Reducing local optima in single-objective problems by multi-objectivization

Abstract

One common characterization of how simple hill-climbing optimization methods can fail is that they become trapped in local optima - a state where no small modification of the current best solution will produce a solution that is better. This measure of `better' depends on the performance of the solution with respect to the single objective being optimized. In contrast, multi-objective optimization (MOO) involves the simultaneous optimization of a number of objectives. Accordingly, the multi-objective notion of `better' permits consideration of solutions that may be superior in one objective but not in another. Intuitively, we may say that this gives a hill-climber in multi-objective space more freedom to explore and less likelihood of becoming trapped. In this paper, we investigate this intuition by comparing the performance of simple hill-climber-style algorithms on single-objective problems and multi- objective versions of those same problems. Using an abstract building- block problem we illustrate how `multi-objectivizing' a single-objective optimization (SOO) problem can remove local optima. Then we investigate small instances of the travelling salesman problem where additional objectives are defined using arbitrary sub-tours. Results indicate that multi-objectivization can reduce local optima and facilitate improved optimization in some cases. These results enlighten our intuitions about the nature of search in multi-objective optimization and sources of difficulty in single-objective optimization.