Basic research questions

The main research questions addressed in this topic area, in the context of Telluride are

1. Can we bring in practical work on advanced control engineering to Telluride to better bridge the understanding gap from technology to biology? In particular, the hierarchy present in the human system suggests a natural structure for combining learning and control--especially the use of central pattern generators in the spinal cord that play an essential role in locomotion. Biological systems seem to be able to quickly assemble prototypes for control from prior primitives, although mastery of any particular skill occurs by reinforcement learning that can take hundreds of thousands of trials. Yet this learning typically occurs within the spinal cord (in the context of locomotion), which is similar to the proposed approach: MPC will play the role of the nominal patterns present in the spinal cord which will be modulated in a safe manner through episodic learning.
2. Can we explore the use of deep learning technology to bring It more into control? In particular, can we understand how to rapidly learn from small amount of training data sufficient generality and stability to address real control problems? This learning strategy will require an understanding of the coupling between the learning and control feedback loops, and how they interact. Special attention will have to be given to learning with limited data in a way that preserves the beneficial properties and behaviors of the system. Thus, we will need to explore the exploration vs. exploitation tradeoff in systems that are not inherently stable.