Conformal Policy Selection
A framework for selective deployment of AI policies based on conformal inference and false discovery rate control. It provides statistical safety guarantees for individual-level decisions.
Timeline (last 90 days)
Conformal Policy Selection reformulates selective deployment as a multiple testing problem, using false discovery rate (FDR) control to provide statistical guarantees for individual-level AI deployment decisions.
The technical approach of Conformal Policy Selection consists of two components: weighted conformal p-values and a two-step selection and pruning procedure.
The two-step selection and pruning procedure controls FDR in finite samples and establishes a doubly robust asymptotic guarantee: FDR control still holds as long as either the outcome model or the propensity model is correctly specified.
The paper covers four model misspecification scenarios in synthetic experiments and validates the theoretical guarantees.
The effectiveness of the framework depends on the outcome threshold set by the user, and the choice of threshold itself involves domain judgment.
All Facts (5)
Conformal Policy Selection reformulates selective deployment as a multiple testing problem, using false discovery rate (FDR) control to provide statistical guarantees for individual-level AI deployment decisions.
50%UnverifiedThe technical approach of Conformal Policy Selection consists of two components: weighted conformal p-values and a two-step selection and pruning procedure.
50%UnverifiedThe two-step selection and pruning procedure controls FDR in finite samples and establishes a doubly robust asymptotic guarantee: FDR control still holds as long as either the outcome model or the propensity model is correctly specified.
50%UnverifiedThe paper covers four model misspecification scenarios in synthetic experiments and validates the theoretical guarantees.
50%UnverifiedThe effectiveness of the framework depends on the outcome threshold set by the user, and the choice of threshold itself involves domain judgment.
50%