Modifications And Applications Of Genetic Algorithm And Particle Swarm Optimization

It is learning from the phenomena of adaptive optimization in the nature that creatures and natural ecological systems make the optimization problems with high degree of complexty solved perfectly. Under the background, biologically-inspired intelligent computations which are different from the classed optimization methods has emerged. The paper gives a comprehensive study on genetic algorithm (GA) and particle swarm optimization (PSO) from the aspects of algorithm mechanism, modifications and their applications. The main contents and contributions given in this dissertetion are as follows:To improve performance of genetic algorithm and avoid trapping to local optima, a new genetic algorithm based on predatory search strategy is proposed. It simulates the animal predatory search in running, and adjusts the crossover and mutation probability by the best individual fitness in every generation, so can balance the ability of global exploring and local search. The effectiveness and practicability are demonstrated by the simulation results.The reason of premature convengence is large lost in population diversity. The measure of population diversity for measuring poplation maturing degree is presented and its calculation is given on the basis fuzzy system theory. Finally, the method of adaptively adjusting inertia weight with poplation maturing degree is proposed to prevent premature convegence. The algorithm and genetic algorithm based on predatory search strategy have been applied successfully to the parameter estimation of heavy oil thermal cracking model.The convengence of particle velocity in PSO and effect on optimization performances are analyzed. A new PSO is proposed with dynamical adjust inertia weight using information defined as the average absolute value of velocity of all of the particles, which can avoid premature convengence for the velocity is closed to 0 in the early search part.To improve performance of PSO algorithm and avoid trapping to local minima, a multi-population parallel particle swarm optimization (DPPSO) algorithm is proposed. In the algorithm, sub populations are divided into exploration and exploitation types. The global version PSO is used in the exploration population to enhance ability of exploring the best individual, and the local version PSO is used in the exploitation population to enhance ability of local search and find the best global result in the local range. Simultaneously, keep communication with sub populations in running. The experimental results show that the restraining premature convergence is enhanced for maintaining the individual diversity.A new dynamic population structure PSO is proposed, in which the small world network model is introduced into population structure, and the topology structure is dynamical with the evolvement of small world network. So the population diversity is enhanced. The individual value is applied to evolution of small world network for particles can distribute in different search space. In order to perform a better local search, a global PSO version is used in the end of the search. The effectiveness and practicability is demonstrated by the simulation results.A dynamic clustering algorithm based on PSO is proposed, in which a novel coding and operation on the basis of standard PSO is introduced and DB Index rule is used to determine the validity of clustering. The proper fuzzy rule number and exact premise parameters can be obtained by using the dynamic clustering algorithm to identify fuzzy models, and result parameters by the least squared method. It is used in fuzzy modeling for thermal processes.GA and PSO are used in training radial basis (RBF) function network. Firstly, a new RBF network algorithm is introduced by combining genetic algorithm of chromosomes with changeable length and least-square method which can determine the structure and parameters of network. The algorithm is used to model heat loading forecasting for co-generation power plants and compares with BP neural network and RBF neural network based on a trainning algorithm of automatic increase in hidden nodes, the simulation results are satisfied. Second, two methods of RBF neural network algorithm based on PSO are proposed. First is to determine units'number in RBF layer using subtractive clustering method, and optimize central position and directional width based on PSO algorithm, and then train RBF neural network combining with least-square method. Second is to introduce a control gene into standard PSO, and determine network structure and parameters, such as centers and widths of hidden units by combining with least square method. Simulation results illustrate their efficiency.Finally, the whole research contents are summarized, and further research directions are indicated.