Models and methods for the computational study of human decision-making using choice-response time data

Abstract

Decision-making plays a critical role in our everyday lives. Some choices are simple, like choosing which way to turn at an intersection, whiles others are complex, like deciding which house to purchase. Decision-making is also a tool for understanding human cognition at large, and when connected to computational models of the brain, it allows us to investigate mechanisms underlying perception, information processing, planning, deliberation, context dependent strategy, and cognitive function in general. In this work, we specifically study cognition through the lens of evidence accumulation models (EAMs) applied to two-alternative forced choice experiments. To do so, we first develop new methods for connecting EAMs to experimental data. In Chapter 2, we introduce the Python package PyBEAM (Bayesian Evidence Accumulation Models). PyBEAM is a general probabilistic framework for predicting the choice and response time distributions of a wide class of two-alternative evidence accumulation models using Bayesian inference. In chapter 4, we develop an R port of this package called REAM which expands upon the methods developed in PyBEAM. Additionally, it adds a larger library of precoded EAMs, including new solutions for dual-process EAMs which has previously been mathematically intractable. Overall, PyBEAM and REAM substantially expand the class of EAMs accessible to researchers, emphasizing computational speed, model breadth, and ease of use. In the second part of this work, we apply the tools of PyBEAM and REAM to closely examine the structure and interpretability of EAMs. In Chapter 3, we perform a parameter recovery study on a variety of common binary choice EAMs with the goal of identifying the specific challenges each face when being fit to data. We further provide recommendations to researchers about how best to improve their parameter recoverability. In Chapter 5, we develop a scientific machine learning approach to discover the unknown shape of cognitive models from decision-making data. To do so, we generate a library of plausible models, then use sparse Bayesian inference to identify which models fit the data best. We demonstrate that this approach not only fits the data well, but allows us to recover model functions outside of our model library set.