An Experimental Study of Evolutionary Product-Unit Neural Network Algorithm

Abstract

This paper aims to obtain empirical information about the behavior of an Evolutionary Product-Unit Neural Network (EPUNN) in different scenarios. To achieve this, an extensive evaluation was conducted on 21 data sets for the classification task. Then, we evaluated EPUNN on eleven noisy data sets, on sixteen imbalanced data sets, and on ten missing values data sets. As a result of this evaluation process, we conclude that there does not exist a significant difference between EPUNN and the four algorithms assessed; the accuracy of EPUNN rapidly worsen in the presence of noise, so we do not recommend its utilization in noisy environments; we found a tendency to robustness in EPUNN while the imbalance ratio grows; finally, we can state that it is able to handle missing data, but in this kind of data, a significant performance deterioration was manifested. For future work, we recommend to assess the impact of irrelevant attributes on EPUNN performance. In addition, an extension of noisy data set evaluation would be opportune.