Recurrent stochastic configuration networks with block increments

Area of Science:

• Computational neuroscience
• Machine learning
• Nonlinear dynamics

Background:

• Recurrent stochastic configuration networks (RSCNs) are effective for nonlinear dynamic systems with order uncertainty.
• Existing RSCNs offer ease of implementation, reduced human intervention, and strong approximation capabilities.

Purpose of the Study:

• Introduce block recurrent stochastic configuration networks (BRSCNs) to improve learning capacity and efficiency.
• Enhance the modeling of complex nonlinear dynamic systems.

Main Methods:

• Develop BRSCNs capable of adding multiple reservoir nodes (subreservoirs) simultaneously.
• Configure each subreservoir with unique structures using a supervisory mechanism.
• Scale the reservoir feedback matrix to ensure the echo state property.
• Employ online output weight updates via a projection algorithm.
• Establish persistent excitation conditions for parameter convergence.

Main Results:

• BRSCNs demonstrate superior modeling efficiency and learning performance.
• The proposed method shows favorable generalization performance across various tasks.
• Effectiveness validated on time series prediction, nonlinear system identification, and industrial data analysis.

Conclusions:

• BRSCNs offer significant potential for modeling complex dynamics with enhanced efficiency.
• The novel architecture improves upon traditional RSCNs for dynamic system analysis.
• BRSCNs provide a robust framework for tackling challenging nonlinear modeling problems.