ActiGrid benchmark finds learned evidence policies add no gain over simple stopping rules
A new arXiv paper introduces ActiGrid, a benchmark for deciding when power system diagnostic models should request more telemetry, and reports that simple rules match a learned policy.
In plain words
Power grid operators use software that diagnoses faults and cyberattacks from sensor data. Usually such software reads every available data source for every event. The ActiGrid benchmark tests whether a reinforcement learning agent that chooses which sources to inspect works better than simpler strategies, and finds that it does not.
What was announced
The framework pairs a classifier trained to tolerate missing data modalities with a Double Deep Q-Network that either requests another telemetry source or stops and issues a diagnosis; it never operates grid actuators. On the public TU Delft digital substation and MSU/ORNL power system attack benchmarks, the authors compared the learned policy against fixed subsets and orderings, calibrated entropy stopping, a greedy value-of-information rule, and per-subset tree ensembles.
The learned policy did adapt to what it observed, but the paper reports it was statistically indistinguishable from entropy stopping on TU Delft and Pareto dominated by simple controls on MSU/ORNL. Compact subset classifiers outperformed it on both benchmarks while reading fewer sources. Entropy stopping over those subset classifiers reached the highest mean Macro-F1 on TU Delft, 0.928 while reading 57% of sources, but brought no saving over a static three-source subset on MSU/ORNL, where no method exceeded 0.45 Macro-F1. The authors conclude that selective acquisition can pay off, but not through a learned policy on these benchmarks.
What it does not show
- The evidence is the paper's abstract only; full methodology, experimental setup and reproducibility details are not available for verification.
- Results are limited to the TU Delft and MSU/ORNL benchmarks and may not transfer to other power system settings.
- The reported Macro-F1 figures are the authors' own measurements and have not been independently confirmed.
For security and IT teams
- Scope
- ActiGrid addresses diagnosis only; the DQN policy never operates grid actuators, so the benchmark does not cover automated control actions.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.