An Actor-Critic Deep Reinforcement Learning Approach for Transmission Scheduling in Cognitive Internet of Things Systems

The cognitive Internet of Things (CIoT) has attracted much interest recently in wireless networks due to its wide applications in smart cities, intelligent transportation systems, and smart metering networks. However, how to smartly schedule the packet transmission in CIoT systems is still a key challenge, that is, how to design a smart agent to realize the intelligent decision making and effective interoperability. In this paper, we model the system state transformation as a Markov decision process, and an actor-critic deep reinforcement learning algorithm based on a fuzzy normalized radial basis function neural network (called AC-FNRBF) is proposed to efficiently solve the intelligent transmission scheduling problem in CIoT systems under high-dimensional variables. The proposed AC-FNRBF algorithm can better approximate both the action function of the actor and the state-action value function of the critic without requiring the system prior knowledge, and a new reward function is established to maximize the system benefit, which jointly takes the transmission packet rate, the system throughput, the power consumption, and the transmission delay into account. Moreover, the AC-FNRBF has the ability to adjust its learning structure and parameters in dynamic environments. Simulation results verify that the proposed algorithm achieves higher transmission packet rate and system throughput with lower power consumption and transmission delay, compared with other existing reinforcement learning algorithms.