Optimum steepest descent higher level learning radial basis function network

Abstract

Dynamically changing real world applications, demands for rapid and accurate machine learning algorithm. In neural network based machine learning algorithms, radial basis function (RBF) network is a simple supervised learning feed forward network. With its simplicity, this network is highly suitable to model and control the nonlinear systems. Existing RBF networks in literature are applied to static applications and also faces challenges such as increased model size, neuron removal, improper center selection etc leading to erroneous output. To overcome the challenges and handle complex real world problems, this paper proposes a new optimum steepest descent based higher level learning radial basis function network (OSDHL-RBFN). The proposed OSDHL-RBFN implements major components inspired from the human brain for efficient learning, adaptive structure and accurate classification. Higher level learning and thinking components of the proposed network are sample deletion, neuron addition, neuron migration, sample navigation and neuroplasticity. These components helps the classifier to think before learning the samples and regulates the learning strategy. The knowledge gained from the trained samples are used by the network to identify the incomplete sample, optimal center and bond strength of hidden & output neurons. Adaptive network structure is employed to minimize classification error. The proposed work also uses optimum steepest descent method for weight parameter update to minimize the sum square error. OSDHL-RBFN is tested and evaluated in both static and dynamic environments on nine benchmark classification (binary and multiclass) problems for balanced, unbalanced, small, large, low dimensional and high dimensional datasets. The overall and class wise efficiency of OSDHL-RBFN is improved when compared to other RBFN’s in the literature. The performance results clearly show that the proposed OSDHL-RBFN reduces the architecture complexity and computation time compared to other RBFN’s. Overall, the proposed OSDHL-RBFN is efficient and suitable for dynamic real world applications in terms of detection time and accuracy. As a case study, OSDHL-RBFN is implemented in real time remote health monitoring application for classifying the various abnormality levels in vital parameters.