Study On The Extension Of Particle Swarm Evolutionary Valuation

Since it was proposed, Particle Swarm Optimization(PSO) have been received widespread attentions both in theory and application with its strong advantage of easy concepts, few control parameters, good convergence ability and simple implementation. However, as a swarm intelligent optimization algorithm, the requirement on a large number of fitness evaluations poses an obstacle for the PSO to be application to solve complex optimization problems with computationally expensive objective function, which limits the application range of the algorithm. Estimation strategy is the main method for this problem, it reduce the number of fitness evaluation by instead the fitness evaluation with objective function with fitness estimation, thereby the total cost of the algorithm is reduced.Fitness estimation strategy is a new estimation strategy proposed in recent years, in which the estimate formula is deduced based on the evolutionary equation of PSO. Different from the other estimation strategy, the fitness estimation strategy does not need set up surrogate models, and the known information can be adequately used. However, in the fitness estimation strategy the selection of the estimated particle is only according to the relative position of the particles in current generation, which limit the number of estimation and made inaccurate fitness estimation. In order to enhance the estimation accuracy while reducing the real fitness evaluations as much as possible, an evolution control method is introduced into the fitness estimation strategy. In this thesis, a similarity and reliability-based evolution control method is introduced into the fitness estimation strategy to judge which particle to estimate, which ensuring the effectiveness of the strategy.Multi-objective optimization is an important area for PSO, but up to now, almost of the study in Multi-objective Particle Swarm Optimization(MOPSO) is focusing on the improvement of the algorithm performance. And majority of the multi-objective optimization problems in the study is very simple. When it comes to complex problems with computationally expensive objective functions, the large number of the fitness evaluation will be the obstacle as it is in single objective algorithm, or even more worse with the more number of the objective function. Therefor the fitness estimation strategy is extended to MOPSO in this thesis, and to improve the accuracy of the estimation, and the evaluation control method is introduced into the strategy.The proposed single-objective optimization algorithm is tested on 6 standard test functions, and comparison results show that the introduction of similarity and reliability improve the performance of the algorithm while reducing the number of fitness evaluation as much as possible. Besides, the multi-objective fitness estimation strategy is introduced into two MOPSO, and two sets of comparison experiment show that the multi-objective strategy is effectively to reduce the number of the fitness evaluations, and the evaluation control method can obviously improve the performance of the algorithm.