Real-Time Cognitive Modeling Using Simulation-Inversion Networks

Cognitive models allow a user to extract information about latent processes that generatebehavior, but they are typically applied long after data have been collected. In the real world,behavior unfolds in continuous time, and quick inferences about these latent processes areinvaluable in social tasks. In this paper, we use machine learning methods to enable real-timecognitive modeling and apply the approach to a novel cognitive model to explain continuousdyadic goal-pursuit behavior in the Team Spirit task. We then train a transformer-basedsimulation inversion network using simulated data and estimate cognitive model parameters forempirical data that has been length-censored to assess how predictions change over time. Resultssuggest that simulation-inversion networks can enable fast and accurate model fitting forsequences of any length and yield precise predictions from a set of interpretable parameters thatdescribe an individual’s cognitive processes. Real-time cognitive model fitting opens the door tonew applications where cognitive models can go beyond a post-hoc tool to explain past behaviorand instead be integrated directly into social agents that continuously update the cognitive profilesof those around them.