Experiment Design for Set-membership Identification: From Prior Knowledge to Universal Inputs
arXiv:2607.00844
2026
Training
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper develops universal experiment inputs that guarantee identification for every system in a prescribed prior set, rather than relying only on generic persistent excitation. Its transferable asset is a finite-horizon, worst-case view of data informativeness: an input should make plausible dynamics distinguishable despite bounded disturbances. For neural world models and recurrent dynamics learners, this suggests actively selecting interventions or training trajectories by maximizing a robust excitation criterion over an ensemble of plausible models. The extracted set-membership inequality also provides a residual certificate for deciding whether observed trajectories are compatible with a candidate dynamics model.
Ideas from this paper
Unverified
Re-invented
2026
Replace random exploration or generic input noise with interventions selected to separate the predictions of all plausible dynamics models in an ensemble. At each data-collection step, choose the bounded input sequence whose simulated trajectories produce the largest worst-case information matrix, while penalizing unsafe or high-energy actions. This should make a recurrent or state-space world model identify dynamics with fewer real trajectories and reduce uncertainty on long-horizon rollouts.
Useful6/10
Difficulty6/10
Novelty6/10