Recurrent Stochastic Configuration Networks With Hybrid Regularization for Nonlinear Dynamics Modeling.

Area of Science:

• Machine Learning
• Nonlinear Dynamics
• System Identification

Background:

• Recurrent Stochastic Configuration Networks (RSCNs) show promise for modeling complex dynamic systems.
• Existing methods may lack robust generalization and learning capacity for uncertain systems.

Purpose of the Study:

• To enhance the learning capacity and generalization performance of RSCNs.
• To develop a hybrid regularization technique for improved nonlinear system modeling.

Main Methods:

• Utilized the Least Absolute Shrinkage and Selection Operator (LASSO) for significant variable identification in temporal data.
• Introduced an improved RSCN with L2 regularization to model residuals from the LASSO approximation.
• Employed a real-time projection algorithm for output weight updates.

Main Results:

• The proposed hybrid regularization method significantly improved RSCN performance.
• Demonstrated superior accuracy in nonlinear system identification compared to existing models.
• Achieved high performance in two industrial predictive tasks across all datasets.

Conclusions:

• The hybrid regularization approach effectively enhances RSCN capabilities for nonlinear dynamic systems.
• The method offers a robust solution for system identification and predictive modeling under uncertainty.
• Theoretical analysis supports the network's universal approximation property for complex functions.