Design Optimization Using Dynamic Evaluation

Abstract

We describe a search strategy that may be useful for a class of design problems by developing an example from cancer radiation treatment planning. This application problem involves typical features of design problems such as constraints, optimality, a large search space with continuously varying parameters as well as discrete (non-numeric) parameters. There is no known method of comparing elements of the solution space based on a static evaluation function. We have therefore developed a dynamic evaluation function, which attempts to heuristically compare all solutions with one another, as a way of interpreting the evaluation results. This allows us to use an analog of hill-climbing with a simple SELECT-GENERATETEST loop where expert rules are used as "move generators" and a similarity metric is used to control or direct the application of the rules for plan modification. Preliminary tests of these ideas indicate that a practical working system can be built.