Theories And Key Technologies Of Interactive Genetic Algorithms With Individual's Uncertain Fitness

There exist a lot of optimization problems with implicit indices in many fields, such as industrial production, engineering, economy and society. These problems are difficult or even impossible to be expressed with explicitly defined objective functions, hence traditional intelligent optimization methods are not applicable. Interactive genetic algorithms can effectively solve these kinds of problems by embedding user's intelligent evaluation into traditional evolution mechanism. However, user fatigue and evaluation uncertainties greatly influence the performance of interactive genetic algorithms, and restrict their applications to more complicated optimization problems. In order to overcome the above disadvantages, theories and key technologies of interactive genetic algorithms with an individual's uncertain fitness are mainly researched in this dissertation.Firstly, the theory of fitness estimation in interactive genetic algorithms without considering evaluation uncertainties is studied, so as to alleviate user fatigue and establish foundations for the subsequent work. An interactive genetic algorithm with multiple surrogate models based on adaptive space division is proposed. The search space is adaptively separated, and multiple surrogate models are constructed in every subspace and applied to estimate an individual's fitness instead of the user, which effectively alleviate user fatigue. Then a method of evolutionary knowledge extraction and applications based on directed graphs is presented. The fitness estimation of an individual is obtained based on the directed graph. With the estimated fitness, methods of identifying and applying superior individuals are given. The proposed algorithm effectively accelerates convergence and alleviates user fatigue.Secondly, the uncertainties in interactive genetic algorithms are researched, and two kinds of uncertain numbers are adopted to express an individual's fitness. A fuzzy number with Gaussian membership function is adopted to express an individual's fitness in order to characterize the user's fuzzy cognition. Comparison of individuals is performed by using cut sets of fuzzy numbers. This method well reflects the user's evaluation uncertainty. Based on the above study, the stochastic uncertainty in the evaluation process is further considered by using a stochastic variable with normal distribution to express the stochastic uncertainty, and an interactive genetic algorithm with an individual's fuzzy and stochastic fitness is presented. By using cut sets of fuzzy numbers and confidence levels of stochastic variables, a method of individuals' comparison is obtained. The approaches to calculate the stochastic parameters and cut set levels are proposed based on the fuzzy entropy. The above expression method of an individual's fitness simultaneously reflects the user's cognitive uncertainty and stochastic uncertainty, and maintains the diversity of an evolutionary population.Finally, the investigation on building surrogate models to approximate a user's cognition is performed for interactive genetic algorithms whose individual's fitness is expressed with three kinds of uncertain numbers. Neural networks based surrogate models of interactive genetic algorithms with an individual's interval fitness is proposed based on our previous studies. Two RBF neural networks are adopted to approximate an interval's upper limit and lower limit, respectively. When an individual's fitness is a fuzzy number, a support vector classification (SVC) and a support vector regressor (SVR) are both adopted to construct surrogate models. A selection strategy of training data is presented based on the fuzzy entropy of the best individual's fuzzy fitness, and applications of these models are also demonstrated. As for the interactive genetic algorithms with an individual's fuzzy and stochastic fitness, the surrogate models are obtained based on directed fuzzy graphs and support vector machines (SVMs). Directed fuzzy graphs are used to extract evolutionary knowledge, and based on the graphs, an individual's fuzzy and stochastic fitness is converted into a precise one. The SVMs are then trained as surrogate models based on these precise fitness. The above three algorithms are quantitatively analyzed. By applying surrogate models to estimate an individual's fitness instead of the user, user fatigue is effectively alleviated and the performance of interactive genetic algorithms with an individual's uncertain fitness in searching is improved.The above researches are successfully applied in fashion evolutionary design systems, and the applied results show that our work effectively expresses the user's evaluation uncertainties, alleviates user fatigue, and improves the performance of interactive genetic algorithms, which provide guarantees for applying these algorithms in more complicated fields.