Fitting Cognitive Simulators to People: Distributed and Hierarchical Parameter Estimation in PsyNeuLink

This poster describes extensions to PsyNeuLink, an open-source Python library for computational models of cognition, that enable distributed model fitting, hierarchical parameter estimation and learned likelihoods. Fitting these models to behavioral data can require thousands of repeated simulations for each participant. We developed a reusable Dask backend that distributes this work across local processes or Slurm clusters while preserving researchers’ existing models and fitting workflows. Workers construct and compile models locally, then reuse them across likelihood evaluations, participant fits and training simulations. The poster presents the execution architecture, the setup required from researchers, and results for distributed scaling, hierarchical parameter estimation and learned likelihoods. Hierarchical fitting shares information across participants while retaining individual differences. Learned likelihoods approximate simulator likelihoods with a trained neural network, reducing the cost of subsequent evaluations. This work was carried out through a research software engineering fellowship at Princeton University.