A Novel Hybrid Chemical Reaction Optimization Algorithm with Adaptive Differential Evolution Mutation Strategies for Higher Order Neural Network Training

In this paper, an application of a hybrid Chemical Reaction Optimization (CRO) algorithm with adaptive Differential Evolution (DE) mutation strategies for training Higher Order Neural Networks (HONNs), especially the Pi-Sigma Network (PSN) is presented. Contrasting to traditional CRO algorithms, the reactant size (population size) remains fixed throughout all iterations, which makes it easier to implement. In addition, four DE mutation strategies (DE/rand/1, DE/best/1, DE/rand/2 and DE/best/2) with adaptive selection of control parameters as inter-molecular reactions and one intra-molecular reaction have been used. The proposed algorithm combines the diversification property of inter-molecular reactions following DE/rand mutation strategies and intensification property of intra-molecular reaction as well as inter-molecular reactions following DE/best mutation strategies, thereby glorifying the chances of reaching the global optima in less iteration. The performance of the proposed algorithm for HONN training is evaluated through a well-known neural network training benchmark i.e., to classify the parity-p problems. The results obtained from the proposed algorithm to train HONN have been compared with results from the following algorithms: Basic CRO algorithm, CRO-HONN Training (HONNT) and the most popular variants of DE algorithm (DE/rand/1/bin, DE/best/1/bin). It is observed that the application of the proposed hybridized algorithm to (DE-CRO-HONNT) performs statistically better than that of other algorithms considering both classification accuracy and number of generation taken to attain the solutions.