Rationality of Heuristics in a Changing Environment
Allgemeine, Kognitive und Mathematische Psychologie
Zusammenfassung der Projektergebnisse
This project was part of a Priority Programme that provided an interdisciplinary framework for research on human rationality, including approaches based on logic, probability, and heuristics. The current project explored the rationality of simple decision heuristics, building on the view that the rationality of a decision depends on the circumstances under which it takes place, known as ecological rationality. Two questions were of primary interest: What type of environmental structures enable decision heuristics to be effective? And, how prevalent are these structures in our world? The project was motivated by two sets of findings in the literature. The first include empirical results showing that simple heuristics compete well with complex statistical methods on natural data sets. The second include theoretical explorations of the conditions under which simple heuristics perform well. The most theoretically well-developed of these are certain dominance relationships among decision alternatives and noncompensatoriness among decision cues. The most important finding of this project is that these theoretical conditions, on dominance and noncompensatoriness, have high practical relevance. On a large, diverse collection of natural data sets, these conditions were found to be prevalent to such a high degree that decisions guided by these conditions approached, and occasionally exceeded, the linear decision model in predictive accuracy while using substantially less information and computation. This finding is unexpected and changes the narrative on why decision heuristics perform well. Prediction error has two components, bias and variance. The bias component comes from how well our models can represent the world, while the variance component comes from sensitivity of our models to fluctuations in the training data. The predominant view in the literature was that simple heuristics, while incurring high bias, can compete with more complex statistical decision methods by keeping variance low. The results of this project show that simple decision heuristics need not incur high bias either, at least not compared to linear models. Prior to the work conducted in this project, dominance and nocompensatoriness, which are conditions that allow heuristics to incur low bias, were not well known outside of a small community of theoretical researchers. In particular, they did not have a presence in overview articles that communicate important findings on decision heuristics to a varied, interdisciplinary audience. This is no longer the case. The most recent overview article discusses these results extensively, along with the results obtained in this project. A second important finding of this project is on single-cue, a heuristic that uses a single piece of information. This heuristics is not entirely new but was neglected in the literature to such a degree that there were no empirical results on its performance. In an analysis conducted during this project, single-cue performed remarkably well in a large, diverse collection of natural data sets. Its performance, on average, was close to that of logistic regression, almost as good as take-the-best, and better than tallying. An additional result was that an appropriate choice of an ordinal attribute yielded, on average, higher accuracy than various ways of integrating a number of binary attributes. One implication of this result is that ordering the cues correctly for a lexicographic heuristic may not be nearly as important as identifying the best cue, or one close to it in accuracy. Another implication is that single-cue deserves attention as a candidate model for how people make decisions.
Projektbezogene Publikationen (Auswahl)
-
(2013). Linear decision rule as aspiration for simple decision heuristics. In C.J.C. Burges, L. Bottou, M. Welling, Z. Ghahramani, & K.Q. Weinberger (Eds.), Advances in Neural Information Processing Systems 26. Red Hook, NY: Curran Associates
Şimşek, Ö.
