Neural network models for anytime use

Nowadays, the role of anytime and situational models and algorithms has become important because they offer a way to handle atypical situations and to overcome problems of resource, time, and data insuffiency in changing and time-critical systems and situations. Soft computing, in particular fuzzy and neural network based models are serious candidates for usage in such systems, however their high complexity, and in some cases unknown accuracy, can limit their applicability. In this paper, special neural network structures are introduced which (1) complexity can adaptively be chosen according to the temporal situation (resource, time, and data availability), (2) the accuracy is always known, and (3) monotonously decreases parallel with the increase of the complexity of the used model/algorithm.