Novel hybrid compact genetic algorithm for simultaneous structure and parameter learning of neural networks

The automatic simultaneous selection of structure and parameters of an artificial neural networks is an important area of research. Although many variants of evolutionary algorithms (EA) have been successfully applied to this problem, their demanding memory requirements have restricted their application to real world problems, especially embedded applications with memory constraints. In this paper, structure and parameter learning of a neural network using a novel hybrid compact genetic algorithm (HCGA) is proposed. In the HCGA, each string combines real and binary segments together. For a feedforward neural network, the real segment encodes it weights, while the binary segment encodes the presence/absence of a connection of the network. The proposed hybrid compact genetic algorithm (HCGA) has several advantages: low computational cost, controllable weight regularization leading to automatic architecture discovery. The HCGA is tested on two benchmark problems of Ripley´s synthetic 2-class problem and Mackey glass time series prediction problem. Experimental results show that the proposed algorithm exhibits good performance with low computation cost and controllable network structure.