Kernel Method Based Human Model for Enhancing Interactive Evolutionary Optimization

Abstract EN

A fitness landscape presents the relationshipbetween individual and its reproductive success in evolutionarycomputation (EC).

However, discrete and approximatelandscape in an original search space maynot support enough and accurate information for ECsearch, especially in interactive EC (IEC).

The fitnesslandscape of human subjective evaluation in IEC is verydifficult and impossible to model, even with a hypothesisof what its definition might be.

In this paper, wepropose a method to establish a human model in projectedhigh dimensional search space by kernel classificationfor enhancing IEC search.

Because bivalent logicis a simplest perceptual paradigm, the human modelis established by considering this paradigm principle.

In feature space, we design a linear classifier as a humanmodel to obtain user preference knowledge, whichcannot be supported linearly in original discrete searchspace.

The human model is established by this methodfor predicting potential perceptual knowledge of human.

With the human model, we design an evolutioncontrol method to enhance IEC search.

From experimentalevaluation results with a pseudo-IEC user, our proposed model and method can enhance IEC searchsignificantly.