Yokogawa and NAIST jointly developed a reinforcement learning algorithm

[ China Instrument Network Instrument Development ] Yokogawa Electric Corporation and Nara Institute of Advanced Science and Technology Research (NAIST) announced the joint development of an enhanced learning* algorithm for automatic optimization of plant operations. Reinforcement learning is a basic technology in the field of artificial intelligence (AI). The joint development of this algorithm provides a practical solution to improve the production quality and output of the factory.

Artificial intelligence and machine learning (ML) are a subset of artificial intelligence. Recently, it is expected to achieve breakthroughs in technological change in various fields, which has aroused widespread concern. AI is being used in real life, for example, autonomous vehicles and boats. Although ML has been put into plant data analysis, it must be further studied by companies and academic institutions before it can be applied to automation control.

Over the years, Yokogawa has provided control systems for various industries such as oil, natural gas, chemicals, steel, pulp and paper, medicine and food, and has acquired a large amount of technology and expertise related to plant operations. NAIST has been researching and developing ML-based technologies such as probabilistic reasoning and systems engineering techniques, optimization control and reinforcement learning, as well as developing intelligent robots and systems that perform specific functions in a dynamic environment.

Yokogawa and NAIST have successfully developed a new algorithm that uses Yokogawa's plant control technology and Yokogawa's knowledge and expertise of interdependence between control loops to improve kernel dynamic strategy programming (KDPP) and NIST reinforcement learning. technology. Traditional reinforcement learning algorithms require a large amount of search processing to ensure proper control, which is a challenge for practical applications. The newly developed algorithm significantly reduces the amount of training that must be done and is therefore highly practical. Yokogawa and NAIST have confirmed on the plant simulator that by using a new algorithm to simultaneously control four different valves during the distillation process at the vinyl acetate production plant, the optimization operation far exceeds what is possible with conventional control algorithms or manual operations.

Yokogawa and NAIST will conduct a (POC) concept test in an up-to-date factory environment to confirm the reliability of actual use. The newly developed algorithm was released at the IEEE International Conference on Automation Science and Engineering held in Germany from August 20th to 24th.

(Original title: Yokogawa and Nara Institute of Advanced Science and Technology (NAIST) jointly develop a reinforcement learning algorithm for automatic optimization of plant operations)