTRACT Algorithmic generation of musical sounding music is an interesting but challenging task, because machines do not inherently possess any form of creativity, which is neces- sary to create music. The Automated Composer of Style Sensitive Music II (ACSSM II) system generates music by searching for a sequence of music segments that best sat- isfy various constraints, including length and pitch range, harmonic backbone, and consistency with a probabilistic model of a composer’s style. As with all optimization prob- lems, our problem requires the construction of a search space; we utilize a clustering space produced by grouping together music segments having similar musical features. The output sequence is simply a path passing through these clusters. In order to produce such a sequence, we utilize a genetic algorithm. To evaluate the system, we have con- ducted an experiment, involving five subjects who possess at least three years of musical training, in which the over- all musical quality of produced music was assessed. Our results show that the automatically generated music achieved a mean satisfaction score of 7.5/10, which is sig- nificantly higher than that given to the music produced by the earlier ACSSM system. Hence, the results suggest that ACSSM II is a better system than its predecessor and is capable of generating reasonably musical sounding music. Keywords Computer music, algorithmic music composition, artificial intelligence, machine learning, clustering, genetic algo- rithm, Markov chain, generative theory, musical intelli- gence. INTRODUCTION One of the main goals of artificial intelligence (AI) is to develop systems that imitate human intelligence and behav- ior, and its success inspires researchers to focus on making machines also imitate human creativity. One approach is to develop a system that generates quality music, with little or no supervision from the user. If such an automated system could create musical sounding music, it might then be use- ful in various ways; for example, human composers could seek inspiration for their own compositions. More gener- ally, from a computer science perspective, such a system would represent a breakthrough in the application of AI technology. Our system, called the Automated Composer of Style- Sensitive Music II (ACSSM II), aims to automate music composition by exploiting an input corpus of music and incorporating machine-learning techniques, including ge- netic algorithms (Goldberg, 1989) and Markov chains (Brémaud, 1999). It is based on an earlier system ACSSM (Chan and Potter, 2005) which it extends with more exten- sive AI techniques. Both of these systems have a common fundamental concept, which is to generate new-sounding, stylish music by “intelligently” rearranging segments of music found in the input corpus. One interesting aspect of In: M. Baroni, A. R. Addessi, R. Caterina, M. Costa (2006) Proceedings of the 9th International Conference on Music Perception & Cognition (ICMPC9), Bologna/Italy, August 22-26 2006.©2006 The Society for Music Perception & Cognition (SMPC) and European Society for the Cognitive Sciences of Music (ESCOM). Copyright of the content of an individual paper is held by the primary (first-named) author of that pa- per. All rights reserved. No paper from this proceedings may be repro- duced or transmitted in any form or by any means, electronic or me- chanical, including photocopying, recording, or by any information retrieval systems, without permission in writing from the paper's primary author. No other part of this proceedings may be reproduced or transmit- ted in any form or by any means, electronic or mechanical, including photocopying, recording, or by any information retrieval system, without permission in writing from SMPC and ESCOM.