Bayesian Treaty Monitoring: Preliminary Report
CALIFORNIA UNIV BERKELEY
Pagination or Media Count:
Our project has initiated and will develop and evaluate a new Bayesian approach for nuclear test monitoring. We anticipate that the new approach will yield substantially lower detection thresholds, possibly approaching a theoretical lower bound that we hope to establish. We will also develop new techniques to implement such monitoring capabilities within a general-purpose Bayesian modeling and inference system that may eventually support a wide range of information-system needs for arms treaties. In ongoing work that is moving towards possible deployment, we have completed a prototype seismic monitoring system based on a generative, vertically integrated statistical model linking hypothesized events to detections extracted from raw signal data by commonly used algorithms. On test data sets of naturally occurring events curated by human experts, our system exhibits roughly 60 fewer detection failures than the currently deployed automated system, SEL3, that forms part of the International Monitoring System. The current phase of the project moves away from hard-threshold detections altogether. Instead, the generative model spans the full range from events to measured signal properties. Given the observed signal traces, the statistical inference algorithm attempts to maximize a whole-network statistical measure of the likelihood that an event - or collection of events - has occurred. Specialized techniques such as waveform matching and double differencing are realized within our framework as special cases of probabilistic inference our initial experiments using 2D simulated data indicate that a full Bayesian analysis can provide more accurate absolute and relative locations than double differencing, while simultaneously estimating the velocity structure of the observed region.
- Operations Research