Decisions as Ill-Posed Problems: A Scoping Review of Regularization Methods in Decision Science

Abstract

Real-life decisions often require individuals tomake inferences based upon a large number of possible predictive features, but only limited observations. These so called ill-posed problems can be translated into well-structured ones using regularizationmethods.While regularization methods using penalty terms have gained popularity in machine learning, decision science has increasingly adopted Bayesian approaches to regularization, such as hierarchical Bayesian estimation. Yet, theories of human decision making appear to have less often embraced the original framing of regularization methods as “solutions to ill-posed decision problems.” In a scoping review, I investigate how regularization methods have influenced decision science in applications, contributed to methodological advances, and informed theoretical debates. Within the 117 reviewed articles, most studies applied regularization methods as a tool to deal with heterogeneity and nested data structures, to select the most relevant problem features, or to improve interpretability. Methodological advances through regularization have rendered cognitive choice models testable on high-dimensional data or allowed to map descriptive models onto normative models of human decision making. Theoretical innovations were, however, confined to modeling constraints in information processing, re-integrating heuristic with rational choicemodels, and debating the role of prior beliefs. Looking ahead, I discuss how future applications in decision may harness regularization techniques to communicate insights from the analysis of decision policies, to improve forecasting tools, and to innovate theoretical approaches in decision science.