Category Learning by Clustering with Extension to Dynamic Environments

摘要

This project focuses on how humans master new categories by learning from examples with extension to dynamic environments. Decision making tends to take place in dynamic environments in which successive decisions are contingent on one another, and in which the rewards associated with actions can be delayed, yet most tasks that have been studied in the laboratory are broken up into brief, independent trials (e.g., classification of a stimulus) in which responses are determined only by the immediate context and have no bearing on future states of the task environment. Thus, this project narrows the gap between the range of mental processes typically addressed by cognitive scientists and the mental processes that underlie performance in Air Force relevant activities. We find that people's performance profiles are generally consistent with modern reinforcement learning models. For example, including perceptual information that disambiguates a person's current state within a task improves performance. Additionally, consistent with model-based predictions, people appear to hill climb on reward gradient, as opposed to globally optimize performance and show other suboptimal behavior, such as poorer performance under certain circumstance when given more information about response options.