A1062
Title: Practical performative policy learning with strategic agents
Authors: Bo Li - Tsinghua University (China) [presenting]
Abstract: The purpose is to study the performative policy learning problem. Existing approaches often rely on restrictive parametric assumptions: micro-level utility models in strategic classification and macro-level data distribution maps in performative prediction, severely limiting scalability and generalizability. This problem is approached as a complex causal inference task, relaxing parametric assumptions on both micro-level agent behavior and macro-level data distribution. Leveraging bounded rationality, a practical low-dimensional structure is uncovered in distribution shifts, and an effective mediator is constructed in the causal path from the deployed model to the shifted data. A gradient-based policy optimization algorithm is then proposed with a differentiable classifier serving as a substitute for the high-dimensional distribution map. The algorithm efficiently utilizes batch feedback and limited manipulation patterns. The approach achieves high sample efficiency compared to methods reliant on bandit feedback or zero-order optimization. Theoretical guarantees for algorithmic convergence are also provided. Extensive and challenging experiments on high-dimensional settings demonstrate the method's practical efficacy.
