New delay-dependent stability criteria for neural networks with time-varying interval delay

Jie Chen, Jian Sun, Shuai Liu, D. ReesPhysics 2010 · DOI

TL;DR

This paper develops new, less conservative stability criteria for neural networks with time-varying delays. By introducing novel Lyapunov functionals that better utilize information about the delay's lower bound and incorporating triple-integral terms, the authors provide improved conditions for ensuring network stability, which is crucial for their reliable application in various fields.

What Problem Does This Paper Address?

Neural networks are powerful tools used in diverse applications like pattern recognition and adaptive control. However, their real-world implementation often involves inherent delays in signal transmission and processing. These delays can significantly impact the network's stability, potentially leading to unpredictable behavior or failure. While existing methods exist to analyze the stability of delayed neural networks, many are overly conservative, especially when the delay is small or varies within a specific range. This conservatism limits the practical applicability of these networks. This paper specifically addresses the challenge of analyzing the stability of neural networks with time-varying delays that fall within a defined interval (i.e., $h_1 < d(t) < h_2$), aiming to provide more accurate and less conservative stability conditions.

Methods

The core of this paper lies in the development of novel Lyapunov functionals and the application of Linear Matrix Inequality (LMI) techniques.

  • Lyapunov Functionals: The authors propose two main Lyapunov functionals.
  • The first (Eq. 9) is an improvement over existing functionals by more effectively incorporating the lower bound of the time-varying delay ($h_1$). Previous methods, while using $h_1$, did not fully leverage its information, leading to conservatism.
  • The second (Eq. 28), an "augmented Lyapunov functional," further enhances the analysis by including novel triple-integral terms. These terms are designed to capture more intricate dynamics related to the delay, leading to even less conservative stability conditions.
  • Linear Matrix Inequality (LMI): The stability conditions derived from these Lyapunov functionals are formulated as LMIs. LMIs are a powerful tool in control theory because they can be efficiently solved using convex optimization algorithms. By ensuring that certain matrix inequalities hold, the authors can guarantee the asymptotic stability of the neural network.
  • Free-Weighting Matrices Method: This technique is employed to handle the cross-terms that arise when differentiating the Lyapunov functional, helping to reduce conservatism.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and validation through numerical examples.

Key Insights and Ideas

  • The Importance of the Delay's Lower Bound: A key insight is that simply knowing the upper bound of a time-varying delay is insufficient for precise stability analysis. Effectively utilizing information about the lower bound of the delay ($h_1$) is crucial for reducing conservatism.
  • Lyapunov Functionals as a Powerful Tool: The paper demonstrates that the design of the Lyapunov functional is paramount. By carefully constructing functionals that capture more information about the system's dynamics (like the triple-integral terms), significantly less conservative stability criteria can be achieved.
  • LMIs for Practical Stability Analysis: Formulating stability conditions as LMIs makes them computationally tractable and allows for systematic verification of stability for a given set of parameters.
  • Trade-off Between Conservatism and Complexity: The paper implicitly highlights a trade-off: more complex Lyapunov functionals (with more terms) lead to less conservative results but also increase the computational burden.

Key Results

The paper presents two main theorems (Theorem 1 and Theorem 4) that provide delay-dependent stability criteria for neural networks with time-varying interval delays.

  • Theorem 1: Introduces a new stability criterion based on a refined Lyapunov functional that better utilizes the lower bound of the delay.
  • Theorem 4: Further improves upon Theorem 1 by introducing an augmented Lyapunov functional with triple-integral terms, leading to even less conservative results.
  • Numerical Examples: Two examples are provided to demonstrate the effectiveness of the proposed methods. The results show that the upper bounds on the delay interval ($h_2$) that guarantee stability are larger when using the proposed theorems compared to existing methods in the literature (e.g., [11, 27, 28]). This indicates that the new criteria are indeed less conservative. For instance, in Example 1, Theorem 4 provides a larger upper bound on $h_2$ for various $h_1$ and $\mu$ values compared to previous works.

Limitations and Future Work

  • Computational Complexity: While LMIs are solvable, the complexity of the matrices involved increases with the number of terms in the Lyapunov functional. Theorem 4, with its triple-integral terms, leads to larger LMIs and thus longer computation times (e.g., ~150.3s for Example 1).
  • Unknown $\mu$: The paper notes that Theorem 1 can be extended to the case where the upper bound of the delay ($\mu$) is unknown (Corollary 3), but the details are omitted due to space limitations.
  • Unknown Input ($u$): The current criteria are derived for systems with a known input $u$. The paper mentions that for the case of an unknown input, a rate-independent result can be obtained (Corollary 3).
  • Further Improvements: The authors suggest that combining their approach with other techniques, such as the delay decomposition approach proposed in [26], could lead to even further reductions in conservatism. They also hint at the possibility of extending their work to handle more complex network structures or delay patterns.

Summary updated

BibTeX
@inproceedings{chen2010new,
  author = {Jie Chen and Jian Sun and Shuai Liu and D. Rees},
  title = {{New delay-dependent stability criteria for neural networks with time-varying interval delay}},
  booktitle = {Physics},
  year = {2010},
  doi = {10.1016/j.physleta.2010.08.070},
}
Figure