Load Balancing In Mobile Networks Using Deep Reinforcement Learning And Traffic Prediction

Abstract

Wireless communication networks are advancing at a
rapid pace, driven by various challenges and ambitious
goals. This rapid growth is driven by a range of
applications, including technologies like the Internet
of Things (IoT), as well as innovations in smart
cities, autonomous vehicles, and more. Different
applications demand specific performance criteria
such as high data throughput, low latency, robust
reliability, and efficient energy usage. In this thesis,
we investigate two enhancements that can be adopted
in wireless networks to tackle the challenges of
resource optimization and network management.
The motivation behind this is the fact that future
networks will face challenges like severe congestion
and varying traffic demands. The objective is to
achieve higher network throughput and more data
transmission by adjusting the network parameters.
The first proposed approach introduces an enhanced
self-optimization framework using deep
reinforcement learning (RL) to dynamically adjust
network parameters such as handover parameters,
power levels, and MIMO technology. The proposed
approach offers significant gains in network
throughput by effectively balancing the load
distribution. The proposed framework explores the
trade-off between system complexity and
performance improvement, demonstrating that
adopting a scenario-aware optimized agent can
outperform generalized agents under specific
network conditions. The second approach we tackle
is to adopt a proactive concept while controlling the
network. The proposed approach is based on the
ARIMA model used to predict the next states of the
environment so that the RL agent considers them in
the decision-making process. The simulation results
demonstrate that the proposed approach leads to
higher throughput and improved network
performance, which underscores its potential as a
robust alternative to the conventional agent existing
in earlier works.