Sound Classification in Hearing Aids by the Harmony Search Algorithm - Music-Inspired Harmony Search Algorithm: Theory and Applications - page 178

inspire novel compositions. Thus, in the effort of exploring utterly different melodies,
the musician is able to generate 'creative' compositions.
The HS algorithm is an optimization method that mimics the composition rules in
order to search for the best solution to a mathematical problem by minimizing some
kind of 'fitness function'. In this music-inspired framework, each solution is consid-
ered a 'melody', and its fitness represents its musical quality. The musician creates
new combinations of sounds, and memorizes the best M melodies, i.e. those with the
highest quality. When the musician has to improvise a novel composition, he/she ei-
ther creates an entirely novel melody with a given probability, let's say, for instance,
P , or selects one of the previous memorized melodies with probability

1 .
Please note that, in the effort of producing new variations, any piece of music is in
fact randomly modified in some of its parts. The iterative application of this process
allows searching for those compositions that minimize the fitness function, or in other
words, those that 'sound better' [16-21].
The question now is: What is the analogy to the problem of selecting those more ap-
propriate sound-describing features for the application at hand? In the music-inspired
framework previously mentioned, the key point consists in defining the basic structure
of a melody. In the problem of selecting features for sound classification in hearing aids,
the set of features to be selected must be coded in an appropriate way. In this work they
have been encoded by using a bitstream so that each bit determines if a given feature is
used or not: bit '1' means that the considered feature will be selected, while bit '0'
represents the contrary. As a consequence, and if there is N F available features, each
melody/composition will consist of a set of N F binary values. With this in mind, the way
the HS algorithm works for the problem at hand can be described as follows:

−

P

Step 1. Preprocessing stage: The harmony memory is filled with random composi-
tions and the associated fitness values are evaluated.

Step 2. A new melody is created. Any melody can be generated either at random
(this case is selected with a probability P ), or by copying it from one memorized
melody (this alternative case is selected with a probability equal to 1 − P n ).

Step 3. The components of the novel melody are then modified with a probability

of

P .

Step 4. The new melody is then evaluated and scored, taking into account the fit-
ness function defined in order to carry out the selection process. In this chapter, the
validation error rate has been selected as fitness function.

Step 5. If the novel melody scores better than the worse one included in the memory,
then it is stored, replacing the worse one. This process iterates beginning in Step 2.

Please note that using the HS optimization algorithm may potentially provide a
good combination of features. The great advantage is that it prevents the process from
evaluating all the possible 2 N F combinations of N F features, which, in some cases,
could demand the use of excessive computational resources.