Preliminary Results on Adaptively Trained Neural Networks
Carlos Eduardo Pedreira, N.M. Roehl · 1993
Real world applications often involve time varying models. There is an intrinsic difficulty in dealing with this sort of models, specially when one is concerned with nonlinear syster. Layered neural networks have been successfully used in a variety of relevant problems when invariance assumptions can be properly made. On the other hand, very little can be found in the literature concerning time varying systems. We propose a new procedure to adjust weights in Neural Networks suitable for this type of models with no necessity of retraining. One of the main features of our approach concerns the designer flexibility to control a trade off problem between fitting new incoming data and causing minimum damage to the information related to the original data set. We desire to keep the error related to the late incoming data within a pre established tolerance, while maximizing the information incorporated by original training data. In this way, the designer is able to judge the relevance he or she wants to attribute to the latest data.