Improving Human Sequential Decision Making with Reinforcement Learning.

This article focuses on a novel reinforcement learning algorithm designed to infer simple, interpretable tips that improve human performance in sequential decision-making tasks. Using a virtual kitchen management game modeled as a job shop scheduling problem, the study conducts large-scale randomized controlled experiments with Amazon Mechanical Turk participants under two scenarios: a fully staffed kitchen and a disrupted understaffed kitchen. The algorithm identifies tips by analyzing discrepancies between human and optimal policies, prioritizing actions that most impact long-term performance, and conveys these as concise if-then rules. Experimental results show that tips generated by the algorithm significantly enhance participant performance and accelerate learning compared to control groups, human-suggested tips, and baseline algorithmic tips based on descriptive statistics. The study also finds that participants gradually increase compliance with tips over time and combine them with personal experience to discover additional effective strategies, highlighting the importance of both tip quality and human operationalization in improving sequential decision making.