MetaCompose:

A Compositional Evolutionary Music Composer

Marco Scirea, Julian Togelius, Peter Eklund and Sebastian Risi

Vision

By creating a cognitive model of the player I hope to be able to identify the player's emotional state and be able to reinforce or manipulate it through the use of generative music to improve the player’s experience.

Motivation

In games (and other interactive media) we need dynamic music

The techniques currently used are not very adaptive and are expensive

There’s a lot of research on procedural music generation but not a lot of focus on the expression of emotional significance

MetaCompose

Small music compositions connected to particular scene or level

Possibility of variations for added interest

Real-Time change in the music triggered by game events

Demo Time




High-Level Architecture

Composition Generation

Melody Generation

Genotype

[8,7,11,10]

Constrained to key

Melody Generation

FI-2POP + NSGA

Melody Generation

FI-2POP + NSGA

Constraints:

  • No big leaps (more than a 5th)
  • No repetitions
  • Not all 2nd intervals

Objectives:

  • Counter stepwise approach to leaps
  • Leap notes part of chord
  • First note of bar part of chord

Research Question


Do all parts of Metacompose add to the perceived quality of the music produced?


Evaluation of composition generation

Quantitative study, participants asked to pick a preference between two clips on pleasantness, randomness, harmoniousness and interestingness

1,291 answers from 298 participants.

Composition Generation

Music clip generation

Five groups:

  • MetaCompose
  • Random chord sequence
  • Random unconstrained melody
  • Random constrained (to the key of the piece) melody
  • Random accompaniment

10 music pieces created for each group. msci.itu.dk/evaluationClips/

MetaCompose vs all other groups

Complete generator vs all other groups

  • Participants prefer the complete system in three of the four criteria
  • Interestingness not conclusive, yet we notice a high amount of neutral answers are given (circa 26%)

MetaCompose vs random chords

  • Similar results as before, but less significant
  • Our generator mitigates the disruptive element introduced
  • Uncommon chord sequences increase interestingness

MetaCompose vs random constrained melody

  • No statistical significance for pleasantness and interestingness
  • Participants can tell the lack of structure, as shown by the randomness and harmoniousness criteria

Conclusion

Each part of our music generation method assists creating music that the listener finds more pleasant and structured

Presented a novel GA method for constrained multi-objective optimization

Future work

Conduct a Turing test using human-composed and generator-composed pieces

Conduct new study to adjust our theory to new expressive capabilities of the generator

Investigate how to improve player experience through interaction between music and gameplay

Thank you :)


You can contact me through my webpage: msci.itu.dk or on twitter @MarcoScirea