計算論的アプローチにおける不確実性下の意思決定・学習課題：スコーピングレビューによる選定指針

抄録

In psychology and psychiatry, decision-making and learning under uncertainty are crucial for understanding the mechanisms of human adaptive behavior. Cognitive tasks are fundamental for computational research in this field, as they shape the nature and quality of behavioral data for computational modeling. Despite a variety of cognitive tasks available, a systematic selection and design guide for researchers is lacking. This study aims to provide a comprehensive selection guide through a scoping review of 229 articles published since 2020. We classified the cognitive tasks based on a novel framework consisting of three axes: (1) target cognitive processes, (2) estimation difficulty (the degree of uncertainty in estimating outcomes), and (3) applied computational models. Tasks with no estimation difficulty (e.g., decision-making under risk) that primarily target static preferences rather than learning processes are modeled using valuation/preference or sequential sampling models. Low estimation difficulty tasks (e.g., basic bandit task) focus on experience-based value updates in stable environments, predominantly using reinforcement learning models. In high estimation difficulty tasks, modeling approaches diverged based on the source of uncertainty. Tasks with environmental volatility or rule instability were analyzed with Bayesian inference models to capture higher-order estimation (e.g., Hierarchical Gaussian Filter). By contrast, tasks with ambiguity (hidden information or feedback) were typically analyzed with valuation/preference models. We propose this framework as a comprehensive guide for researchers to align their research interests, task designs, and modeling strategies.