Design of multiobjective fuzzy control system using reinforcement learning

Abstract

In practical control area, there are many examples with multiple objectives which may conflict or compete with each other like overhead crane control, automatic train operation, and refuse incinerator plant control, etc. These kinds of control problems are called multiobjective control problems, where it is difficult to provide the desired performance with control strategies based on single-objective optimization. Because the conventional control theories usually treat the control problem as the single objective optimization problem, the methods are not adequate to treat the multiobjective control problems. Particularly, in case of large scale systems or ill-defined systems, the multiple objective control problem is more difficult to solve due to the uncertainty in them. Therefore, the efficient method is required to solve the multiobjective control problem in large scale or ill-defined uncertain systems. The multiobjective control method that uses the conventional multi-objective optimization method required the exact model of the plant. However, the requirement is difficult to satisfy in large scale or ill-defined uncertain systems. On the other hand, the reinforcement learning changes the control rule on the basis of the evaluative information about the control results rather than the exact information about the environment. Therefore, the paper proposes the multiobjective controller design method using reinforcement learning for the multobjective control in large scale or ill-defined uncertain systems. In large scale or ill-defined uncertain systems, the traditional control methods are not applicable due to uncertainty. However, in many cases, human operators operate the system well based on their experience and knowledge. These mean it is necessary to design the controller using the human operator`s experience and knowledge in large scale or ill-defined uncertain systems. Fuzzy logic makes it easy to convert the human knowledge to the fuzzy rules. Theref...