Directed control is crucial for implementing controllers on programmable logic controllers. A directed controller is one that actively selects at most one controllable event (control command) to execute at any instant. This study investigates the numerical optimization problem of directed control for discrete-event systems, such as traffic systems and robotic systems. By integrating supervisory control theory and reinforcement learning, an algorithmic procedure is designed to synthesize an optimal directed controller, which minimizes the total cost associated with event execution while ensuring the safety and liveness of the controlled system. Two improved Q-learning algorithms incorporating dynamic parameter adaptation strategies are developed to enhance the global search ability of basic Q-learning. To demonstrate applicability and effectiveness, we apply the proposed method to control a guideway system and a multi-train traffic system, respectively. The experimental results indicate the superiority of the proposed method over the comparison methods.
Read full abstract