The Research Of Intelligent Learning And Optimization Methods Based On Evolutionary Algorithms

Evolutionary algorithm(EA) simulates the evolutionary process and uses the generalized hierarchical computer programs which can dynamically change the structures according to the environment to describe the problem. It’s suitable for solving the complex searching and optimization problems and has broad application prospects in the field of intelligent learning.The work of paper mainly falls into three parts.In order to avoid the phenomenon of premature convergence in EA(especially in Genetic Programming) and accelerate the convergence speed at the same time, six different diversity measures, i.e. genotypes, phenotypes, entropy, pseudo-isomorph, edit distance 1 and 2 are adopted to examine the effects that genetic operators have on the population diversity during the evolution process. The focus is to analyze the changes of individual structure and behavior made by different genetic operators and their combinations in discrete and continuous fitness space, then the resulting influence on population diversity and ultimately the ability of the total algorithm using the benchmark problems. Tries to measures the correlation between individual fitness and population diversity are also made. There is no doubt that low diversity will make EA converge to local optimal. It can be discovered that the selection and crossover operators make the diversity decrease in different degree while the mutation operator can maintain large amount of diversity and even increase it. The simulation experiments show that it is feasible to alter the genetic operators in order to control the population diversity.According to the former work, two kinds of edit distance are adopted respectively to alter genetic operators to make the evolution switch between the period of exploration(global search) which increases the diversity and the period of exploitation(local search) which is to decrease the diversity according to the thresholds set in advanced. Based on this mechanism,an improved Genetic Programming is presented. The same benchmark problems used before are adopted to test the feasibility of the algorithm based on diversity control. The results show that in the optimization problems with continuous fitness space, maintaining a right amount of population diversity by changing the operators can not only improve the performance of the EA so as to obtain better solutions to the problem with lower fitness values, but also reduce large amount of computation and save lots of time.Random Drift Particle Swarm Optimization(RDPSO) algorithm is introduced to solve multi-objective optimization problems(MOP). In order to maintain the diversity of particle swarm, two kinds of multi-objective RDPSO algorithm(MORDPSO) based on crowding distance and adaptive grid are proposed. Based on the same diversity maintenance,comparative experiments among multi-objective Quantum-behaved PSO(MOQPSO) and classical PSO with inertia weight(MOPSO-In) in the aspects of algorithm convergence, the distribution of solutions and running time are made. The experimental results denote that the proposed algorithms converge fast and have good ability of global search which cost less time.