Bio-Inspired Hybridization of Artificial Neural Networks for Various Classification Tasks

Abstract

: Recently, in order to optimize artificial neural networks (ANNs), several bio-inspired metaheuristic algorithms have been successfully applied. Moreover, these hybrid ANNs were operated using no more than two or three metaheuristic algorithms at a time. Additionally, the classification field is so rich that some issues were not sufficiently addressed. The main contribution of this paper is related to the use of several ANN hybridizations at the same time, while taking into account the datasets for which the ANNs or their hybridizations have been rarely explored. Thus, seven hybridized ANNs with bio-inspired metaheuristic algorithms such as particle swarm optimization (PSO-ANN), genetic algorithm (GA-ANN), differential evolution (DE-ANN), cultural algorithm (CA-ANN), harmony search (HS-ANN), black hole algorithm (BH- ANN) and ant lion optimizer (ALO-ANN) were considered for classifying four kinds of datasets. After a back-propagation neural network (BPNN) was designed, the connection weights and biases of neurons were optimized by using the seven metaheuristic algorithms mentioned above. The four selected data types belong to different domains and differ with regard to the number of classes, variables and examples. As performance measurement is concerned; the efficiencies, purities and F-measure are analysed. For all simulation runs, it can be noticed that metaheuristic algorithms were able to reach optimal efficiencies and that all the PSO-ANN-based networks obtained higher values for efficiency. For this analysis, the dependence of the obtained results on certain metaheuristic parameters was taken into account.