A TLBO based gradient descent learning-functional link higher order ANN: An efficient model for learning from non-linear data

All the higher order ANNs (HONNs) including functional link ANN (FLANN) are sensitive to random initialization of weight and rely on the learning algorithms adopted. Although a selection of efficient learning algorithms for HONNs helps to improve the performance, on the other hand, initialization of...

Full description

Bibliographic Details
Main Authors: Bighnaraj Naik, Janmenjoy Nayak, H.S. Behera
Format: Article
Language:English
Published: Elsevier 2018-01-01
Series:Journal of King Saud University: Computer and Information Sciences
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S1319157816300040
Description
Summary:All the higher order ANNs (HONNs) including functional link ANN (FLANN) are sensitive to random initialization of weight and rely on the learning algorithms adopted. Although a selection of efficient learning algorithms for HONNs helps to improve the performance, on the other hand, initialization of weights with optimized weights rather than random weights also play important roles on its efficiency. In this paper, the problem solving approach of the teaching learning based optimization (TLBO) along with learning ability of the gradient descent learning (GDL) is used to obtain the optimal set of weight of FLANN learning model. TLBO does not require any specific parameters rather it requires only some of the common independent parameters like number of populations, number of iterations and stopping criteria, thereby eliminating the intricacy in selection of algorithmic parameters for adjusting the set of weights of FLANN model. The proposed TLBO-FLANN is implemented in MATLAB and compared with GA-FLANN, PSO-FLANN and HS-FLANN. The TLBO-FLANN is tested on various 5-fold cross validated benchmark data sets from UCI machine learning repository and analyzed under the null-hypothesis by using Friedman test, Holm’s procedure and post hoc ANOVA statistical analysis (Tukey test & Dunnett test).
ISSN:1319-1578