Particle Swarm Optimization Based On Space Contraction And Its Application In Investment Prediction

Optimization is a kind of applied technique to find optimal solutions in various engineering problems. For its wide application in these years, it is highly concerned and rapidly developed. With the expansion of application fields, the time and space complexity of optimization problems makes it more and more difficult to solve them, so that traditional optimization algorithms cannot meet the needs in practice. In recent years, the birth of intelligent optimization algorithms provides new ideas and means to optimization techniques and the intelligent algorithms, which are applied widely and develop rapidly in science research, economics and engineering technology. Particle Swarm Optimization is a new kind of random search swarm intelligent algorithm. As an intelligent optimizer, it shows potential in solving complex optimization problems, and becomes a hot topic in scientific researches.In this dissertation, firstly we briefly introduce some optimization knowledge related to PSO. Secondly, we systematically discuss the basic principle, the process of implementation and primary research results. Then, we do some deep research in improvement of the algorithm and its application. The main contributions given in the dissertation are as follows.1. The dissertation points out the reason for premature convergence by the empirical analysis from numerical experiments. Then we introduce the space contraction mechanism to PSO, and according to the different ways of space contraction, we present two models of PSO algorithm based on space contraction, that is, the elite evolution model with population extinction (EEPV) and elite evolution model with bad particles eliminated(EEBPE). We expound the principles and the influence of important parameters to the algorithm performance of both algorithms presented in the dissertation.2. We give reasonable suggestion to the choice of the parameters in both algorithms by series of numerical experiments. In order to study the performance of our algorithms, we did series of comparison experiments with some classical improved algorithms on five representative benchmark test functions. The experimental results show that, both algorithms can effectively avoid premature convergence and meanwhile enhance local search capability in late of algorithms, thus improve algorithm performance greatly, and both algorithms show obvious advantages in search for optima in the high-dimensional complex functions, especially in 30-D Rosenbrock function. In addition, the second model outperforms than other algorithms in optimization effects such as the best-quality of the optima, search speed etc.3. Both improved algorithms are applied to training feed-forward network as learning algorithm. The principle and implementation process based on space contraction PSO are studied, finally the feed-forward network based on space contraction PSO are successfully applied in the prediction of investment problem. The experimental results illustrate the efficiency. The feed-forward network based on space contraction PSO has faster training speed and better generalization capacity than BP learning algorithm, and it has much better prediction accuracy in the model of investment prediction compared with BP algorithm.