Reducing redundancy with unit merging for self-constructive normalized Gaussian networks

Abstract

In this paper, a Normalized Gaussian Network (NGnet) is introduced for online sequential learning that uses unit manipulation mechanisms to build the network model self-constructively. Several unit manipulation mechanisms have been proposed for online learning of an NGnet. However, unit redundancy still exists in the network model. We propose a merge mechanism for such redundant units, and change its overlap calculation in order to improve the identification accuracy of redundant units. The effectiveness of the proposed approach is demonstrated in a function approximation task with balanced and imbalanced data distributions. It succeeded in reducing the model complexity around 11% on average while keeping or even improving learning performance.