Gain Scheduled PI controller design using Multi-Objective Reinforcement Learning

作     者：Kranthi Kumar P Ketan P Detroja 

作者机构：Department of Artificial Intelligence Indian Institute of Technology Hyderabad India 

出 版 物：《IFAC-PapersOnLine》 

年 卷 期：2024年第58卷第7期

页      面：132-137页

主　　题：Gain Scheduling Multi-Objective RL PI controller PPO algorithm Steam Turbine System Step response 

摘      要：Gain scheduling, a widely adopted control design approach, performs tasks by decomposing into sub-problems. It has found successful applications across diverse fields, including aerospace and industrial process control (Leith and Leithead (2000)). Reinforcement Learning (RL) has become increasingly significant in process control applications, thanks to the advancement of sophisticated algorithms and the rise in computational speeds. An attempt has been made to enhance the performance of a Proportional-Integral (PI) controller by integrating the benefits of both gain scheduling and RL. The manuscript focuses on designing a Gain Scheduled PI controller using Multi-Objective RL (GS-MORL). Proposing the training of multiple RL-based PI controllers with distinct objectives with customized reward functions to derive tuning parameters for each goal. The switching between these parameter sets is facilitated by a gain schedule variable. The controller parameters are adjusted based on the error between actual and desired responses. The aim is to enhance overall performance across various performance indices, including the integral of absolute error and time response specifications. The performance improvement from the proposed work has been demonstrated using two simulation case studies, one involving an underdamped second-order system and an integrating type steam turbine system. Comparative analysis against existing literature revealed superior performance across all indices.