Leveraging automatic strategy discovery to teach people how to select better projects

Lovis Heindrich, Falk Lieder

Keywords: metareasoning, automatic strategy discovery, decision-making, reinforcement learning, partially observable Markov decision process, project selection, metalevel Markov decision process, judge advisor systems, resource rationality

Abstract: Human decisions are often suboptimal due to limited cognitive resources and time constraints. Prior work has shown that errors in human decision-making can in part be avoided by leveraging artificial intelligence to automatically discover efficient decision strategies and teach them to people. So far, this line of research has been limited to simplified decision problems that are not directly related to the problems people face in the real world. Current methods are mainly limited by the computational difficulties of deriving efficient decision strategies for complex real-world problems through metareasoning. To bridge this gap, we model a real-world decision problem in which people have to choose which project to pursue, and develop a metareasoning method that enables us to discover and teach efficient decision strategies in this setting. Our main contributions are: formulating the metareasoning problem of deciding how to select a project, developing a metareasoning method that can automatically discover near-optimal project selection strategies, and developing an intelligent tutor that teaches people the discovered strategies. We test our strategy discovery method on a computational benchmark and experimentally evaluate its utility for improving human decision-making. In the benchmark, we demonstrate that our method outperforms PO-UCT while also being more computationally efficient. In the experiment, we taught the discovered planning strategies to people using an intelligent tutor. People who were trained by our tutor showed a significant improvement in their decision strategies compared to people who tried to discover good decision strategies on their own or practiced with an equivalent tutor that did not reveal the optimal strategy. Project selection is a very consequential high-stakes decision regularly faced by organizations, companies, and individuals. Our results indicate that our method can successfully improve human decision-making in naturalistic settings similar to the project selection decisions people face in the real-world. This is a first step towards applying strategy discovery methods to improve people's decisions in the real-world.

Submission Number: 13389