Task scheduling for control system based on deep reinforcement learning

We investigate the control system's computational task scheduling problem within limited time and with limited CPU cores in the cloud server. We employ a neural network model to estimate the runtime consumption of linear quadratic regulators (LQR) under varying numbers of CPU cores. Building upon this, we model the task scheduling problem as a two-dimensional bin packing problem (2D BPP) and formulate the BPP as a Markov Decision Process (MDP). By studying the characteristics of the MDP, we simplify the action space, design an efficient reward function, and propose a Double DQN-based algorithm with a simplified action space. Experimental results demonstrate that the proposed approach has improved training efficiency and learning performance compared to other packing algorithms, effectively addressing the challenges of task scheduling in the context of the control system.