At Parsa Research Laboratory (PRL) we are focused on developing full stack novel brain-inspired paradigms that enable accurate, energy efficient, fast, and reliable intelligence at the edge. These span from algorithm-hardware codesign to physics-informed neuromorphic computing, distributed learning, and safe lifelong learning.
Title: Learning Neuromorphic Physics-Informed Stochastic Regions of Attraction through Bayesian Optimization (single-PI)
Sponsor: Intel Neuromorphic Research Community (INRC)
Collaborator: Dr. Joe Hays, Naval Research Laboratory
Goal: The ultimate goal of this project is to enable agents, who are learning to control themselves, to have an inherent sense of their own stability, and thus enable them to learn and expand their capabilities safely. This work aims to use Bayesian optimization (BO) to learn a distribution of Lyapunov regions of attraction (ROA). ROA defines the boundary between stability and instability. Once learned offline, the policy will be able to discriminate which specific parameter set defines the specific system under control from the state/action data and therefore refine the ROA for the specific agent.
Impact: Endowing agents with a sense of their own stability through an estimate of their own ROA (pulled from an offline characterized stochastic ROA distribution) is a strong step forward in our community’s pursuit of having safe lifelong motor learning. The ROA will enable the agent to maximize the exploration needed for learning while remaining safe.