A Review of Artificial Music
Analysis and Composition
Jacob Adams
CS438 Spring 2008
Topic Paper
Research in the field of artificial intelligence has primarily dealt with performing tasks
that are of a very mathematical or logical nature. Obviously, modeling environments that adhere
to a very logical set of rules is much easier than modeling one that does not. However, research
in AI has begun branching out in to more aesthetic and creative fields. Artificial agents are now
performing tasks which were often thought to only be able to be performed by humans. Two
such tasks are music analysis and composition. Manaris et al., have successfully implemented a
system consisting of “artificial music critics” and “artificial music composers” that are able to
determine a piece of music’s “pleasantness” and compose novel pieces that are deemed
“pleasant” as well [1]. Although there are still shortcomings presented in their research, is it a
pivotal step in the field of artificial intelligence and music.
The first step taken in the work of Manaris et al. was to develop the artificial art critics.
This was done by employing an artificial neural network to map pieces of music to their
popularity. Unlike previous research which had taken music notes from the piece as direct inputs
into the neural network, Manaris et al. used algorithms to extract certain musical “features” from
the piece, which where than used as inputs[1]. These algorithms were variations of logarithmic
functions known as power laws which, based on previous research, provide a decent model for
music [1]. The algorithms were used to extract primary features relating to musical properties
such as pitch, duration, harmony intervals, melodic intervals, and chords. Derivative and
standard deviations based on primary features were also calculated. All of these calculations
were then fed into the neural network.
In order to provide test data for the artificial music critic, download statistics were
tracked for an online repository of classical music. The most downloaded songs were considered
to be the most aesthetically pleasing. The most popular and unpopular pieces were then analyzed
and fed into the neural network, which was adjusted based on the pieces actual “aesthetics”. The
resulting network was over 87% accurate when applied to the remainder of the songs [1].
Upon creating reasonably accurate artificial critics, Manaris et al. were able to create an
artificial music composer. The composer agent would be given three inputs, a general template
for a music piece, small musical elements called “musical genes” and a music critic agent. The
“musical genes” could consist of anything from a few notes to an entire musical piece, each of
which could be broken up into smaller intervals of notes or phrases. Using the music template,
1
the composer would create several random arrangements of the musical genes. These
arrangements would then be evaluated by the music critic. The results were used as the fitness
function for genetic programming principles, where the higher rated pieces were crossed over to
form a new hybrid piece. After the crossover and random mutations, the resulting generation of
pieces was evaluated by the critic. This process was continued until a “pleasantness” threshold
was achieved [1].
The test performed by Manaris et al. took the entirety of J.S. Bach’s Invention #13 in A
minor as a template for its artificial composer. The composer then generated 17 novel variations,
2 of which were intentionally created to be “unpleasant”. The pieces were then rated by the 23
students. The 15 “pleasant" works were all assigned high aesthetic values, while the two
“unpleasant” pieces were assigned very low aesthetic values.
These developments and finding certainly represent a large step for the field of artificial
intelligence. It is a success for intelligent agents in a field often considered to be purely creative,
and requiring human intelligence. However, the research of Manaris et al. does have some
shortcomings which leave things to be desired of future research.
The first point worth noting is that the research assumed a direct correlation between the
popularity of a piece of music and its aesthetics. Although it is often the true that aesthetic music
is well liked and gains popularity, it is not always the case. Additionally, much of what is
deemed popular is often considered to be quite unaesthetic. Human data on the aesthetic quality
of music would have been better captured by polling actual music critics on several musical
works. Although it would be more difficult to acquire, this data would probably yield an
artificial critic that more closely judges aesthetics and not popularity.
A second shortcoming of the research is that the composer agent required snippets and a
template with which to work with. A composer that is given large enough building blocks of
pleasing music and a rigid enough template is almost guaranteed to create a fairly pleasing work.
The example even went as far as to use an entire work of music as its template. All variations
that closely resembled the original Bach piece should have been deemed at least somewhat
pleasing based on the virtue that the original piece was pleasing. The current findings
theoretically would allow humans to develop simple music elements, and use the artificial
composer to create entire works form them. However, future research should strive to create an
agent that can create purely original works with no guidelines.
In addition to extending the creative capabilities of the composer, the research also
should have subjected the new pieces to a larger number of human critics. While the download
statistics of thousands of users were tracked when creating the critic, only 23 students were
polled on the works created by the composer. Not only was this a very small sampling, it was a
very narrow one, with the subject pool only having a four year span in age, with 73% of subjects
being female. In order to truly get a grasp of the effectiveness of the composer, a much larger
and broader sampling of human critics should have been used.
2
Finally, a problem arises from the music collection used to train the artificial critic. As a
result of only training on published music, the critic is never exposed to truly “bad” music.
Manaris et al. aknowledge this limitation and mention it as a potential shortcoming [1].
However, one would think that it would be more problematic that what they led on. A critic that
has little or no exposure to “bad” music could potentially behave erratically upon seeing a very
bad piece because it would be very alien to its classifications. Such a critic could give a
composer a favorable review of very bad music, a lot of which would be generated by the
composer in the very early stages of its evolutionary process. It probably would have been
beneficial to the training of the critic if it could have been presented with a broader array of
music, especially it the aspect of aesthetic quality.
Despite its shortcomings, the research of Manaris et al. is still quite remarkable. By
creating a fairly accurate artificial music critic and composer, it has provided a profound success
for artificial intelligence in the field of music. This will hopefully provide a key step toward
further advancements in the analysis and development of creative mediums by artificial agents.
References
[1] Manaris, B., Roos, P.,Machado, P., Krehbiel, D., Pellicoro, L., Romero, J. 2007. A
Corpus-Based Hybrid Approach to Music Analysis and Composition
[2] Johanson, B, Riccardo, P. 1998. GP-Music: An Interactive Genetic Programming System
for Music Genration with Automated Fitness Raters
3