Edge Computing Optimization for 5G Networks: A Machine

Learning Approach

Kevin Chang, PhD¹, Dr. Priya Sharma², and Prof. Antonio Rossi¹
¹Department of Electrical Engineering, Georgia Tech, Atlanta, GA | ²Wireless Systems Lab, Nokia Bell
Labs, NJ

Abstract

The proliferation of Internet of Things (IoT) devices and latency-sensitive applications

necessitates intelligent resource allocation at the network edge. This paper presents a
novel machine learning framework for optimizing edge computing resource allocation
in 5G networks. Our approach employs deep reinforcement learning to dynamically
assign computational tasks between edge servers and cloud infrastructure based on
real-time network conditions, application requirements, and energy constraints.
Evaluation using real-world network traces demonstrates a 43% reduction in average
latency compared to baseline algorithms, while improving energy efficiency by 31%.
The framework adapts to traffic fluctuations and maintains quality of service
guarantees for critical applications including autonomous vehicles and augmented
reality services.

1. Introduction

Fifth-generation (5G) wireless networks promise unprecedented data rates and ultra-low
latency, enabling transformative applications in autonomous systems, industrial automation,
and immersive media. However, realizing these capabilities requires fundamental rethinking
of network architecture and resource management.

Edge computing extends computational capabilities to the network periphery, reducing latency
by processing data closer to end users. Yet optimal placement and allocation of edge
resources remains challenging due to dynamic network conditions, heterogeneous application
requirements, and constrained resources at edge nodes.

This work addresses these challenges through a deep reinforcement learning framework that
learns optimal resource allocation policies from network experience. Unlike traditional
optimization approaches that rely on simplified models, our method adapts to real-world
complexity and evolving traffic patterns.

2. System Model and Problem Formulation

We model the edge computing system as a three-tier architecture consisting of IoT devices,
edge servers, and cloud data centers. Each tier has distinct computational capabilities,
latency characteristics, and energy profiles.
The resource allocation problem is formulated as a Markov Decision Process (MDP) where
the state space encompasses network conditions (channel quality, congestion), application
characteristics (computational demand, latency requirements), and resource availability.
Actions correspond to offloading decisions and resource allocations.

The objective function balances multiple criteria: minimizing end-to-end latency, maximizing
energy efficiency, and maintaining application QoS requirements. We introduce a weighted
reward function that dynamically adjusts priorities based on network state and application
criticality.

3. Deep Reinforcement Learning Framework

Our solution employs a dueling Double Deep Q-Network (D3QN) architecture enhanced with
prioritized experience replay. The neural network consists of four fully connected layers with
512 units each, using ReLU activation and batch normalization.

Key innovations include a multi-timescale learning approach that handles both rapid channel
fluctuations and longer-term traffic patterns. Fast timescale decisions (milliseconds) address
immediate offloading choices, while slow timescale adaptations (minutes) optimize resource
provisioning.

We incorporate domain knowledge through shaped rewards that guide exploration toward
promising regions of the action space. Safety constraints ensure that critical applications
always receive sufficient resources, even during exploration phases.

The framework includes online learning capabilities, continuously refining policies as network
conditions evolve. Transfer learning techniques enable rapid adaptation when deploying to
new network configurations.

4. Experimental Evaluation

Experiments used real-world network traces from a major metropolitan 5G deployment,

covering three weeks of operations with over 10,000 base stations and 500,000 active users.
Application mix included autonomous vehicle coordination (15%), augmented reality (25%),
smart city sensors (40%), and general mobile broadband (20%).

Our D3QN approach was compared against four baseline algorithms: greedy local
processing, cloud-only offloading, static edge allocation, and model-based optimization. All
methods were evaluated under identical traffic conditions using discrete-event simulation
validated against deployment data.

Results show significant improvements across all metrics. Average end-to-end latency
decreased from 47ms (best baseline) to 27ms with our approach, a 43% reduction. The 95th
percentile latency improved from 182ms to 89ms, critical for ultra-reliable low-latency
communications.

Energy consumption per successfully served request decreased by 31%, primarily through
intelligent workload consolidation and sleep scheduling of underutilized edge servers. Carbon
footprint analysis indicates this translates to approximately 12 tons CO2 reduction per base
station annually.
Quality of service metrics exceeded requirements consistently. Autonomous vehicle
applications experienced zero deadline violations compared to 3.4% for the best baseline.
Augmented reality sessions maintained target frame rates (90 fps) 99.7% of the time versus
94.2% for static allocation.

5. Analysis and Discussion

Detailed analysis reveals that performance gains derive from three key capabilities: (1)
accurate prediction of resource demands, (2) proactive resource provisioning before
congestion occurs, and (3) effective load balancing across heterogeneous edge infrastructure.

The multi-timescale learning approach proves essential. Fast timescale decisions respond to
channel variations and immediate congestion, while slow timescale adaptations handle
diurnal traffic patterns and gradual infrastructure changes.

Ablation studies demonstrate that each component contributes meaningfully. Removing

prioritized replay degrades performance by 12%, while eliminating the safety constraints
results in occasional catastrophic failures for critical applications.

Computational overhead of the learning framework is manageable. Inference requires 0.3ms

per decision on standard server hardware, well within the 1ms requirement for 5G control
plane operations. Online learning adds negligible latency as training occurs asynchronously.

6. Conclusions and Future Work

This work demonstrates that deep reinforcement learning can effectively optimize edge
computing resource allocation in complex, dynamic 5G networks. The substantial
improvements in latency, energy efficiency, and reliability support deployment in production
systems.
Future research directions include extending the framework to handle multi-access edge
computing across multiple operators, incorporating federated learning for privacy-preserving
optimization, and addressing emerging 6G requirements including integrated sensing and
communications.

The methodology developed here generalizes beyond 5G networks, with potential

applications in content delivery networks, distributed database systems, and other edge
computing scenarios requiring real-time resource orchestration.

References

[1] Mao, H., Alizadeh, M., Menache, I., & Kandula, S. (2016). Resource management with
deep reinforcement learning. ACM HotNets.

[2] Wang, S., et al. (2019). Deep reinforcement learning for dynamic computation offloading in
mobile edge computing. IEEE Transactions on Mobile Computing, 18(3), 771-784.

[3] Mnih, V., et al. (2015). Human-level control through deep reinforcement learning. Nature,
518(7540), 529-533.
[4] Schaul, T., Quan, J., Antonoglou, I., & Silver, D. (2015). Prioritized experience replay.
ICLR.

[5] Li, L., et al. (2021). Edge intelligence: The confluence of edge computing and artificial
intelligence. IEEE Internet of Things Journal, 8(16), 12435-12455.