Predictive Uncertainty Estimation with Temporal Convolutional Networks for Dynamic Evolutionary Optimization
Predictive Uncertainty Estimation with Temporal Convolutional Networks for Dynamic Evolutionary Optimization
复制标题
DOI：
10.1007/978-3-030-30484-3_34
复制
发表时间：
2019-09
期刊：
影响因子：
--
通讯作者：
Almuth Meier;Oliver Kramer
中科院分区：
文献类型：
--
作者：
Almuth Meier;Oliver Kramer
动态进化优化中的预测策略旨在通过改变适应度函数来估计移动最优解。考虑种群重新初始化的预测最优值，将进化策略引入下一个最优值的方向。我们提出了一种新的方法来控制预测的影响，依赖于其估计的不确定性。此外，我们还为动态优化问题构造了一种新的基准生成器--动态正弦基准，为预测方法量身定做。对于移动最优预测和不确定性估计，我们应用了一种带有蒙特卡罗丢弃的时间卷积网络(TCN)。在实验研究中，我们将我们的方法与已知的预测策略和重新初始化策略进行了比较。实验结果表明了新的重新初始化策略和具有不确定性估计的TCNs对于维达一定的复杂问题的优越性。
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.