Incremental learning and model selection under virtual concept drifting environments

This paper presents an incremental learning and model selection method under the virtual concept drifting environments, where their prior distribution of inputs is changing over time. In the previous work, a statistical model of the virtual concept drift was constructed, and the model-selection criterion for radial basis function neural networks (RBFNNs) under such environments was built with the environmental model (Yamauchi 2009). However, in the previous model, no consideration was given to reducing the computational complexity and storage space for storing learned samples used in future re-learning. This study extends the previous model to a new one that uses less storage space. The extended model uses pseudo-learning samples generated by its RBFNN predecessor instead of using the real old learning samples.