Meta-Learning-Based Adaptive Stability Certificates for Dynamical Systems

Amit Jena, Dileep Kalathil, Le Xie

Introduction

In the field of control theory, ensuring the stability of a dynamical system—guaranteeing it will return to a desired equilibrium rather than diverging—is a fundamental requirement. This is typically achieved by identifying a Lyapunov function, a scalar function that acts as a mathematical "energy" metric. If this energy consistently decreases along the system's trajectories, the system is certified as stable. The set of initial states from which the system is guaranteed to converge is known as the Region of Attraction (ROA).

While traditional methods like Sum-of-Squares (SOS) provide rigorous certificates, they struggle to scale with high-dimensional, complex systems. Recent advances in Neural Lyapunov Functions (NLFs) leverage deep learning to find these certificates for complex dynamics. However, NLFs are notoriously brittle; a slight change in physical parameters (e.g., a change in a drone's mass or a pendulum's length) renders the pre-trained function invalid. Re-training from scratch is often too slow for safety-critical, real-time applications.

Methodology: Meta-Neural Lyapunov Functions

To address this brittleness, this research introduces Meta-Neural Lyapunov Functions (meta-NLFs). This framework treats the search for a stability certificate as a meta-learning problem, specifically utilizing Model-Agnostic Meta-Learning (MAML).

The "Master Key" Initialization

Instead of training a neural network for one specific system configuration, meta-NLFs optimize for a "meta-parameter" initialization. This initialization serves as a mathematical starting point that is strategically positioned to be close to the Lyapunov functions of many different system variations. When a new parametric configuration is encountered, the system can adapt this "master key" using only a few gradient updates and a small amount of local data.

Visualization of the learned Lyapunov surfaces (a, c) and their corresponding stability regions (b, d) during the adaptation process.

Verification and Safety Margins

To ensure the adapted function provides a valid certificate, the authors employ a Lipschitz-based verification technique. By calculating the maximum possible rate of change of the neural network, they can guarantee that if Lyapunov conditions hold on a discrete grid, they hold throughout the continuous state space. Furthermore, the loss function incorporates safety margins to maintain robustness against discretization errors.

Experimental Validation

The meta-NLF approach was evaluated on several standard benchmarks, including inverted pendulums, multi-microgrid power systems, and the Caltech Ducted Fan. In these tests, physical constants like mass, inertia, and friction were varied to simulate real-world parametric shifts.

Key Results

  • Data Efficiency: Meta-NLFs adapted to new systems using as few as 50 training samples and 10 gradient steps, whereas standard NLFs require thousands of samples.
  • Superior ROAs: The framework consistently produced larger certified Regions of Attraction compared to traditional Quadratic Lyapunov Functions (QLF) or SOS methods.
  • Reliability: While standard transfer learning often failed to find any valid stability region after a significant parameter shift, meta-NLFs successfully adapted in every test case.
Phase plot for an inverted pendulum showing system trajectories (colored lines) starting from various initial conditions (blue dots) and converging to the origin within the certified ROA (red boundary).

Conclusion and Future Work

Meta-NLFs represent a significant step toward autonomous systems that can certify their own safety in changing environments. By combining the flexibility of deep learning with the rapid adaptation of meta-learning, this framework provides a computationally tractable path toward verifying high-dimensional systems without the need for slow SMT solvers.

Future research aims to integrate meta-NLFs with Meta-Reinforcement Learning to develop control policies that are inherently aware of their stability boundaries. Additionally, the authors plan to explore meta-learning for "Barrier Functions" to provide formal guarantees for avoiding specific unsafe regions of the state space.

Summary updated

Abstract

This paper addresses the problem of Neural Network (NN) based adaptive stability certification in a dynamical system. The state-of-the-art methods, such as Neural Lyapunov Functions (NLFs), use NN-based formulations to assess the stability of a non-linear dynamical system and compute a Region of Attraction (ROA) in the state space. However, under parametric uncertainty, if the values of system parameters vary over time, the NLF methods fail to adapt to such changes and may lead to conservative stability assessment performance. We circumvent this issue by integrating Model Agnostic Meta-learning (MAML) with NLFs and propose meta-NLFs. In this process, we train a meta-function that adapts to any parametric shifts and updates into an NLF for the system with new test-time parameter values. We demonstrate the stability assessment performance of meta-NLFs on some standard benchmark autonomous dynamical systems.
BibTeX
@inproceedings{jena2024meta,
  author = {Amit Jena and Dileep Kalathil and Le Xie},
  title = {{Meta-Learning-Based Adaptive Stability Certificates for Dynamical Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i11.29176},
}
Figure