New algorithm for improving prediction performance in modified radial basis function network

In neural networks, the accuracies of the networks are primarily relying on two critical factors, which are the centers and networks weight values. The feed-forward network known as Radial basis function network (RBFN) capable of performing nonlinear approximation on an unknown dataset, classification, pattern recognition, control system, and image processing. However, there are some disadvantages of the RBFN network, such as longer computation time for large datasets, less efficient weight updating, and center selection algorithms that cause low accuracy are identified. Limited data points or overload data points can affect the training of RBFN. Hence, proper size for dataset is required to ensure RBFN is trained using suitable dataset size to lessen the computational time without a significant influence on the accuracy. For RBFN weight updating, the gradient descent (GD) algorithm easily trapped in local minima by random weight generated during the initial stage of training. Meanwhile, the center's selection using the K-means algorithm is known for its sensitivity and high dependency to initial center selection from the input dataset. Therefore, this work proposed solutions for these mentioned disadvantages through modification on a few parts of the RBFN algorithm to improve their performance. First, this work proposed a new dataset reduction formula to obtain a suitable number of a dataset for network training. Next, a modified steepest descent algorithm was proposed for RBFN weight updating during training. Then, a new distance-weighted K-means algorithm is proposed for obtaining more accurate initial centers for RBFN. Finally, this work proposed a new model through a combination of quantum evolutionary algorithm (QEA) and RBFN known as QRBFN. This proposed RBFN demonstrated its abilities in global search and local optimization to effectively provide better accuracy in prediction results. All proposed modified RBFN was tested against the standard RBFN in predictions accuracy on four nonlinear models from literature, and four real-world datasets that consist two time-series datasets (Air pollutant dataset and forex pair EURUSD dataset), and other two datasets are Biochemical Oxygen Demand (BOD) dataset, and Phytoplankton growth dataset. The proposed dataset reduction formula was conducted through experiments where data was tested by a 5 percent step size reduction. The results of this proposed RBFN are compared for root mean square error (RMSE) and area under curve (AUC) values with standard RBFN. The proposed dataset reduction case yielded average results over a 50 percent decrease in time usage and a 20 percent reduction in RMSE. Meanwhile, all proposed RBFN yielded better results and robustness with an average improvement percentage of more than 40 percent in RMSE and AUC results.