An RBF online learning scheme for non-stationary environments based on fuzzy means and Givens rotations

Abstract

Learning on non-stationary environments is laden with many challenges, as the procedure is usually characterized by drifts and data unavailability; on the other hand, it is of great importance with regard to data stream modeling, where an effective and up-to date model is required as the data stream evolves. This work presents a new method for producing highly accurate models for learning on such environments. The method is based on artificial neural networks (ANNs) and more specifically on the efficient architecture of radial basis function (RBF) networks. A novel RBF online training scheme for real time adaption of both the network structure and parameter values is proposed based on the fuzzy means (FM) algorithm and the Givens rotations technique. Within this integrated framework, it is guaranteed that for each update of the network structure, the optimal values for the synaptic weights are calculated efficiently by maintaining low order matrix updates. The proposed approach is evaluated on 9 real-word and artificial benchmark data streams, including a challenging real-world application which involves precipitation prediction, and is compared to other well-known methodologies from the literature. The results show that the FM-Givens algorithm produces models with highly competitive online accuracy for non-stationary environments in the presence of drifts, while maintaining low model-updating times.