WIP: Demand-Driven Power Allocation in Wireless Networks with Deep Q-Learning
Anastasios Giannopoulos, Sotirios Spantideas, N. Capsalis, Panagiotis K. Gkonis, Panagiotis Karkazis, Lambros Sarakis, Panagiotis Trakadas, Christos N. Capsalis
Abstract
Power allocation is strongly related to the coverage and capacity of wireless networks, playing a critical role in the development of 5G networks. This paper proposes a Demand-Driven Power Allocation (DDPA) algorithm aiming to fulfill the requested throughput of individual users and accommodate their needs. DDPA is based on model-free Deep Reinforcement Learning (DRL) approaches and has the ability to proactively adjust the power levels of network transmitters. The performance of the developed algorithm is evaluated for a variety of simulation parameters and variable user demands. According to the presented results, the DDPA scheme exhibits a near-optimal performance for up to 50 users in the network area (i.e. satisfaction percentage exceeds 95%), with each one requesting 1 Mbps. Moreover, performance comparison between DDPA and two typical baseline methods reveals that the former results into enhanced total allocated throughput solutions (i.e. a performance increase by a factor of approximately 9% against baseline methods).