Learning ambiguous functions by neural networks
This addresses reliability assessment in predictive models for domains like physics and finance, but it is incremental as it builds on existing neural network methods.
The paper tackles the problem of model ambiguity in systems with hidden variables by proposing a scheme with two coupled neural networks: one for prediction and another for evaluating output reliability based on error values, applied to tracking slopes in a straw chamber and credit scoring.
It is not, in general, possible to have access to all variables that determine the behavior of a system. Having identified a number of variables whose values can be accessed, there may still be hidden variables which influence the dynamics of the system. The result is model ambiguity in the sense that, for the same (or very similar) input values, different objective outputs should have been obtained. In addition, the degree of ambiguity may vary widely across the whole range of input values. Thus, to evaluate the accuracy of a model it is of utmost importance to create a method to obtain the degree of reliability of each output result. In this paper we present such a scheme composed of two coupled artificial neural networks: the first one being responsible for outputting the predicted value, whereas the other evaluates the reliability of the output, which is learned from the error values of the first one. As an illustration, the scheme is applied to a model for tracking slopes in a straw chamber and to a credit scoring model.