Teaching people how to make their own wise decisions is one of the most challenging goals of education. This project leverages artificial intelligence to address two key challenges in teaching decision-making skills: discovering smart decision strategies and teaching them at scale.
To discover smart decision strategies, we have developed a mathematical theory of optimal decision-making with limited time and bounded cognitive resources [ ] and cognitively inspired reinforcement learning methods for computing such strategies [ ][ ]. To make it possible to teach the resulting automatically discovered strategies at scale, we develop called intelligent cognitive tutors. Our intelligent cognitive tutors help people internalize those strategies by having them practice decision-making with optimal feedback [ ][ ], demonstrate the optimal strategies to them [ ], and present them with automatically generated natural-language descriptions of those strategies [ ]. Our recent work has made our methods significantly more scalable to larger and more complex decision problems [ ], more robust to errors in the model of the real world [ ], and more interpretable [ ]. We are currently extending this approach to partially observable environments. Moving forward, we will apply our approach to improve human performance in real-world problems, such as in hiring decisions.
Testing our intelligent cognitive tutors in large-scale online experiments demonstrated that our intelligent cognitive tutors can significantly improve the strategies and outcomes of human decision-making not only in the trained task but also in more complex decisions and in similar tasks in other domains [ ][ ]. Unfortunately, we also found that people rarely apply the taught strategies to tasks that look very different from the trained task [ ].
Our work has primarily focussed on helping people overcome short-sighted biases in decision-making by discovering and teaching far-sighted planning strategies [ ][ ]. However, our methods are very general. We have already successfully applied them to other types of sequential decision problems and the problem of choosing between multiple alternatives based on their attributes [ ]. In addition, we have developed a desktop app that helps people train their executive functions by giving them feedback on how well they succeed at staying focused on their goals as they work on their computers [ ].