Thursday 02 Feb 2017: Evolutionary multi-objective optimisation: effectiveness, scalability and efficiency
Dr. Ke Li - University of Exeter
Harrison 170 14:30-15:30
Evolutionary multi-objective optimization has been studied for nearly three decades. Many effective methods have been developed for solving problems with two and three objectives. However, with the development of science and technology, real life applications pose more significant challenges in optimization, e.g., the complicated problem structure and a large number of objectives. In this talk, I will present some of my recent advances towards these challenges. Different from the traditional evolutionary multi-objective methods, the original multi-objective optimization is at first decomposed into a set of subproblems, either a single-objective scalarization or a simplified multi-objective optimization problem. Then, a population-based technique is applied to solve these subproblems in a collaborative manner. In particular, we model the environmental selection process as a matching process and the final selection result is guided by the stable matching between subproblems and solutions. I will show the effects of the length of the preference list to the selection result. Experimental results demonstrate our idea on problems with complicated properties and a large number of objectives. Apart from the problem solving effects, I will also discuss our recent progress on how to accelerate the evolutionary multi-objective optimization by replacing the non-dominated sorting with a non-domination level update method.