Human Decision-Making with AI Assistance under Correlated Features

The paper presents a new approach to AI-assisted human decision-making under correlated features, demonstrating that traditional stationary policies are suboptimal in this context. It introduces an explore-then-commit strategy where the AI initially recommends diverse tests to enable human learning before committing to a specific set, with exploration length influenced by feature correlation. The study proves the NP-hardness of computing the optimal policy and offers a dynamic programming algorithm for finite horizons, along with an approximation for shorter planning, highlighting the practical implications of feature correlation on decision-making quality and learning efficiency.