Neural Network Models for Time-Varying System Solutions

Neural network models have become indispensable tools for solving dynamic problems in which system parameters evolve over time. By recasting time-varying equations or optimisation tasks as zero-finding or minimisation problems, specialised architectures such as zeroing neural networks and recurrent structures can track solutions in real time. Continuous-time formulations ensure smooth adaptation, while discrete-time variants allow efficient implementation on digital hardware. Advances in activation-function design and adaptive compensation yield faster convergence and enhanced stability, even in the presence of noise or perturbations. Key applications span robotics path tracking, adaptive control of chaotic systems, signal processing under fluctuating conditions and financial optimisation with time-dependent constraints. Global interest in these methods arises from their ability to embed physical or mathematical models directly within learning architectures, thereby combining theoretical guarantees of convergence with the flexibility of data-driven adaptation. Recent work has emphasised finite-time convergence, robustness to harmonic or random disturbances and integration with classical numerical techniques such as singular value decomposition and regularisation. The evolving landscape of time-varying neural solvers underscores a trend towards unifying rigorous stability analysis with practical demands for speed and resilience in complex, non-stationary environments.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent studies have refined zeroing neural networks for matrix pseudoinversion by integrating singular value decomposition and Tikhonov regularisation, yielding continuous-time models that adaptively compute inverses or pseudoinverses of arbitrary time-varying matrices with high accuracy. Other work has addressed the pervasive challenge of harmonic noise in industrial settings by embedding adaptive compensation terms and novel activation functions into zeroing-based architectures, achieving rapid convergence and robustness when controlling robotic manipulators under noisy conditions. In parallel, enhancements to recurrent neural network designs for time-varying quadratic programming problems have introduced non-linear activation functions to overcome convexity constraints and redundant formulations, thereby improving solution quality and convergence speed for optimisation tasks with evolving equality and inequality constraints. Together, these contributions demonstrate a cohesive progression: from foundational zeroing principles through noise-tolerant adaptations to advanced activation schemes, all aimed at delivering reliable, real-time solutions for a breadth of dynamic systems.

Neural Network Models for Time-Varying System Solutions publication trend

The graph below shows the total number of articles in neural network models for time-varying system solutions across all publications each year (not limited to Nature Index journals).

Technical terms

Zeroing Neural Network (ZNN): A continuous-time or discrete-time neural architecture designed to drive an error function to zero, thereby tracking solutions of time-dependent equations in real time.

Activation Function: A non-linear mapping applied within neural units to shape response characteristics, influence convergence speed and ensure stability in dynamic solvers.

Convergence Rate: A measure of how quickly a neural solver’s output approaches the true solution of a time-varying problem, often quantified by exponential or finite-time bounds.

Pseudoinverse: A generalised matrix inverse computed for possibly non-square or singular matrices, enabling least-squares solutions in time-varying environments.

Regularisation: A technique, such as Tikhonov regularisation, used to stabilise inverse computations and mitigate the effects of noise or ill-conditioned data.

References

1. Zeroing neural networks: A survey. Neurocomputing (2017).
2. Zeroing Neural Network for Pseudoinversion of an Arbitrary Time-Varying Matrix Based on Singular Value Decomposition. Mathematics (2022).
3. Harmonic Noise-Tolerant ZNN for Dynamic Matrix Pseudoinversion and Its Application to Robot Manipulator. Frontiers in Neurorobotics (2022).
4. Design and analysis of recurrent neural network models with non‐linear activation functions for solving time‐varying quadratic programming problems. CAAI Transactions on Intelligence Technology (2021).
5. Finite-Time Convergence and Robustness Analysis of Two Nonlinear Activated ZNN Models for Time-Varying Linear Matrix Equations. IEEE Access (2019).
6. Rejecting Chaotic Disturbances Using a Super-Exponential-Zeroing Neurodynamic Approach for Synchronization of Chaotic Sensor Systems. Sensors (2018).
7. A Velocity-Level Bi-Criteria Optimization Scheme for Coordinated Path Tracking of Dual Robot Manipulators Using Recurrent Neural Network. Frontiers in Neurorobotics (2017).
8. Portfolio Insurance through Error-Correction Neural Networks. Mathematics (2022).