Robust Evolutionary Optimization Based on Coevolution

Abstract

A way to deal with uncertainties in the fitness function of an optimization problem is robust optimization, which optimizes the expected value of the fitness. In the context of evolutionary optimization, it is a common practice to compute the expected value of the fitness approximately with the help of Monte-Carlo simulation. This approach requires a lot of evaluations of the fitness function in order to evaluate an individual and thus it can be very compute-intensive. In the present paper, we propose a coevolution-based approach for the robust optimization of problems with a fitness function basically depending on discrete random variables, which conditionally depend on the decision variables. Experiments on three benchmark functions show that the approach yields a good trade-off between the number of required fitness function evaluations and the quality of the results.