Fitness Calculation Time Consuming Optimization Problem Oriented Pso Prediction Strategy Research

Particle Swarm Optimization (PSO) has been widely applied in variouskinds of optimization problems for its simplicity in concept， fast convergenceand strong global optimization capability. However， as a population-basedoptimization approach， high computation cost significantly limits itsapplicability for the solution of real-world engineering optimization problemswith computationally expensive objective functions. To alleviate this problem，computationally efficient models can be constructed to approximate the fitnessfunction. Such models are often known as approximate models， meta-modelsor surrogates.Model management is an important problem needed to be solved whenusing a surrogate in a population-based optimization approach. Which modelwill be chosen to estimate the fitness， how and when to estimate fitness is alsoa problem needed to be solved when using a surrogate. No matter which modelto choose， the selection of the sample directly affects the correctness of theestimated fitness when using a surrogate. Generalized Regression NeuralNetwork (GRNN) is a transformation of the Radial Basis Function network. Animportant advantage of the GRNN is its simplicity and fast approximationprocedure. Comparison with back-propagation-based neural networks， GRNNdoes not converge to local minima and the training process of GRNN algorithmis much more efficient than with BP-NN algorithm. Hence， in this paper，generalized regression neural network is adopted to be used as the predictionmodel， and different selection strategy of sample is studied deeply anddifferent samples of single-model update strategy and multi-model sampleupdate strategy are proposed.So far, little of the research work has concentrated on Data SamplingTechniques on GRNN-based fitness prediction model for PSO to solvecomputationally expensive problems. Firstly, Two Data Sampling Techniques isproposed in two aspects. One of the Data Sampling Technique is the last positional and fitness information of each particle are selected to update theworse sample data which was selected to build prediction mode in lastgeneration， the other one is to replace the crowded degree sample which wasselected to build prediction mode in last generation. The above two samplingapproaches are used in the early stage of the algorithm， and later in thealgorithm， using the personal best position of each particle help to improve thefinal accuracy of the algorithm. Simulation results show that the aboveprediction strategy reduces the computational cost in the premise of eperformance of guaranteed. Next， this paper proposed a strategy based onmulti-model to predict fitness value for PSO. The personal best position of eachparticle was selected to build one of the prediction model， and the lastpositional and fitness information of each particle are selected to build one ofthe prediction model. Then the two predicted linear weighted value as thefitness value of particles. Simulation results show that the algorithm has highefficiency and performance.