Real-Time Progressive Learning: Accumulate Knowledge From Control With Neural-Network-Based Selective Memory.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Memory is fundamental to learning, influencing knowledge storage, updating, and forgetting.
• Conventional neural network learning control (NNLC) often relies on Lyapunov-based methods, prioritizing stability and performance.
• NNLC may suffer from gradual knowledge forgetting over time.

Purpose of the Study:

• To propose a novel radial basis function neural network (RBFNN)-based learning control scheme, real-time progressive learning (RTPL).
• To enhance learning speed, robustness, generalization, and knowledge retention compared to traditional NNLC.
• To guarantee system stability and closed-loop performance while learning unknown system dynamics.

Main Methods:

• Developed the real-time progressive learning (RTPL) scheme utilizing a radial basis function neural network (RBFNN).
• Employed the selective memory recursive least squares (SMRLS) algorithm for neural network weight updates.
• Validated the approach through theoretical analysis and simulation studies.

Main Results:

• RTPL demonstrated improved learning speed without filtering and robustness to hyperparameter settings.
• The scheme exhibited good generalization ability, allowing learned knowledge reuse across different tasks.
• RTPL ensured guaranteed learning performance under parameter perturbation and continuous knowledge accumulation.

Conclusions:

• RTPL offers significant advantages over conventional NNLC, particularly in learning speed, robustness, and knowledge retention.
• The SMRLS algorithm effectively manages memory, preventing the forgetting of learned information.
• RTPL provides a stable and high-performance learning control solution for systems with unknown dynamics.