PPO-Based deep reinforcement learning framework for dynamic resource allocation and network slicing in 5G mobile networks

Abstract

This study proposes a new intelligent framework to
cope with the challenges involved with dynamic resource allocation
in the 5G network environment based on Proximal Policy
Optimization (PPO), which is one of the most successful Deep
Reinforcement Learning (DRL) techniques. We have reformulated
resource allocation as a Markov Decision Process (MDP). Here, the
"state" represents the current status of the network in terms of
demand, interference, and channel quality. At the same time, the
"Action" represents the allocation decision made for each service
slice in terms of spectrum, capacity, and time. The proposed model
focuses on balanced dynamic resource allocation across three main
segments: eMBB, URLLC, and mMTC, through ensuring that
QoS requirements for each segment are met without impact to the
overall system performance. Our simulation results have
demonstrated excellent performance by the proposed algorithm
when compared to traditional algorithms (i.e., GA, PSO, QLearning,
and Round Robin). In our results, we showed a
throughput increase of approximately 180 Mbps, energy efficiency
of 0.91 bps/joule, a Fairness Index of 0.88 overall performance
improvement between 12% to 15%. As a result of the simulation
results, we believe that the PPO-MDP Framework is a good,
realistic option for optimizing the use of resources within a
dynamically segmented environment, thus improving the ability of
a 5G system to efficiently and sustainably respond to a variety of
service demands.