Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
Improving Human Sequential Decision Making with Reinforcement Learning
Citation. Bastani, H., Bastani, O., & Sinchaisri, W. P. (2026). Improving human sequential decision making with reinforcement learning. Management Science, 72(1), 733–755. https://doi.org/10.1287/mnsc.2022.02455
Summary. Bastani, Bastani, and Sinchaisri develop an interpretable reinforcement-learning method that extracts concise action tips from behavioral trace data to help people improve sequential decisions. In randomized virtual-kitchen experiments, participants receiving the algorithm’s tips outperformed intuitive-baseline tip conditions. The authors also find that participants combined the tips with their own experience rather than simply following them blindly, framing design choices for human-AI decision-support interfaces.
Key claims
- The method selects an interpretable tip that targets consequential differences between a worker’s actions and an optimal policy, rather than merely reporting a generic recommendation.
- Across randomized experiments in a virtual-kitchen task, algorithm-generated tips significantly improved human performance relative to intuitive baselines.
- Participants did not simply defer to the system’s recommendations; they used tips together with their own experience to discover additional performance-improving strategies.
Evidence & limitations
- The article reports randomized controlled experiments in which participants managed a virtual kitchen, giving causal evidence for the intervention within that simulated sequential-decision setting.
- The task, tips, and performance measures are context-specific; transfer to consequential workplace decisions, other domains, or longer-term learning remains to be established.