In this paper, we introduce a self-constructive Normalized Gaussian Network (NGnet) for online learning tasks. In online tasks, data samples are received sequentially, and domain knowledge is often limited. Then, we need to employ learning methods to the NGnet that possess robust performance and dynamically select an accurate model size. We revise a previously proposed localized forgetting approach for the NGnet and adapt some unit manipulation mechanisms to it for dynamic model selection. The mechanisms are improved for more robustness in negative interference prone environments, and a new merge manipulation is considered to deal with model redundancies. The effectiveness of the proposed method is compared with the previous localized forgetting approach and an established learning method for the NGnet. Several experiments are conducted for a function approximation and chaotic time series forecasting task. The proposed approach possesses robust and favorable performance in different learning situations over all testbeds.

The approximation expression about error accumulation of a long-term prediction is derived. By analyzing this formula, we find that the factors that can affect the long-term predictability include the model parameters, prediction errors and the derivates of the used basis functions. To enlarge the maximum attempting time, we present that more suitable basis functions should be those with smaller derivative functions and a fast attenuation where out of the time series range. We compare the long-term predictability of a non-polynomial based algorithm and a polynomial one to prove the success of our method.