Modified orthogonal chaotic imperialist competitive algorithm and its application to improve pattern recognition in the multilayer perceptron neural network

In spite of the success of the imperialist competitive algorithm (ICA) in solving optimization problems, this algorithm still suffers from frequent entrapment in the local minimum and low rate convergence. In this paper, a new version of this algorithm, called the modified chaotic orthogonal imperialist competitive algorithm (COICA), is proposed. In the absorption policy of our proposed version, each colony searches the space of movement toward to imperialist through the definition of a new orthogonal vector. The probability of choosing powerful empires is also defined through the Boltzmann distribution function and selection is done through the roulette wheel method. The proposed algorithm is used to training of multilayer perceptron neural network (MLP) to classify standard data sets, including ionosphere and sonar. The K-Fold cross validation method was used to performance evaluation of this algorithm and generalizability assessment of the trained neural network with the proposed version. The results obtained from the simulations confirm reduction of neural network training error and generalizability improvement of our proposed algorithm.