A Distributed Neural Hybrid System Learning Framework in Modeling Complex Dynamical Systems

In this article, a distributed neural network modeling framework including a novel neural hybrid system model is proposed for enhancing the scalability of neural network models in modeling dynamical systems. First, high-dimensional training data samples will be mapped to a low-dimensional feature space through the principal component analysis (PCA) featuring process. Following that, the feature space is bisected into multiple partitions based on the variation of the Shannon entropy under the maximum entropy (ME) bisecting process. The behavior of subsystems in the prespecified state space partitions will then be approximated using a group of shallow neural networks (SNNs) known as extreme learning machines (ELMs), and then it can further simplify the model by merging the redundant lattices based on their training error performance. The proposed modeling framework can handle high-dimensional dynamical system modeling problems with the advantages of reducing model complexity and improving model performance in training and verification. To demonstrate the effectiveness of the proposed modeling framework, examples of modeling the LASA dataset and an industrial robot are presented.