Incremental adaptation of resource-allocating network for non-stationary time series

A major restriction on a traditional artificial neural network (ANN) is the approximation capability will be frozen after the completion of the training process. This results in a gradual degradation of estimation performance when applied to a non-stationary environment. Accordingly, the paper suggests an incremental refinement approach (IRA), which enables a resource-allocating network (RAN) to learn online from environmental variances during the prediction period. During IRA development, the key challenge is the requirement to maintain a compromise between robustness toward interference and the adaptivity to environmental changes. This problem is known as the stability-plasticity dilemma. RAN-IRA is basically composed of three ingredients to achieve the incremental learning process. They are principal kernel selection, noise filtering and incremental refinement. An experiment is provided to evaluate the performance of RAN-IRA by predicting the closing price of Hang Seng Index time series. Finally, empirical results are briefly discussed and provide evidence to indicate that RAN-IRA considerably outperforms the traditional RAN model with a more promising estimation under a non-stationary environment