Title: The Advantage of Human-AI Synergy: Super Reinforcement Learning Amidst Confounding Factors

Abstract:

As artificial intelligence permeates various sectors of society, a critical objective has emerged: developing robust strategies to merge human and AI capabilities, thereby capitalizing on their distinct advantages while minimizing potential hazards. This study presents a new framework known as "super policy learning," which harnesses the dynamics of human-AI interaction to enhance data-driven sequential decision-making. The core mechanism involves utilizing observed actions—whether executed by humans or algorithms—as inputs to construct a superior oracle for the decision-maker.

In scenarios characterized by unmeasured confounding, historical actions provide crucial clues regarding hidden variables. By incorporating these insights into the policy search process through a novel and valid methodology, the proposed super policy learning approach generates a "super-policy." This resulting policy is theoretically proven to surpass both the conventional optimal policy and the behavior policy (such as the actions of previous agents). We term this enhanced performance the "blessing" derived from human-AI collaboration.

To tackle the challenge of unmeasured confounding when deriving super-policies from batch data, we establish several nonparametric and causal identification results within the context of proximal causal inference. Leveraging these innovative identification findings, we design multiple algorithms for super-policy learning and rigorously analyze their theoretical attributes, including finite-sample regret guarantees. The efficacy of our proposed method is demonstrated through comprehensive simulations and applications in real-world settings.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC