Proceedings of the Annual Meeting of the Cognitive Science Society

Finding computational models that give rise to human-like heuristic decision-making is an open challenge in cognitive science. Meta-learning has been proposed as a promising computational framework for obtaining algorithmic models of cognition by learning from environmental interactions rather than designing normative models. Recent work by Binz et al. (2022) demonstrates that bounded meta-learned inference (BMI) discovers human-like heuristic decision strategies in paired comparison tasks. However, the reasons and mechanisms underlying this discovery remain unclear. Here, we extend research on the BMI framework to better understand its behavior and reverse-engineer its computational principles. We first systematically varied model parameters that influence the network's representational capacity, finding that BMI robustly discovers context-appropriate heuristics across configurations. The model consistently uses single-cue strategies when feature rankings are known and equal weighting strategies when feature directions are known, even when network capacity is strongly reduced compared to the original work. We then reverse-engineer the computational principles underlying this behavior using hierarchical Bayesian modeling. Our analyses reveal that regularization, not network size, drives heuristic discovery. Bayesian models with heuristic priors provide the best fit for regularized network behavior when environmental cues are available. Additionally, we discovered a stronger inductive bias toward equal weighting over single-cue strategies. Thus, we conclude that inductive biases, in the form of priors, have a greater impact on the emergence of heuristics than resource rationality, in the form of networks' representational capacity.