THE NEURAL ARCHITECTURE OF REINFORCEMENT LEARNING IN PARTIALLY OBSERVABLE ENVIRONMENTS
Abstract
This project will investigate the neural basis of belief computation in the dopamine system and affiliated brain regions. Using the computational framework of temporal difference reinforcement learning in combination with rodent experiments, the project will test several mechanistic hypotheses about reinforcement learning under state uncertainty. First, we will use high-density multi-electrode arrays (Neuropixels) to characterize components of belief state computations in frontal cortex and hippocampus. Second, we will causally manipulate belief state inputs to the dopamine system by inactivating frontal, insular and hippocampal regions while simultaneously monitoring dopamine activity using fiber photometry. Third, we will parametrically manipulate belief states while monitoring dopamine to test quantitative predictions of the theory. This work will advance the Air Force’s mission by advancing our understanding of fundamental neurocomputational mechanisms underlying learning and decision making.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2021
- Source ID
- FA95502010413
Entities
People
- Samuel J Gershman
Organizations
- Air Force Office of Scientific Research
- President and Fellows of Harvard College
- United States Air Force