Research On The Multi-Objective Evolutionary Algorithms Based On Gradient Growding Diversity Maintenance Strategy

Multi-objective optimization problem (MOP) has been concerned a lot in the fields of science and engineering. In the past decades, due to the progress of evolutionary algorithms, multi-objective evolutionary algorithms have become the focus in multi-objective optimization problem.In MOP, one objective always conflicts with another. Optimizing one objective often depresses the performance of another. Moreover, unlike single optimization problems, there are many optimal solutions called Pareto optimal solutions in MOP and hence some trade-off solutions are always considered. How to find solutions converging to the true Pareto-optimal front with diversity and uniform distribution is the most important issue to multi-objective optimization problem. The population diversity maintenance is important to the convergence of MOP solutions. This thesis focuses on the research of multi-objective evolutionary algorithms based on diversity maintenance strategies.The thesis introduces the background, the state-of-the-art and significance of MOP. Then, focusing on diversity maintenance strategies, the thesis analyzes existing diversity maintenance strategies especially density estimation methods, and a new strategy called Gradient Crowding Algorithm is addressed for pruning non-dominated solutions as well as preserving a wide-spread solution set and maintaining population diversity. Later on, inspired by the conception of Entropy in Information Theory, the Entropy Metrics is defined and applied to assess the new strategy.After the addressing of the new strategy, a multi-objective evolutionary algorithm based on Gradient Crowding diversity maintenance strategy is put forward. In the process of designing the MOEA, a diverse operation is put respectively into initialization and selection. Considering the initialization, most MOEAs use random initial population strategy, but in this thesis, a diverse initial population strategy is proposed so that the population can be widely spread at first. In selection, Gradient Crowding Algorithm is applied to be part of the crowding factor and proceeds density estimation. By this method, solutions can be widely spread on the true Pareto optimal front and the population diversity is kept. In order to improve the adaptation of the algorithm, sectionalized elitism and two-level fitness model for selection are established to provide a unified model.Experimental results show that the proposed algorithms in this thesis improve in diversity maintenance and convergence of MOP solutions.