An architecture of interval random vector function-link networks and its numerical analysis

摘要

This paper extends the random vector functional-link (RVFL) networks with single-hidden-layer to interval ones (IRVFLNs) with interval model parameters. The analytic solutions are derived for the interval network parameters using the well-known least square methods, which can overcome the problems such as local minimal, slow convergence. In order to evaluate the performance of IRVFLNs, we choose two data sets in different levels of complexity to be modeled, and compare the aspects of generalization and train time with the interval feed-forward BP neural networks (IBPNNs). The simulation results show that the proposed IRVFLNs have the better properties than the IBPNNs in the network converging and approximating.