Skip to content

Abstract Search

Primary Submission Category: Bayesian Causal Inference

Bayesian Safe Policy Learning with Chance Constraint Optimization: Application to Military Security Assessment in the Vietnam War

Authors: Zeyang Jia, Eli Ben-Michael, Kosuke Imai,

Presenting Author: Zeyang Jia*

Algorithmic recommendations have become an integral part of our society, being utilized in high-stake decision making settings. In those applications, it is essential to control the risk before putting data-driven policies into practice. A prominent frequentist approach assumes a model class for the conditional average treatment effect (CATE) and finds an optimal policy within a pre-specified policy class by maximizing the worst-case expected utility. However, when both model class and policy class are complex, the resulting optimization and uncertainty quantification are often intractable. We propose a Bayesian safe policy learning method that controls the risk via chance constraint optimization while decoupling the estimation and optimization steps. We first estimate the CATE with a Bayesian nonparametric model, then derive a safe policy by maximizing the posterior expected utility while limiting the posterior probability that the new policy negatively affects a group of individuals on average. We also show that the chance constraint optimization can be efficiently solved as a constrained linear programming problem. Our motivating application is the military security assessment policy used during the Vietnam War. By adopting ideas in graph theory, we solve the optimization over complex decision tables, which are widely used in public policy. We find that economic and social development factors should be given greater weights.