Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Sampling as a Resource-Rational Learning Mechanism

Muhammad Hashim Satti¹, Matthew R. Nassar², Sebastian Gluth¹, Radoslaw Martin Cichy³, Peter Dayan⁴, Rasmus Bruckner^3,1; ¹Universität Hamburg, ²Brown University, ³Freie Universität Berlin, ⁴Max-Planck Institute

Presenter: Muhammad Hashim Satti

Adaptive learning under uncertainty depends critically on how cognitive resources are allocated. While Bayesian models provide a normative account of belief updating, representing full probability distributions is computationally demanding. Sampling offers a tractable approximation, allowing resource-limited agents anytime belief estimation. However, due to the inherent randomness of sampling, fitting such models to human data is challenging. We show how to use simulation-based inference to fit a sampling-based model of a predictive inference task to lifespan behavioral data. The model accurately reproduced trial-by-trial learning trajectories and characteristic behavioral biases observed in humans. Importantly, it captured group differences: children and older adults relied on fewer samples than younger adults, leading to stronger perseveration and anchoring biases. These findings are consistent with the view that humans rely on sampling as a resource-rational learning mechanism.

Topic Area: Memory, Learning & Knowledge Structures