Information Technology Reference
In-Depth Information
7 Conclusions
This chapter has centered on the novel application of the harmony search algorithm
on selecting features for sound classification in hearing aids.
After motivating the social demand for hearing aids exhibiting sound classification,
it has been shown how the problem of feature selection in hearing aids is very compli-
cated in the sense that hearing aids suffer from strong constrains that make difficult the
design of the application. It has been explained how these limitations arise mainly from
the fact that hearing aids must work at very low clock frequency in order to minimize
power consumption and maximize battery life. This necessitates to the use of design
algorithms that require a small number of operations per second. It has been illustrated
that, as a consequence, implementing sound classification algorithms embedded in
hearing aids is a very challenging task. The underlying reason is that it requires opti-
mizing each parameter for the sound classifier in order to reduce its computational
complexity while maintaining a low error probability, and thus, a good dynamic adap-
tation to the sound surroundings.
In particular, the feature extraction process, essential to properly characterize the
sound to be classified, is arguably one of the most time-consuming tasks. Therefore,
reducing the number of features has been shown to be one of the most successful
ways of diminishing the complexity of the classification system. However, an exces-
sively small number of features may likely increase the error probability. Thus, select-
ing those more appropriate features is a crucial point for the application at hand.
The feasibility of the harmony search algorithm for solving this complicated prob-
lem has been explored. Harmony search algorithm is an optimization method that
mimics the music composition rules in order to search for the best solution to a
mathematical problem by minimizing some kind of 'fitness function'. Just in this re-
spect, the first step has been establishing a fitness function based on the classification
error. For this purpose, a mean square error linear classifier has been designed for
each combination of features, and the classification error rate over the validation set
has been used as fitness criterion. So, the harmony search algorithm aims to reduce
the number of combinations of features to be evaluated for obtaining the subset of
features that minimize this fitness criterion.
The main results obtained by the harmony search algorithm demonstrate the good
performance of this feature selection technique. For comparative purposes, those re-
sults obtained applying a sequential forward search and a random search process are
also included. Considering the same number of iterations of the algorithms, the pro-
posed feature selection technique outperforms these more classical approaches both
terms of fitness performance and test error rate.
The harmony search algorithm can reduce the number of feature combinations to
be evaluated in the feature selection process for a sound classification system. For the
same number of iterations, the final test error rate achieved by the proposed search
process is 13% lower than the best test error rate achieved by a random search proc-
ess. These results make the harmony search algorithm a viable and promising choice
in the feature selection problem for sound classification in digital hearing aids.
Search WWH ::
Custom Search