Q-Learning Based Optimisation Framework for Real-Time Mixed-Task Scheduling

Mixed periodic and aperiodic tasks with explicit deterministic or probabilistic timing requirements are becoming increasingly deployed in real-time industry control systems. Such systems pose significant challenges to the scheduling algorithms because the failure of scheduling can be catastrophic. In the past decades, significant research effort has been dedicated on the scheduling problems, and various scheduling algorithms were proposed to meet various system requirements and task loads. However, a single fixed scheduling algorithm usually cannot fully satisfy the requirements for a dynamic mixed-task-set, which is commonly found in modern complex real-time control systems. It is thus extremely hard for engineers to design a set of scheduling solutions to guarantee the correctness and optimality under all conditions. Aiming at optimising the scheduling performance in a real-time control system, this paper proposes a Q-learning-based optimisation framework to select proper scheduling algorithms for the mixed-task-set. Built on a three-layer perceptron network, our Q-learning framework is able to efficiently and effectively choose scheduling algorithms that dynamically adapt to the characteristics of task-sets. Experimental results using real-world data proved the effectiveness of the proposed framework.