This paper presents a new algorithm for cooperative co-evolution of random connections (Co-Co-RC) among neurons in a neural network (NN). Search for an optimal NN is decomposed into search for its two subcomponents, i.e. a connection matrix (C) and a weight matrix (W). Each subsystem is evolved by two separate genetic algorithms and through special evolutionary operations. Optimum connection matrix is obtained by considering a minimum number of neurons in which any neuron may have a connection with any neuron of any layer. The Mackey-Glass chaotic time series is used as a benchmark to test the performance of the proposed evolutionary strategy. Statistical ttest analysis reveals the significant improvement of the method as compared with previous co-co strategies as well as standard back-propagation

Keywords