Speaker:    

Yibei Li博士后, 新加坡南洋理工大学
Inviter:      
Title:
    Cost Learning in Autonomous Dynamical Systems: An Inverse Optimal Control Approach
Language:     English
Time & Venue:
    2022.11.29 16:00-16:30 腾讯会议：566894742
Abstract:
    

Optimal control, reinforcement learning and other optimization-based methods have been widely applied to control synthesis for real systems in physics, biology, and robotics. However, in many complicated tasks especially with spatially or temporally dynamical contexts, defining a cost function that can be optimized effectively and encodes the correct task is quite challenging in practice. It is always much easier to learn and imitate an optimal behavior from natural phenomena or human demonstrations. Driven by a revolutionary increase in the availability of data, a variety of mathematical methodologies have been developed to model the optimization mechanisms underlying such intellectual behaviors. In this talk, some of our results on the inverse optimal control problem will be introduced. Given a dynamical system and observations of its optimal policies for a specific task, the goal is to recover the underlying optimization criterion based on which the optimal controller is generated. Feasibility and uniqueness of corresponding linear quadratic cost functions are studied and the space of equivalent cost functions are characterized analytically. Other issues also include approximate cost learning for sub-optimal demonstrations and scenarios with partial state observations.
--
FROM 182.51.86.*