Abstract:

We consider the problem of predicting a risk index of an event on the base of several predictors. The specificity of the problem lies in the fact, that there are no training data; instead we have a set of logical rules describing in a fuzzy way several classes of the risk (’very low’, ’low’, ’medium’, ’high’ and ’very high’ risk). We present several problems and doubts we had met when trying to construct a predictor. After considering several alternatives we stated that a simple neural network architecture (with one perceptron) calculating logistic regression has given the best agreement with expert opinion. Thus,
based on the rules, we are able to mimic the expert opinion by a neural network model.