Computational, Cognitive, and Neural Models of Decision-making Biases

The question for the symposium is how best to understand biases in decision-making, going beyond traditional judgment and decision-making (JDM) accounts such as prospect theory to take a more modern reverse-engineering perspective bridging rational computational, algorithmic, and neural levels of explanation, and viewing decision-making under risk and uncertainty not just as a simple matter of evaluating lotteries but in the context of cognition more broadly, taking seriously learning, perception, motor control, memory, and action planning. The dominant normative approach to studying decision-making under risk is axiomatic expected utility theory, which argues that any agent obeying seemingly reasonable axioms of choice consistency can be modeled as maximizing the expected utility of its decisions. From decades of research that analyzes people’s choices between simple gambles in the lab, it is known that humans routinely violate these axioms. This has forced decision theorists to adopt descriptive models of choice that lack a normative rational in order to account for observed patterns of choice, the most prominent of which is Khaneman’s and Tversky’s prospect theory for one-shot decisions under risk with immediate outcomes and hyperbolic discounting for decisions involving delayed outcomes. There are several challenges not addressed by prospect theory and its variants. First, they are silent on the issue of the cognitive mechanisms that are actually responsible for human choice behavior. Second, they do not seem as a practical matter to scale to real-world decision problems, where the space of possible outcomes and actions is not sharply defined, the effects of actions are highly uncertain, and the explicit calculation of expected values is impractical. Third, they do not strongly constrain or give an underlying rationale for the probability weighting function, temporal discounting function, or utility function featured in prospect theory. Thus these models cannot explain why these functions’ estimated forms and parameters seem to be greatly affected by seemingly irrelevant factors of the task framing and setup, such as whether the outcome probabilities are presented numerically in tables or learned through experience and why the evaluation of individual gambles seems to be highly effected by the properties of other gambles in the choice set. More broadly, these theories fail to explain why in day-to-day life human decision-making seems to generally be highly robust and effective while sharply contrasting with normative predictions in the simple, stylized decision tasks commonly used in JDM experiments. This symposium brings together researchers who represent a variety of perspectives on ways cognitive science can inform our understanding of decision biases to address these challenges, with relevance at all three Marr levels of analysis. Malmaud and Tenenbaum, and Dayan both offer computational-level Bayesian accounts that explain decision-making biases as resulting from reasoning with priors that are adapted for real-world or evolutionary-relevant decision tasks. Malmaud and Tenenbaum explain choices in terms of advanced models from the AI planning literature and animal foraging theory. Dayan offers a neurobiological implementation of inference that spans the Marr levels. Other approaches relate to algorithms levels of the Marr hierarchy with links to lower and higher levels. Vul offers an algorithmic description of biases as resulting from cognitive limitations associated with reasoning using only a limited number of samples from a posterior over decision parameters. Maloney and Chater link high-level decision-making to known properties of perception and cognition, such as scale-invariance. Maloney gives a unifying account of the probability weighting function as arising from the same principles as perception of continuous quantities in psychophysics. Chater explores the origin of subjective utility and temporal discounting through connections to broader cognitive processes. One general idea that cuts across all these approaches is that human decision-making can be modeled in a unified way as the result of general cognitive principles that offer principled explanatory accounts of biases in decision-making, rather than via a series of descriptive utility-maximizing models that have undergone ad hod adjustments to account for a mélange of deviations from a narrow normative standard.