Determine whether targeted rejection auditing reduces the cost of value recovery

Determine whether targeting audits at rejected actions whose certified lower bounds lie near zero can recover beneficial actions more cheaply than uniformly auditing rejected actions.

Background

The paper audits observationally rejected actions uniformly and finds that recovering beneficial actions requires substantially more experiments than eliminating harmful executions. It notes that a targeted strategy could prioritize rejections most likely to reflect graph misspecification, such as those with certified bounds close to zero.

The effectiveness and cost savings of such targeted rejection auditing are not evaluated in the paper, leaving unresolved whether it can recover value more efficiently than uniform sampling.