REINFORCEMENT LEARNING BASED WIRELESS BASE STATION PARAMETER OPTIMIZATION

Abstract

Nowadays, in order to provide customers with the best surfing experience through the wireless network. Companies started building modern wireless base stations with a large amount of algorithm-based parameters that can optimize the performance of a single base station. However, tuning these base stations to reach their best performance is not only a time-consuming task but also requires experts for tuning. Trying to make this tuning procedure more efficient, we introduced deep reinforcement learning and built a policy that can optimize a single KPI with a similar performance as the human experts. In our paper, we claimed to achieve the following accomplishments, • Built a simulator that can accurately describe a wireless base station’s parameter tuning scenario in the real world. The simulator enables the estimation of specific Key Performance Indicator (KPI) while can give rewards as feedback to the tuning actions made by a human or the policy. • Employed deep reinforcement learning, together with imitation learning and prioritized experience replay, to build an agent that can automatically tune the parameters for the base station with the performance better than human experts.