Abstract
Ensuring the stability of dynamical systems under parametric uncertainty is a critical challenge in control theory. Traditional Neural Lyapunov Functions (NLFs) provide large Regions of Attraction (ROA) but lack the flexibility to adapt to changing physical parameters without extensive retraining. This paper introduces Meta-Neural Lyapunov Functions (meta-NLFs), a framework that leverages Model Agnostic Meta-Learning (MAML) to enable rapid adaptation of stability certificates. By meta-training across a distribution of systems, the proposed method can generate a task-adapted NLF with minimal data and gradient updates. Experimental results on benchmarks like inverted pendulums and microgrids demonstrate that meta-NLFs significantly outperform classical and transfer-learning approaches in both ROA size and adaptation speed.
Introduction
In control theory, stability is typically certified using a Lyapunov function—a scalar function that decreases along the system's trajectories toward an equilibrium. The set of states where this property holds defines the Region of Attraction (ROA). While recent deep learning approaches have successfully used neural networks to represent complex ROAs, these models are often brittle. A change in a robot's mass or a power grid's impedance can invalidate a pre-trained NLF. Re-training from scratch is too slow for real-time safety, creating an "adaptation gap." This work addresses how to quickly obtain a valid stability certificate for a new system instance using minimal data.
Methodology: Meta-Neural Lyapunov Functions
The authors propose integrating MAML into the NLF training pipeline to create meta-NLFs. The process involves two main stages:
- Meta-Training: Instead of optimizing for a single system, the model learns a meta-parameter initialization across a variety of similar systems. This initialization is designed to be "one step away" from the optimal parameters for any specific task in the distribution.
- Fast Adaptation: Upon encountering a new system, the meta-NLF undergoes a few gradient descent steps using a small local dataset to produce a task-adapted NLF.
- Verification: To guarantee mathematical validity, a Lipschitz-based verification technique checks the Lyapunov conditions (positivity and negative-definite derivative) across a discretized grid.
Experimental Evaluation
The framework was tested on four benchmark systems:
- Inverted Pendulum (IP): A standard nonlinear control task.
- 3-Microgrid and 5-Microgrid Systems (3-MG, 5-MG): Power network models with varying droop constants.
- Caltech Ducted Fan (CF): A flight control model in hover mode.
Parameters such as mass, length, and friction were treated as stochastic variables to simulate real-world uncertainty.
Key Results and Insights
The meta-NLF approach demonstrates a significant advantage over Sum-of-Squares (SOS), Quadratic (QLF), and standard Transfer-Learning (T-NLF) methods.
- Superior ROA Coverage: In the 3-microgrid benchmark, the task-adapted NLF achieved an ROA area of 52.89, vastly exceeding the 1.60 achieved by QLF and 6.82 by SOS.
- Robustness and Efficiency: While SOS methods often failed to converge for complex systems like the 5-microgrid or Caltech Fan, meta-NLF consistently produced valid regions. Adaptation required as few as 50 samples and 10 gradient steps.
- The Failure of Naive Transfer: Standard transfer learning often fails because the starting point is too biased toward the original "nominal" system. Meta-learning provides a neutral starting point optimized for rapid shifting.
Conclusion and Future Work
Meta-NLFs bridge the gap between the descriptive power of neural networks and the need for rapid safety certification in changing environments. However, the current grid-based verification method faces the "curse of dimensionality" as system complexity grows. Future research will explore Meta-Reinforcement Learning to co-learn adaptive control policies and stability certificates, as well as extending the framework to barrier functions and contraction metrics.
Summary updated