Using regression to improve local convergence

Traditionally Evolutionary Algorithms (EAs) choose candidate solutions based on their individual fitnesses, usually without directly looking for patterns in the fitness landscape discovered. These patterns often contain useful information that could be used to guide the EA to the optimum. While an EA is able to quickly locate the general area of a peak, it can take a considerable amount of time to refine the solution to accurately reflect its true location. We present a new technique that can be used with most EAs. A surface is fitted to the previously-found points using a least squares regression. By calculating the highest point of this surface we can guide the EA to the likely location of the optimum, vastly improving the convergence speed. This technique is tested on Moving Peaks, a commonly used dynamic test function generator. It was able to significantly outperform the current state of the art algorithm.