Reinforcement Learning for Scheduling Wireless Powered Sensor Communications

In a wireless powered sensor network, a base station transfers power to sensors by using wireless power transfer (WPT). Inadequately scheduling WPT and data transmission causes fast battery drainage and data queue overflow of some sensors who could have potentially gained high data reception. In this paper, scheduling WPT and data transmission is formulated as a Markov decision process (MDP) by jointly considering sensors' energy consumption and data queue. In practical scenarios, the prior knowledge about battery level and data queue length in MDP is not available at the base station. We study reinforcement learning at the sensors to find a transmission scheduling strategy, minimizing data packet loss. An optimal scheduling strategy with full-state information is also investigated, assuming that the complete battery level and data queue information are well known by the base station. This presents the lower bound of the data packet loss in wireless powered sensor networks. Numerical results demonstrate that the proposed reinforcement learning scheduling algorithm significantly reduces network packet loss rate by 60%, and increases network goodput by 67%, compared to existing non-MDP greedy approaches. Moreover, comparing the optimal solutions, the performance loss due to the lack of sensors' full-state information is less than 4.6%.