Data-Driven Approaches to Model Predictive Control of Neural Systems

Achieving precise control of neural activity is a major focus of modern neuroscience. This challenge persists due to the diverse nonlinearities in neuronal dynamics and the largely unknown biophysical mechanisms governing large populations of neurons. Increasing our ability to control these systems would be of enormous benefit, both in terms of basic theory and clinical applications. Model Predictive Control (MPC) is a powerful control technique that uses mathematical models of a system to find optimal inputs to get a desired output. Using data-driven methods, models can be empirically fit to neural activity for use in MPC. This allows for control of a neural system even when there is very limited knowledge about the dynamics present. This dissertation provides a framework for how these data-driven models can be fit and utilized for MPC of neural systems without needing extensive a priori information.