Few-Shot Closed-Loop Neural System Identification via Meta-Learning

TMLR Paper8785 Authors

Abstract:

We study few-shot closed-loop neural system identification in a meta-learning setting, where related source systems are used to learn a neural-network-based open-loop dynamics model for a new target system from limited feedback-controlled data.
In closed loop, inputs are generated through output feedback; consequently, the observed trajectories are shaped by both the plant dynamics and the controller.
Under feedback-dependent data generation and scarce target data, existing system identification methods are insufficient for recovering the open-loop dynamics.
Based on meta-learning and neural closed-loop identification, we propose Meta-ICI, which learns an initialization for an intermediate operator to recover open-loop dynamics from limited target closed-loop data.
We further extend Meta-ICI to fragmented target adaptation, where only scattered one-step transitions are available instead of continuous trajectories.
This extension yields Fast Meta-ICI for fully observable systems, using fragmented transitions to support accurate long-horizon rollouts.
To instantiate Fast Meta-ICI, we design a Schur-Koopman model that enforces the latent spectral-radius constraint during unconstrained optimization.
Experiments on partially observable Li\'enard systems and fully observable nonlinear pendulum systems show that Meta-ICI improves few-shot adaptation and Fast Meta-ICI enables non-divergent long-horizon rollouts from fragmented target data.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Amir-massoud_Farahmand1

Submission Number: 8785