Autonomous and Resilient Control of Dynamical Systems Using Reinforcement Learning

Sunday, April 26, 2020 - 10:00am - 10:30am
Keller 3-180
Bahare Kiumarsi (Michigan State University)
Model-free and model-based reinforcement learning (RL) algorithms have
been widely and successfully employed to solve optimal control problems using only
measured data along the system trajectories. However, in an adversarial
environment, sensory information received by the RL agent to learn a control policy
could be compromised, resulting in learning a biased control policy with potentially
catastrophic outcomes. Therefore, it is desired to empower RL algorithms with
resiliency to survive in the presence of sophisticated threats. In this presentation, I
will talk about how to design a resilient model-free RL algorithm for systems under
attacks on sensors. I will show that to cancel out the effects of the compromised
sensors, a sparse least-squares optimization method can be employed to solve the
Bellman equation in the policy evaluation step of the RL algorithm.