Multi-objective evolutionary algorithms of spiking neural networks

Abstract

Spiking neural network (SNN) is considered as the third generation of artificial neural networks. Although there are many models of SNN, Evolving Spiking Neural Network (ESNN) is widely used in many recent research works. Among the many important issues that need to be explored in ESNN are determining the optimal pre-synaptic neurons and parameters values for a given data set. Moreover, previous studies have not investigated the performance of the multi-objective approach with ESNN. In this study, the aim is to find the optimal pre-synaptic neurons and parameter values for ESNN simultaneously by proposing several integrations between ESNN and differential evolution (DE). The proposed algorithms applied to address these problems include DE with evolving spiking neural network (DE-ESNN) and DE for parameter tuning with evolving spiking neural network (DEPT-ESNN). This study also utilized the approach of multi-objective (MOO) with ESNN for better learning structure and classification accuracy. Harmony Search (HS) and memetic approach was used to improve the performance of MOO with ESNN. Consequently, Multi- Objective Differential Evolution with Evolving Spiking Neural Network (MODE-ESNN), Harmony Search Multi-Objective Differential Evolution with Evolving Spiking Neural Network (HSMODE-ESNN) and Memetic Harmony Search Multi-Objective Differential Evolution with Evolving Spiking Neural Network (MEHSMODE-ESNN) were applied to improve ESNN structure and accuracy rates. The hybrid methods were tested by using seven benchmark data sets from the machine learning repository. The performance was evaluated using different criteria such as accuracy (ACC), geometric mean (GM), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), negative predictive value (NPV) and average site performance (ASP) using k-fold cross validation. Evaluation analysis shows that the proposed methods demonstrated better classification performance as compared to the standard ESNN especially in the case of imbalanced data sets. The findings revealed that the MEHSMODE-ESNN method statistically outperformed all the other methods using the different data sets and evaluation criteria. It is concluded that multi objective proposed methods have been evinced as the best proposed methods for most of the data sets used in this study. The findings have proven that the proposed algorithms attained the optimal presynaptic neurons and parameters values and MOO approach was applicable for the ESNN.