Reinforcement learning task design using reinforcement learning models

Abstract

Recent studies have suggested various computational models of human reinforcement learning , but task design for these studies is often difficult, due to the complexity of human reinforcement learning and high individual variability. Furthermore, unlike animal studies, human studies lack brain stimulation methodologies, thus impeding the investigation of the computational mechanisms underlying reinforcement learning. In this thesis, I propose an algorithm, called a task controller , for improving learned task design, using interaction with a standard computational model of human reinforcement learning. The model-based fMRI analysis, involving 21 human subjects, showed that the learned task design has the potential to guide the neural activity of the targeted brain regions.