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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Dynamical Systems Modeling

Background:

• Neural network models face scalability challenges in complex dynamical systems.
• High-dimensional data complicates traditional modeling approaches.

Purpose of the Study:

• To propose a distributed neural network framework for scalable dynamical system modeling.
• To enhance model complexity reduction and performance in training and verification.

Main Methods:

• Utilized principal component analysis (PCA) to map high-dimensional data to a low-dimensional feature space.
• Employed maximum entropy (ME) and Shannon entropy for feature space partitioning.
• Applied extreme learning machines (ELMs), a type of shallow neural network (SNN), to approximate subsystem behavior.
• Integrated a model simplification step by merging redundant partitions based on training error.

Main Results:

• The framework effectively handles high-dimensional dynamical system modeling.
• Demonstrated reduced model complexity and improved performance on the LASA dataset and an industrial robot model.
• The novel neural hybrid system model enhances scalability.

Conclusions:

• The proposed distributed neural network framework offers a scalable and efficient solution for modeling complex dynamical systems.
• The integration of PCA, ME partitioning, and ELMs provides a robust approach to reduce complexity and improve accuracy.