Interactive evolutionary computation in modelling user preferences

Abstract

The modelling of user preferences in many applications is very interesting and is one of the problems researched during the last year. We researched the possibilities of neural networks to predict user subjective preferences using human-machine cooperative systems that use Interactive Evolutionary Computation (IEC). In such systems a subjective preference (evaluation) is a response to a system generated proposals. We consider these preferences to present the relative discrete fitness function values. We showed that searching for a preferred solution can be accelerated and evaluation characteristics can be obtained quicker if the target fitness values are converted from relative values to absolute values. We described a formula for a conversion of relative fitness function values to absolute values in IEC algorithms. We used a recurrent neural network to predict user preferences on a problem of the most attractive font face. Our experiments showed a substantial improvement of the error of the neural network in testing phase when using absolute fitness function values.