Almuth Meier

Abstract

Evolutionary algorithms (EAs) are a good choice to solve dynamic optimization problems. Objective functions changing over time are challenging because after a change the EA has to adapt its population to find the new optimum. Prediction techniques that estimate the position of the next optimum can be incorporated into the EA. After a change, the predicted optimum can be employed to move the EA's population to a promising region of the solution space in order to accelerate convergence and improve accuracy in tracking the optimum. In this paper we introduce a recurrent neural network-based prediction approach. In an experimental study on the Moving Peaks Benchmark and dynamic variants of the Sphere, Rosenbrock, and Rastrigin functions we compare it to an autoregressive prediction approach and an EA without prediction. The results show the competitiveness of our approach and its suitability especially for repeated optima.