Predictive Uncertainty Estimation with Temporal Convolutional Networks for Dynamic Evolutionary Optimization

Abstract

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.