Control System Design Automation Using Reinforcement Learning
Title: “Control System Design Automation Using Reinforcement Learning”
Name: Hamid Mirzaei
Date: Tuesday, November 20, 2018
Time: 1:00 p.m.
Location: Donald Bren Hall 3011
Committee: Professor Tony Givargis (Chair), Professor Eli Bozorgzadeh, Professor Ian Harris
Conventional control theory has been used in many application domains with great success in the past decades. However, novel solutions are required to cope with the challenges arising from complex interaction of fast growing cyber and physical systems. Specifically, integration of classical control methods with Cyber-Physical System (CPS) design tools is a non-trivial task since those methods have been developed to be used by human expert and are not intended to be part of an automatic design tool.
On the other hand, the control problems in emerging Cyber-Physical Systems, such as intelligent transportation and autonomous driving, cannot be addressed by conventional control methods due to the high level of uncertainty, complex dynamic model requirements and operational and safety constraints.
In this dissertation, a holistic CPS design approach is proposed in which the control algorithm is incorporated as a building block in the design tool. The proposed approach facilitates the inclusion of physical variability into the design process and reduces the parameter space to be explored. This has been done by adding constraints imposed by the control algorithm.
Furthermore, Reinforcement Learning (RL) as a replacement for convection control solutions are studied in the emerging domain of intelligent transportation systems. Specifically, dynamic tolling assignments and autonomous intersection management are tackled by the state-of-the-art RL methods, namely, Trust Region Policy Optimization and Finite-Difference Gradient Descent. Additionally, Q-learning is used to improve the performance of an embedded controller using a novel formulation in which cyber-system actions, such as changing control sampling time, is combined with the physical action set of the RL agent. Using the proposed approach, it is shown that the power consumption and computational overhead of the embedded control can be improved.
Finally, to address the current lack of available physical benchmarks, an open physical environment benchmarking framework is introduced. In the proposed framework, various components of a physical environment are captured in a unified repository to enable researchers to define and share standard benchmarks that can be used to evaluate different reinforcement algorithms. They can also share the realized environments via the cloud to enable other groups perform experiments on the actual physical environments instead of currently available simulation-based environments.