Yokogawa and NAIST Jointly Develop Reinforcement Learning Algorithm Applicable to Automatic Optimization of Plant Operations

Tokyo, Japan - August 22, 2018

Yokogawa Electric Corporation
Nara Institute of Science and Technology

Yokogawa Electric Corporation (TOKYO: 6841) and the Nara Institute of Science and Technology (NAIST) announce that they have jointly developed a reinforcement learning* algorithm that can be utilized for the automatic optimization of plant operations. Reinforcement learning is a fundamental technique employed in the artificial intelligence (AI) field, and the joint development of this algorithm holds promise as a practical solution for the improvement of production quality and volume at plants.

AI and machine learning (ML), a subset of AI, have been drawing keen attention as the result of recent breakthroughs that hold promise for a technological transformation in various fields. AI is being put to practical use to, for example, automate the driving of automobiles and the maneuvering of ships. While ML is already being put to use in plant data analytics, further research must be done by companies and academic institutions before it can be put to practical use in automated control applications.

Over the years, Yokogawa has provided control systems to a wide range of industries, including oil, gas, chemicals, iron and steel, pulp and paper, pharmaceuticals, and food, and has acquired a wealth of technologies and expertise related to plant operations. NAIST, on the other hand, has been engaged in the research and development of ML-based techniques such as probabilistic inference as well as system engineering techniques such as optimization control and reinforcement learning, while aiming to develop intelligent robots and systems that are capable of carrying out specific functions in a dynamic environment.

Yokogawa and NAIST have succeeded in developing a new algorithm, leveraging Yokogawa's plant control technologies as well as its knowledge and expertise on the interdependencies between control loops, to improve Kernel Dynamic Policy Programming (KDPP), NAIST's reinforcement learning technology. Conventional reinforcement learning algorithms require an enormous amount of search processing to ensure appropriate controls, which poses challenges for practical use. The newly developed algorithm significantly decreases the amount of training that must be done, and is thus highly practical. Yokogawa and NAIST have already confirmed on a plant simulator that, by using the new algorithm in a distilling process at a vinyl acetate production plant to simultaneously control four different valves, optimal operation far beyond that what is possible with a conventional control algorithm or through manual operation can be achieved.

Yokogawa and NAIST will conduct a proof of concept (PoC) test in an actual plant environment to confirm its suitability for practical use.

The newly developed algorithm will be announced today at the IEEE International Conference on Automation Science and Engineering, which is being held in Germany from August 20 to August 24.

* A trial-and-error method by which software agents learn which actions will maximize a reward


About Yokogawa

Founded in 1915, Yokogawa engages in broad-ranging activities in the areas of measurement, control, and information. The industrial automation business provides vital products, services, and solutions to a diverse range of process industries including oil, chemicals, natural gas, power, iron and steel, and pulp and paper. With the life innovation business the company aims to radically improve productivity across the pharmaceutical and food industry value chains. The test & measurement, aviation, and other businesses continue to provide essential instruments and equipment with industry-leading precision and reliability. Yokogawa co-innovates with its customers through a global network of 113 companies spanning 61 countries, generating US$3.8 billion in sales in FY2017. For more information, please visit www.yokogawa.com.


Nara Institute of Science and Technology (NAIST) is a Japanese national university located in Kansai Science City, a border region between Nara, Osaka, and Kyoto. Founded in 1991, NAIST consisted of graduate schools in three integrated areas: information science, biological sciences, and materials science. In 2018, NAIST underwent an organizational transformation to continue research in these areas while promoting interdisciplinary research and education across traditional fields. With this new single graduate school organization, NAIST strives forward with the objectives of conducting cutting-edge research in frontier areas and training students to become tomorrow's leaders in science and technology.

The names of corporations, organizations, and products herein are either trademarks or registered trademarks of their respective holders.