Knowledge-Embedded Deep Reinforcement Learning for Autonomous Network Decision-Making Algorithm

This paper proposes a multi-critic deep Reinforcement learning framework (MCDRL) and a knowledge-embedded multi-critic deep reinforcement learning(KE-MCDRL) Decision-making method, the method can ensure users' real-time QoS delay requirements. Compared with implementing the deep reinforcement learning algorithm directly in the communication system, this method can accelerate the convergence and guarantee the initial QoS performance of the system. Simulation results show that the design method can significantly reduce the convergence time compared with traditional deep reinforcement learning, and has nearly optimal decision delay compared with existing decision-making methods, which can actualize real-time decision-making in a time-varying channel environment.