Study On Improved Particle Swarm Optimization And Its Application In Hybrid Artificial Neural Network

With the overlapping, penetration and promotion between some subjects, theresearchers have focused their considerable attention on all kinds of intelligentoptimization algorithm to artificial neural network (ANN) training. The particleswarm optimization (PSO) algorithm is a classical swarm intelligence algorithm, withthe advantages of simple structure, less parameters, easy to describe and implementand with good global search capability, and so on., which is widely used in manyfields such as function optimization, multi-objective solving, and pattern recognition.But the standard PSO algorithm has some shortcomings such as prematureconvergence and poor local search capability. The algorithm has been prematureconvergence when the population may have gathered in a stagnation point and havenot the global optimal if it is applied to the high dimensional optimization problem.The emergence of premature convergence does not ensure the algorithm can convergeto the global minimum point. At the same time, the convergence speed becomesslowly when the particles have searched the area near the global values, it is that poorlocal search capability is shown at the end of iteration. Aiming at the shortcomings ofthe PSO algorithm, researchers have proposed many improvement strategy, althoughthe performance and efficiency are improved, therefore, research on intelligentalgorithm with higher precision, better performance, efficiency, correlation and wideradaptability are still the first-line goal in academia and industiral community.In this thesis, in order to improve the convergence speed and the performance ofPSO algorithm, an improved PSO algorithm is proposed and applied to the ANNtraining, called hybird neural network model. Finally, the model is applied to theforeign trade export forecast. The main research content are as follows:(1) A novel improved particle swarm algorithm based on self-adaptive weightadjustment strategy, chaos theory, and simulated annealing algorithm (SA) isproposed, called CSAPSO-SA. In the CSAPSO-SA algorithm, the adaptive weightadjustment strategy is used to improve the convergence speed; and the chaoticsequence produced by chaos theory is used to tune the learning factor with thepurpose of balance the exploitation and exploration, and improved the prematureconvergence problem. The local search speed and convergence precision of PSOalgorithm have been greatly improved by the SA algorithm. Finally, the CSAPSO-SAalgorithm is tested by four multi-objective test function, and compared with the traditional multi-objective algorithm such as NSGA II and multi-objective PSOalgorithm. The results show that the C SAP SO-SA algorithm has faster convergencespeed, higher precision, and better diversity.(2) A hybird artificial neural network model (HANN) is proposed combinedCSAPSO-SA algorithm and RBF ANN, called CSAPSO-SA RBF ANN, or HANN.In the CSAPSO-SARBF ANN, CSAPSO-SA algorithm is used to tune the functionhidden centers and spreads. HANN has combined the advantages of each algorithm,so as to improve the performance of HANN model.(3) Prediction model of foreign trade export is established based on HANN.Through the foreign trade export instances of zhejiang and chongqing, the HANNmodel is approved that it is feasible and reliable for foreign trade export prediction.Compared with conventional RBF ANN and PSO ANN, the HANN shows betterperformance with better accuracy and correlation.In this thesis, a high performance algorithm and a HANN model are proposedbased on PSO algorithm, SA algorithm, chaos theory, and ANN technology, whichprovide a viable and effective methold for foreign trade export prediction. At the sametime, the proposed CSAPSO-SA algorithm may be used for reference for design inmany industrial and research fields, it has a good application prospect.