Emulating fuzzy mappings with a neural network architecture

We propose a neural network model for the processing of fuzzy data. The network parameters (weights) are standard real numbers and the spreads at the output level result exclusively from uncertainty in the input data. Our network model performs 'intelligent' inference calculations on the basis of fuzzy data and minimizes uncertainty in the final output. The number of free parameters (weights) in our network model coincides with the number of connections emanating from the various nodes. In this paper, we refer only to the single node case. As usual, the learning mechanism corresponds to a nonlinear regression over the network parameters. The local transfer functions associated with the nodes are more sophisticated than the standard ones. The difference is mainly in the preliminary integration module, in which the various local inputs contribute to the single total input to the node. Instead of the standard linear combination with the parameters /spl omega//sub i/, our model weights each local input in a way which is inversely proportional to its spread (uncertainty). As a result, the more precise data are dominant in the local network computation.