An Evolutionary Training Algorithm for Artificial Neural Networks with Dynamic Offspring Spread and Implicit Gradient Information

摘要

Evolutionary training methods for Artificial Neural Networks can escape local minima. Thus, they are useful to train recurrent neural networks for short-term weather forecasting. However, these algorithms are not guaranteed to converge fast or even converge at all due to their stochastic nature. In this paper, we present an algorithm that uses implicit gradient information and is able to train existing individuals in order to create a dynamic reproduction probability density. It allows us to train and re-train an Artificial Neural Network supervised to forecast weather conditions.