In this lecture, we expand upon the first two classes of multi-objective genetic algorithms (i.e., those based on random weights and those based on alternating fitness objectives) and add the third and most modern class of approaches – Pareto-ranking approaches. After reviewing the first two classes, we discuss how a new "fitness" can be used that represents how many other solutions a focal solution dominates. By selecting according to this fitness, solutions are driven toward a Pareto frontier without collapsing to any particular region of that frontier. To further improve diversity, "fitness sharing" is introduced. We close with a brief introduction to distributed GA's and multi-modal optimization and hint at a connection to Sewall Wright's "shifting-balance theory" from population genetics.
Whiteboard notes for this lecture can be found at:
No comments:
Post a Comment