Improving the fit of locally weighted regression models

Introduces an algorithm for approximating a function by means of local models. We assume that the data arrives pattern-by-pattern. The shapes and locations of receptive fields are changed in an adaptive manner. With each pattern, not only are the model equations updated but also the regions for which each model contributes significantly to the forecast. This error-dependent step is based on competition: models with worse forecasts in the long term retire in favour of superior models. Areas with higher errors attract models, so there is a tendency to balance local errors. We assume that the distribution of input vectors is known, and therefore we use a fixed number of models, distributed randomly according to the known distribution during the initialisation phase. Although our algorithm was developed in the framework of continuous learning, an even better performance can be achieved by presenting the patterns repeatedly. The learning capabilities are demonstrated by means of an example.