Description

Humans frequently need to predict the future state of partially observable systems, such as tracking objects that temporarily disappear from view. However, the ability to make accurate predictions is limited by uncertainty that accumulates over time, giving rise to a prediction horizon beyond which forecasts become unreliable. This project aims to develop computational models that characterize human prediction behavior in such settings. In collaboration with the Biological Psychology lab led by Prof. Ricarda Schubotz (University of Münster), we will study a grid-world paradigm in which a target (e.g., a “mole”) moves according to structured but partially hidden dynamics. Participants observe the target intermittently and are asked to predict its future location after varying delays. On the computational side, the project will simulate the behavioural task using probabilistic models of latent state inference, such as Hidden Markov Models or related state-space models. These models maintain a belief distribution over the hidden state and propagate it forward in time to generate predictions at different horizons. A key focus will be on how uncertainty evolves during periods without observation and how this limits predictive performance. Different decision strategies will be evaluated, including maximum a posteriori prediction and probabilistic choice rules. The models will be used to generate quantitative predictions for human behavior, such as accuracy as a function of prediction horizon and sensitivity to the statistical structure of the environment. The project will provide insight into the computational principles underlying human predictive inference and establish a link between probabilistic modeling and behavioral data.