Evolutionary Dendritic Neuron Model Learned by A State-of-the-art Evolutionary Learning Algorithm

Recently, the role of dendrite structures in nerve calculation has attracted wide attention. As a single neuron model, Dendritic Neuron Model (DNM) is usually built to transmit information through imitating the mechanism and process network of biological nerves. The branches of a dendrite that distribute corresponding to the three coordinates are used to classify the training data based on demands. Contrarily, traditional artificial neural networks which use a couple of McCulloch and Pitts' neurons are still difficult to be understood and trained. Commonly, evolutionary computing is adopted to solve nonlinear problems. In this paper, a recently proposed spherical search algorithm (SASS) is for the first time introduced as the training algorithm for DNM. It substitutes the traditional error back propagation (BP) learning method to alleviate the local minima trapping problem. Six benchmark classification datasets are tested to verify the accuracy of the well trained neural network. Experimental results suggest that SASS performs better as a learning algorithm for DNM in terms of solution accuracy.