Surrogate enhanced interactive genetic algorithm with weighted Gaussian process

Abstract

Interactive genetic algorithm (IGA), combining a user's intelligent evaluation with the traditional operators of genetic algorithms, are developed to optimize those problems with aesthetic indicators. The evaluation uncertainties and burden, however, greatly restrict the applications of IGA in complicated situations. Surrogate model approximating to the evaluation of the user has been generally applied to alleviate the evaluation burden of the user. The evaluation uncertainties, however, are not taken into account in existing research, therefore, a weighted multi-output gaussian process is here proposed to build the surrogate model by incorporating the uncertainty so as to enhance the performance of IGA. First, an IGA with interval fitness evaluation is adopted to depict the evaluation uncertainty, and the evaluation noise is defined based on the assignment. With the evaluation noise, the weight of each training sample is calculated and used to train a gaussian process which has two outputs to approximate the upper and lower values of the interval fitness, respectively. The trained gaussian process is treated as a fitness function and used to estimate the fitness of individuals generated in the subsequent evolutions. The proposed algorithm is applied to a benchmark function and a real-world fashion design to experimentally demonstrate its strength in searching.