Acquiring Plant Operation Knowledge through Learning Classifier Systems

We are conducting a research and development project on agent technologies to enrich "Field Forces" in steel industries under the support of the Iron and Steel Institute of Japan. The objective of the project is to explore the applicability of recent software agent technologies to practical task domains performed by both human experts and plant systems. Under the support, this paper proposes a new method to extract plant operation knowledge from time varying plant data. The method is characterized by the use of Learning Classifier Systems (LCSs), which is one of machine learning methods with rule generation, modification by evolutionary algorithms. We have equipped plural learning components, each of which consists of XCS (eXtended Classifier System), a recent advanced version of LCSs. We have designed each XCS as a software agent with communication capabilities among the other agents and the operation environments. This paper describes the basic principles and implementation of the method, then explains how the proposed method can be applied to plant operation tasks for hot strip mills of a steel plant. In our methodological frame work, we do not use any plant specific knowledge but only rely on the plant operation data virtually generated by the simulator of hot strip mills operations developed by Konishi et al. The objective of the proposed method is, therefore, to explore the feasibility of the data oriented method with XCS agents. The proposed method is combined with the simulator of hot strip mills operation and shows the effectiveness.