To make automatic strategy discovery robust to potential errors in the description of the decision environment, we developed a computational method that performs Bayesian inference on what kind of environment could have given rise to the provided description. Automatic strategy discovery then optimizes the expected performance of the strategy across all possible decision environments, weighted by their respective posterior probabilities. We found that this method makes it possible to discover and teach decision strategies that improve human performance even when the model of the environment is somewhat inaccurate [].

To make good decisions in a limited time, people need clever heuristics that exploit the structure of the environment. Given a perfect model of the environment, optimal heuristics can be discovered automatically using the principle of resource-rationality. However, when discovering strategies for real-world problems, we have to depend on people to provide us with a description of the environment. Since people are known to be fallible to cognitive biases, the perfect model assumption may not hold true.

In this project, we, therefore, aim to answer the following question:

How can we discover clever heuristics from biased descriptions of the environment?

As a first step to that end, we formed models of cognitive biases and used Bayesian Inference to obtain candidate true environments (Kemtur et al. 2020). We then sought to discover strategies that work well in expectation over these candidates. Experimental studies show that our robust cognitive tutor significantly improved human decision-making when the model was so biased that conventional cognitive tutors were no longer effective. Motivated by these results, we now seek to scale up our methods and apply them to more real-world scenarios.