Evolutionary Optimization of Multi-step Dynamic Systems Learning

Abstract

This paper develops an optimization framework based on evolutionary computation for the multi-step prediction enhancement of Deep Learning-based Dynamic Systems simulation models. We propose using the Differential Evolution algorithm and an Autoencoder network to find the optimal arrangement that accurately models a nonlinear system. A series of experiments are performed using a nonlinear, chaotic benchmark system: the double pendulum to validate our claims. As a result, the prediction error and confidence level were increased by an average of 20% against conventional parameters selection methods. Furthermore, we found that the training stage relies less on trial and error approaches in favor of a quantitative objective function using an optimization method.