Sound Classification in Hearing Aids by the Harmony Search Algorithm - Music-Inspired Harmony Search Algorithm: Theory and Applications - page 181

where T represents a matrix containing the target classes for each input pat-

tern. If the MSE is assumed to be:

C × N

N C

(18)

1

∑∑
==

2

MSE = e

cn

N C

n 11 c

It is possible to derive with respect to the coefficients w , and minimize the MSE , the

ij

result being:

( )

(19)

− 1

T T

V = T ⋅ Q ⋅ Q ⋅ Q

5.4 Parameters for the Harmony Search Algorithm

In the experiments of this chapter, the HS algorithm has been used in order to deter-
mine the set of features that minimizes the validation error rate. For this purpose, sev-
eral feature sets have been explored by using the optimization process described in
Section 3. Table 1 lists the parameters used for the problem at hand. The parameters
have been explored to select those more appropriate features for sound classification
in digital hearing aids.

Table 1. List of the parameters used for the HS algorithm

Parameter Symbol Value

M M

= 20 Number of melodies in memory

P P

= 0 . 1

Probability of generating a novel melody

P P

= 0 1

Probability of mutation

Regarding the fitness objective, the classification error probability over the validation
set has been used. In those cases in which the classification error over the validation set
is exactly the same, the mean square error over the validation set has also been used.
Each experiment has been repeated 30 times, and the mean and the standard devia-
tion of the classification error rates have been calculated and evaluated.

6 Results

This section summarizes the results obtained by the HS algorithm in terms of error
rate. For comparative purposes, the results applying the sequential search (SS) and
random search (RS) algorithms have been also included.
Table 2 shows the results obtained by these feature selection techniques. For the
RS and the HS algorithms, three different batches of experiments have been carried
out, limiting the number of iterations of the algorithm to 100, 1,000, or 10,000 itera-
tions. These results have been expressed as the mean and the standard deviation for
the classification error rates over the validation and the test sets.