My research asks how people make decisions under uncertainty, both when encountering a choice for the first time and when learning from repeated experience. The goal is not only to explain behavioral regularities, but also to predict behavior accurately enough to inform the design of environments, incentives, and policy.
A recurring idea in my work is that people tend to use decision strategies that worked well in similar past situations. Faced with a new decision, they first identify what kind of situation it resembles and then rely more heavily on strategies that have worked well in that class of situations. I use behavioral research to identify the features people may use to classify situations and the strategies they may consider, and computational models to examine when these mechanisms can explain and predict choice.
Learning from experience and rare events
Much of my research studies how people learn from repeated experience, particularly when environments are uncertain and when important outcomes are rare.
My early work showed that several behavioral regularities commonly attributed to simple reliance on recent outcomes can instead arise from similarity-based learning. This account predicts sensitivity to sequential patterns and a novel wavy recency effect of rare events, and can be adaptive in complex changing environments even when it produces apparently suboptimal behavior in simpler settings.
I have since examined the implications of these learning processes across decisions with partial or biased feedback, social interactions, strategic games, and other environments involving rare events.
From understanding behavior to designing it
A major reason I care about prediction is that useful behavioral theories should help us design environments that reliably influence choice.
In one line of work, I used a model of learning from experience (CATIE) to design the winning reward schedule in the Choice Engineering Competition, outperforming submissions based on standard reinforcement-learning models and qualitative choice architectures.
More broadly, I study how incentives, enforcement, and feedback shape repeated behavior. This work suggests, for example, that frequent enforcement can be more effective than severe punishment, and that small, frequent reinforcements can sometimes outperform much larger but infrequent ones in sustaining behavior.
Predictive and interpretable models of choice
Another major part of my research develops models that are both psychologically interpretable and predictively accurate.
I co-developed BEAST, a behavioral model that captures a broad range of regularities in decisions under risk and uncertainty. More recent work has focused on making such models easier to estimate and scale, including differentiable implementations and hybrid approaches that use behavioral models to construct features for machine learning.
In large-scale tests, these hybrid models, such as BEAST-GB, perform best when machine learning receives both objective information about the decision environment and interpretable behavioral features representing plausible choice strategies. This combination improves prediction while also revealing where different behavioral strategies are more likely to be used.
Current directions
Current projects extend these ideas to new settings, including choices between human and algorithmic advisors, learning under ambiguity, strategic interaction, and the behavior of large language models in decisions-from-experience tasks.
Across these projects, the broader aim remains the same: to understand the processes that generate human choice well enough to predict behavior outside the specific setting in which a theory was developed—and ultimately to use those predictions to design better decision environments.