Improved Bayesian Random Vector Functional-link Networks For Uncertain Data Modelling

This paper presents an improved algorithm,named IB-RVFL,which is a com-plete Bayesian framework combined with the Random Vector Functional-Link(RVFL)networks for uncertain data modeling.In comparison to the existing work,we ad-dress the use of prior distributions on the parameters of basis functions.This additional information helps to enhance the learning power of the model and pro-vides an effective solution for the difficult and significant setting problem of random parameters in existing RVFL-based modelling techniques.A Variational Inference(VI)method is used to obtain an approximation of the intractable posterior dis-tribution,which helps to realize automatic inference of the hyper-parameters and gives a probability estimate for the test data.We experimented from two aspects.One is to observe the robustness of the algorithm by adding different levels of noise to the training set.The results show that IB-RVFL is less sensitive to noise,and its prediction accuracy is better than the other two algorithms in all experimrnts.The other is to observe the performance of the algorithm by comparing it with the other five algorithms on eight different regression data sets.The results show that the IB-RVFL performs well on all data sets.