Handling uncertainty in neural networks: an interval approach

During the last decade there has been a considerable increase of activities in the field of neural network modeling. Trying to override the limitations of the classical deterministic schema, various network models for dealing with uncertainty are continuously being developed. Most of them are based on stochastic and fuzzy data representations. On the other hand, in the sphere of reliable computations, data analysis and knowledge-based systems, there is an increasing popularity of interval analysis as a method for data processing and reasoning based on data with bounded errors. This paper is devoted to the the interval representation of incomplete and network models. The notion of an interval cell is defined as a building block of the model. The cell properties, including activation function and dynamic properties are examined. Attention is given to the problems connected with the explosion of interval uncertainty that results from repeated operations on interval values. The inference method is based on the propagation of the intervals across the network activating functions. At the end of the paper are given initial guidelines for evaluation of the learning properties