Reduced state space representation for reinforcement learning based wireless scheduling

Abstract

As the wireless network structure becomes more complex to handle heavy traffic generated from IoT devices, smart phones, etc., and to satisfy various QoS requirements, existing network scheduling algorithms are showing limitations. Wireless Scheduling based on reinforcement learning is emerging as a method to efficiently allocate limited resources while overcoming the limitations of existing algorithms, since without prior knowledge of complex environment, reinforcement learning finds optimal policy by interaction with the environment. However, to apply reinforcement learning to the complex network of reality, it is necessary to solve the curse of dimensionality problem that occurs as the size of the state space increases. In this study, we analyze previous research and limitations of non-learning and learning-based network scheduling algorithms, and develop a state space representation method by selecting candidate users according to each user’s priority given by specific criterion. With this representation method, we propose a new deep reinforcement learning based wireless network scheduling algorithm.