Developing Real-Time Scheduling Policy by Deep Reinforcement Learning

Designing scheduling policies for multiprocessor real-time systems is challenging since the multiprocessor scheduling problem is NP-complete. The existing heuristics are customized policies that may achieve poor performance under some specific task loads. Thus, a new design pattern is needed to make the multiprocessor scheduling policies perform well under various task loads. In this paper, we investigate a new realtime scheduling policy based on reinforcement learning. For any given real-time task set, our policy can automatically derive a high performance by online learning. Specifically, we model the real-time scheduling process as a multi-agent cooperative game and propose multi-agent self-cooperative learning that overcomes the curse of dimensionality and credit assignment problems. Simulation results show that our approach can learn high-performance policies for various task/system models.