Model-Free Learning Control of Chemical Processes

Learning is the nature for human being. For example, a school-student learns a subject by doing exercise and home-work. Then, a school-teacher grades the school-student’s works. From this student and teacher interaction, the ability of the student mastering the subject is a feedback that the previous teaching method is successful or failure. As a result, the teacher will change the teaching method to improve the student ability for mastering the subject. This is a picture that the reinforcement learning (RL) agent learns the environment. Process control mainly focuses on controlling variable such as pressure, level, flow, temperature, pH, level in the process industries. However, the methodologies and principles are the same as in all control fields. The early successful application control strategy in process control is in evolution of the PID controller and Ziegler-Nichols tuning method (Ziegler and Nichols, 1942). Till nowadays, 95% of the controllers implemented in the process industries are PID-type (Chidambaram and See, 2002). However, as (i) the industrial demands (ii) the computational capabilities of controllers and (iii) complexity of systems under control increase, so the challenge is to implement advanced control algorithms. There have been commercial successes of the intelligent control methods, but the dominating controller in process industries is still by far the PID-controller (Chidambaram and See, 2002). This stands to the fact that a simple and general purpose automatic controller (for example PID) is demanded in process industries. Therefore, designing advanced controllers are to address the industrial user demand. This is the reason that a learning method called model-free learning control (MFLC) is introduced. The MFLC algorithm is based on a well known Q-learning algorithm (Watkins, 1989). Successful applications of RL are well documented in the recent literature, including learning to control mobile robots (Bucak and Zohdy, 2001), sustained inverted flight on an autonomous helicopter (Ng et al., 2004), and learning to minimize average wait time in elevators (Crites and Barto, 1996). However, only few articles can be found regarding RL applications for process control: multi-step actions based on RL was fruitfully applied for thermostat control (Schoknecht and Riedmiller, 2003), and one of the authors successfully applied RL for modeling for optimization in bath reactors by making the most effective use of cumulative data and an approximate model (Martinez, 2000). The reason for the difference between robotics and process control is possibly the nature of the control task in