Why less can be more : a Bayesian framework for heuristics

When making decisions under uncertainty, one common view is that people rely on simple heuristics that deliberately ignore information. One of the greatest puzzles in cognitive science concerns why heuristics can sometimes outperform full-information models, such as linear regression, which make full use of the available information. In this thesis, I will contribute the novel idea that heuristics can be thought of as embodying extreme Bayesian priors. Thereby, an explanation for less-is-more is that the heuristics’ relative simplicity and inflexibility amounts to a strong inductive bias, that is suitable for some learning and decision problems. I will formalize this idea by introducing Bayesian models within which heuristics are an extreme case along a continuum of model flexibility defined by the strength and nature of the prior. Crucially, the Bayesian models include heuristics at one of the Bayesian prior strength and classic full-information models at the other end of the Bayesian prior. This allows for a comparative test between the intermediate models along the continuum and the extremes of heuristics and full regression model. Indeed, I will show that intermediate models perform best across simulations, suggesting that down-weighting information is preferable to entirely ignoring it. These results refute an absolute version of less-is-more, demonstrating that heuristics will usually be outperformed by a model that takes into account the full information but weighs it appropriately. Thereby, the thesis provides a novel explanation for less-is-more: Heuristics work well because they embody a Bayesian prior that approximates the optimal prior. While the main contribution is formal, the final Chapter will explore whether less is more at the psychological level, and finds that people do not use heuristics, but rely on the full information instead. A consistent perspective will emerge throughout the whole thesis, which is that less is not more.