Memory neuron networks for identification and control of dynamical systems

Area of Science:

• Artificial Intelligence
• Control Systems Engineering
• Computational Neuroscience

Background:

• Nonlinear dynamical systems present significant challenges in modeling and control.
• Traditional neural networks often require explicit historical data, limiting their application to systems with unknown dynamics or delays.
• The need for neural network architectures with inherent memory capabilities is crucial for advanced system identification.

Purpose of the Study:

• To introduce and analyze memory neuron networks (MNNs) as a novel approach for identifying and adaptively controlling nonlinear dynamical systems.
• To demonstrate the capability of MNNs to handle systems with unknown orders or time delays.
• To provide theoretical justification for the learning algorithm and present adaptive control strategies.

Main Methods:

• Development of recurrent neural networks with trainable temporal elements (memory neuron networks).
• Utilizing the history-sensitive output of MNNs for system identification without explicit input/output history.
• Implementing adaptive control methods based on MNN models for nonlinear systems.

Main Results:

• MNNs demonstrated effectiveness in identifying nonlinear dynamical systems, even those with unknown orders or delays.
• The learning algorithm provided theoretical justification for the identification accuracy.
• Simulations confirmed the efficacy of MNNs for both identification and model reference adaptive control.

Conclusions:

• Memory neuron networks offer a powerful and flexible framework for modeling and controlling complex nonlinear dynamical systems.
• The inherent memory capability of MNNs overcomes limitations of traditional feedforward and recurrent networks in handling unknown system dynamics.
• MNNs represent a significant advancement in adaptive control and system identification research.