Prediction strategies in dynamic evolutionary optimization aim at estimating the moving optimum after a change of the fitness function. Considering the predicted optimum for re-initialization of the population, the evolution strategy is led into the direction of the next optimum. We propose a new way to control the influence of the prediction depending on its estimated uncertainty. In addition, we construct a new benchmark generator for dynamic optimization problems, Dynamic Sine Benchmark, tailored to prediction approaches. For prediction of the moving optimum and uncertainty estimation we apply a temporal convolutional network (TCN) with Monte Carlo dropout. In the experimental study, we compare our approach to known prediction and re-initialization strategies. The results show the advantage of the new re-initialization strategy and TCNs with uncertainty estimation for complex problems up to a certain dimensionality.
CITATION STYLE
Meier, A., & Kramer, O. (2019). Predictive Uncertainty Estimation with Temporal Convolutional Networks for Dynamic Evolutionary Optimization. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11728 LNCS, pp. 409–421). Springer Verlag. https://doi.org/10.1007/978-3-030-30484-3_34
Mendeley helps you to discover research relevant for your work.