A1791
Title: Orthogonal dynamic programming for sequential treatment allocation
Authors: Heejun Lee - Brown University (United States) [presenting]
Abstract: Sequential treatment allocation with weak treatment effects and rewards confounded by time-varying nuisance parameters is studied. In standard Bayesian frameworks, optimal planning via dynamic programming is hindered by the curse of dimensionality, as the state space grows with the history of observations. A local asymptotic framework overcomes this computational barrier by collapsing the history into low-dimensional sufficient statistics. However, nuisance variation complicates this framework: accommodating nuisance parameters increases computational complexity and requires specifying priors that may be prone to misspecification, while ignoring them introduces bias. To address this, orthogonal planning is proposed, a robust dynamic programming framework based on orthogonality.
