Evolving Feedforward Neural Networks Using a Quasi-Opposition-Based Differential Evolution for Data Classification
Seyed Jalaleddin Mousavirad, Shahryar Rahnamayan
Abstract
One of the most challenging problems in conducting machine learning is the learning process of feedforward neural networks (FFNN), which means finding the proper weights for connections and biases. The performance of FFNNs is mainly dependent on the success of the learning process. Gradient descent-based methods such as back-propagation (BP) are among the most widely employed learning algorithms, whereas they are susceptible to be trapped in local optima. Population-based metaheuristic algorithms such as differential evolution (DE) are a reliable alternative to tackle complex problems. In this paper, we propose a quasi-opposition-based differential evolution approach for FFNN learning to improve the performance of FFNNs (QODE-FFNN). Our proposed algorithm benefits from a variant of opposition-based learning (OBL) to enhance the performance of FFNN. Based on OBL concept, the opposite of a candidate solution is generated. Afterward, OBL selects the best between a candidate solution and its opposite based on their objective function values. In this paper, we employed a variant of OBL to improve the performance of FFNN, called quasi-OBL, which generates a random point between the center of search space and its opposite. Also, in our proposed algorithm, connection weights and biases are encoded as a candidate solution, while the objective function is based on a classification error. Experimental results confirm the performance of QODE-FFNN compared to other recent approaches.