Journal of Robotics and Control (JRC) Volume 2.
Issue 3 ISSN: 2715-5072 DOI: 10.
18196/jrc.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Israa Fars Hassan 1.
Othman Atta Ismael 2 Department of History.
College of Education for Women.
AL Iraqia University.
Baghdad.
Iraq 2 Department of Networks Engineering and Cyber Security.
College of Engineering.
AL Iraqia University.
Baghdad.
Iraq Email: 1 Israa.
hasan@aliraqia.
iq, 2 othman.
atta@aliraqia.
iq, othman.
atta@yahoo.
*othman.
atta@aliraqia.
iq, othman.
atta@yahoo.
AbstractAiThe rapid proliferation of latency-sensitive 5G applications, such as autonomous driving and telemedicine, necessitates highly energy-efficient and adaptive resource management in Mobile Edge Computing (MEC) systems.
Conventional static optimization and isolated machine learning approaches often fail in dynamic, large-scale network To address this problem, an adaptive machine learning framework that integrates Deep Reinforcement Learning (DRL), predictive analytics, and Federated Learning (FL) is proposed for the joint optimization of task offloading and transmission power allocation.
The primary research contribution is a novel, scalable, and privacy-aware architecture that enables real-time, coordinated decisionmaking across distributed edge servers.
The methodological approach employs an Actor-Critic DRL agent for adaptive policy learning, a predictive analytics module for workload and channel forecasting, and a Federated Learning mechanism for distributed model training and privacy preservation.
The experimental results, based on synthetic data .
enerated via a Poisson proces.
and real traffic traces from the 'CRAWDAD' dataset, demonstrate that the proposed framework achieves significant improvements.
Compared to baseline methods (DRL.
FL-RA.
HBO.
GT-ES), it yields up to 35.
2% energy savings .
ean, 95% CI: A2.
1%), a 28.
2% reduction in latency (CI: A1.
8%), and maintains a 92.
1% packet delivery ratio (CI:
A1.
5%).
In conclusion, this work provides a validated, intelligent solution for sustainable MEC deployment in 5G networks and establishes a foundation for future 6G-enabled edge intelligence systems.
KeywordsAiMobile Edge Computing.
Deep Reinforcement Learning.
Federated Learning.
Energy Efficiency.
Resource Allocation.
INTRODUCTION
The development of wireless networks to the fifth generation .
G) systems has made it possible to achieve ultrareliable low-latency communication (URLLC), enhanced mobile broadband .
MBB), and massive machine-type communication .
MTC) .
, .
To facilitate these demanding conditions.
Multi-Access Edge Computing (MEC) has become a paradigm whereby the computation is no longer done on the centralized cloud infrastructures but instead on the edge servers directly in front of the end-users .
Abbrevation Symbol
Abbreviation QoS
DRL
HBO
GT-ES
Definition Task i generated by a mobile user MEC server j System state at time t Action selected at time t .
ask offloading, power Reward value at time t Policy mapping states to actions in reinforcement learning Total energy consumption of the system Computational energy consumption Transmission energy consumption Idle energy consumption Latency per task System throughput Quality of Service Number of tasks in the system Number of MEC servers Available bandwidth Transmission power CPU frequency of MEC server Weighting factor for energyAelatency trade-off Discount factor in reinforcement learning Learning rate in training Federated Learning Deep Reinforcement Learning Heuristic-Based Optimization Game-Theoretic Scheduling Confidence Interval Standard Deviation Such an architectural change minimizes latency and congestion of the backbone and enhances Quality of Service (QoS) of delay-sensitive applications like autonomous driving and healthcare monitoring .
Nevertheless, the implementation of MEC can present a major energy consumption problem since the distributed edge nodes are active, and they respond to adaptive task offloading requests .
Conventional methods of optimization, including convex programming, heuristic allocation, and game-theoretic scheduling, do not provide adaptive solutions to dynamic and large-scale 5G conditions but provide deterministic solutions .
They make a set of assumptions that are not dynamic, which leads to poor use of energy, low level of scalability.
Journal Web site: http://journal.
id/index.
php/jrc Journal Email: jrc@umy.
Journal of Robotics and Control (JRC) ISSN: 2715-5072 and complexity of computation.
Recently, machine learning (ML) techniques like federated learning (FL) .
and deep reinforcement learning (DRL) .
have been used to optimize MEC successfully.
However, most of the research focuses on either task offloading or transmission power allocation alone, whereas a more comprehensive, adaptive, framework that can combine the two is not well studied .
Furthermore, the transparency of datasets and the methodological clarity of most works provide little opportunity to re-use and implement practically.
The following main gaps can be observed in the extant literature:
A Lack of combined frameworks that coordinate the optimization of offloading and resource assignment in dynamic MEC conditions .
A A small methodological rigor, where most studies do not declare the ML architectures, training procedures, and adaptive update schemes.
A The description of datasets is not sufficient, containing abstract notions of what can be referred to as synthetic or real traces, which are not reproducible.
A Unfavorable comparative analysis, in which baseline methods are not tested at the same time, nips claims of superiority .
A Ignoring the importance of both computational overhead and scalability is essential to the implementation of MEC in resource-constrained settings.
The challenges limit the process of translating ML-based optimization of MEC to real-world and sustainable 5G deployments and limit its extrapolation to 6G-enabled MEC systems, including Reconfigurable Intelligible Surfaces (RIS) and UAV-enabled MEC .
Although the recent literature, including .
, has considered hierarchical adaptive federated RL to conduct MEC resource allocation, our approach stands out based on three major features:
A Temporal-spatial coordination, which takes into account the instantaneous conditions of a network and the forecasted future conditions through built-in predictive A Multi-objective reward formulation, which dynamically balances the weight of energy, latency, and throughput according to the needs of the application.
A Scalable federated aggregation which reduces the cost of To address these gaps, a new adaptive ML framework is suggested, in which the DRL, predictive analytics, and FL are combined into one co-optimization framework.
Its main innovation is that, its hierarchical joint decision-making mechanism optimizes simultaneously the offloading of tasks and the transmission power in real-time, and is not holistically considered in previous independent methods.
This unified architecture allows dynamically adapting to the changes in the network with a balance of energy consumption, end-to-end latency, and calculation costs.
To compensate these gaps, this work develops a particular research question: How can an adaptive ML framework be used to jointly optimize task offloading and transmission power distribution within dynamic 5G MEC environments to maximize the energy efficiency whilst meeting latency requirements? In order to respond to it, the following hypotheses can be formulated to be tested:
H1: A unified DRA-FL-predictive analytics system will satisfy statistically significant positive changes in energy efficiency relative to non-integrated or non-inflated baseline procedures.
H2: Under the joint optimization mechanism, the throughput and packet delivery ratio will be maintained or enhanced at the expense of a reduced latency without the need to place a prohibitive computational load.
H3: The federated learning aspect will allow scalable model updates that can be conducted privacy-observant with major convergence performance without a significant decline compared to centralized training.
The main contribution of this paper is designing, implementing and critically evaluating a new adaptive ML framework that can confirm these hypotheses by the use of massive In dealing with these issues, the adaptive ML-based framework in this paper suggests combining both DRL and predictive analytics to share tasks offloading and transmission power allocation.
Compared to fixed or singlepoint solutions, the framework suggested can dynamically adjust to the changes in the network conditions on the fly, so that it can implement energy-efficient utilization of available resources and maintain QoS.
Using simulation experiments, which are founded upon synthetic as well as actual traffic traces, the proposed model is compared to four baselines, namely.
DRL.
FL-based Resource Allocation (FL-RA).
Heuristic-Based Optimization (HBO), and Game-Theoretic Scheduling (GT-ES).
Findings indicate that energy efficiency, throughput and latency are greatly enhanced without causing high computational overhead, hence illustrating the possibility of the framework being used in actual MEC deployments.
Summing up, the paper bridges the gap that is critically needed in dynamic MEC settings joint and adaptive This paper has three research contributions:
A Architectural innovation: The suggestion of an integrated framework to jointly use DRL, predictive analytics, and FL to optimize task offloading and transmission power management simultaneously is an issue that was mostly addressed independently in the literature.
A Methodological Rigor and Reproducibility: An elaborate and transparent experimental plan has been laid down, comprising of dataset description, simulation cross-validation interpretation, making it easy to replicate or benchmark.
Empirical Evidence and Practical Experiences: A largescale evaluation through simulation on a variety of stateof-the-art baselines shows strong statistically significant results on performance superiority.
Moreover, the analysis provides practical implications of trade-offs, limitations, and feasible directions of the real-world implementation and the integration of 6G in the future.
Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 Fig.
Architecture of Multi-Access Edge Computing (MEC) in 5G Networks The rest of this paper is organized as following: Section II is related works and gaps.
Section i system model, problem formulation and proposed adaptive ML framework.
Section IV is the description of the dataset and simulation setup.
Section V is the description of the experimental results and the comparative analysis.
Section VI is the discussion and limitations, and the last section VII presents the future research directions.
II.
RELATED WORKS
Recently, energy efficiency in MEC has been one of the most active topics since it is the key to the sustainability of 5G networks.
Several works have been done so far with the aim of optimizing the energy consumption of MEC environments, focusing on different aspects like task offloading, resource allocation, and dynamic power Traditional energy-efficient approaches mainly rely on heuristic algorithms, convex optimization techniques, and rule-based strategies that balance the computational workload to minimize power usage in computation .
Such methods depend on predefined static models, which cannot adapt well to the dynamic and heterogeneous nature of MEC While effective to a certain degree, such solutions fall short in providing the needed flexibility to deal with realtime fluctuations in network traffic, user mobility, and energy constraints in the case of suboptimal performance in dynamic scenarios .
As MEC systems grow more complex, there has been a growing trend of using ML-based techniques as effective alternatives to the traditional optimization Different ML-based models, including reinforcement learning, deep learning, and neural networkbased frameworks, have been applied to improve energy efficiency in MEC environments.
Reinforcement learning, especially DRL, has gained much attention due to its capability of autonomous decision-making in complex and dynamic environments.
DRL-based approaches can intelligently optimize resource allocation and task offloading strategies by continuously learning from network conditions and adjusting decision-making policies accordingly .
Predictive energy management has considered both the supervised and unsupervised learning techniques while analyzing the historical data to forecast the future workload distribution and pre-acting power consumption .
The performance comparison between conventional optimization techniques and ML-based techniques underlines benefits coming from intelligent decision-making in energy Conventional approaches typically include game theory, linear programming, and evolutionary While these can provide deterministic solutions, they have a high computational cost and can hardly be scaled for large-scale MEC deployment .
On the contrary.
MLdriven models can utilize real-time data for adaptive decisionmaking in enhancing energy efficiency without sacrificing network performance.
Besides.
ML-based frameworks can automatically extract features and recognize patterns, thus dynamically adapting to changing MEC conditions and optimizing power consumption accordingly .
On the other hand.
ML approaches also bring about their own challenges, such as high computational overhead, increased training complexity, and large datasets that are needed for making accurate Despite the promising advances in ML-based energy optimization, several limitations remain in state-of-the-art Most related works focus on either a task offloading or resource allocation perspective toward MEC energy efficiency, not considering both within an integrated framework .
Besides, most ML models require substantial pre-training and may not generalize well to unseen conditions of the network, thus hindering adaptation in actual deployment.
Also, there is an evident lack of common datasets and benchmarking schemes to compare ML-based energy optimization strategies fairly.
this makes comparison of different methodologies quite difficult in .
The challenges brought forth by all these calls for further research work on hybrid machine learning models integrating several learning techniques, enhanced generalization strategies, and the establishment of standardized evaluation frameworks that make the energy-efficient solutions for MEC in 5G networks robust and scalable.
Many approaches have been used to find solutions to resource allocation and offloading of tasks in Mobile Edge Computing (MEC).
Conventional methods, including heuristic algorithms, convex optimization and game-theoretic models were offering solutions with lower complexity but in Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 many cases were based on static assumptions, restricting their ability to adapt to dynamic 5G worlds.
On the same note, the evolutionary algorithms enhanced convergence in special situations, yet it could not generalize.
During the advent of machine learning (ML), more adaptive methods were suggested .
Deep Reinforcement Learning (DRL) allowed systems to learn with evolving network states, which works better on the energy efficiency and latency Nonetheless.
DRL is expensive to compute and needs a significant amount of training data.
Federated Learning (FL) solved the privacy and distributed training issues but added delays in communication and overhead in aggregation.
Recent studies examined hybrid models that integrate DRL with predictive analytics or heuristic models by trying to trade off flexibility and efficiency.
Recent developments .
involve the incorporation of Reconfigurable Intelligible Surfaces (RIS).
UAV-assisted MEC, and 6Goriented hybrid systems, which focus on scalability and However, most of these works are still simulationbased and do not provide an integrated framework to simultaneously optimize the task offloading and transmission power distribution and guarantee reproducibility and transparency of the dataset.
with a minimum power cost without compromising on the network performance.
Besides, it develops an adaptive learning mechanism whereby decision-making policies are updated at runtime continuously, considering real network The proposed mechanism further strengthens the robustness of the optimization process and enables the dynamic response of the model against changing workloads, user mobility, and fluctuating energy demands.
Furthermore, the study has presented comprehensive performance evaluation by comparing the proposed ML-based approach against traditional heuristic and mathematical optimization Extensive simulations demonstrate the efficiency, scalability, and adaptability of ML-driven energy management .
This work also develops a unified framework that incorporates several ML techniques, including DRL and predictive analytics, to achieve optimal energy management in MEC.
These techniques, when put together, enabled smart resource scheduling, proactive power allocation, and enhanced decision-making capability.
Finally, it provides practical implications for real-world deployment, thus helping network operators/service providers implement ML-based energy optimization in existing 5G infrastructures.
This work bridges the gap between theoretical advancements and practical applications in the path to future innovations of TABLE I.
Comparative Analysis of MEC Resource Allocation Approaches (Traditional and Recent Work.
Approach / Technique Game-theoretic optimization Convex optimization DRL-based offloading Actor-Critic DRL DRL Heuristic Federated Learning (FL) FL for MEC Hybrid ML heuristics ML Predictive analytics RIS-assisted MEC UAV-enabled MEC 6G-oriented hybrid framework Dataset Used Synthetic Simulated Synthetic traces Real workload traces Synthetic Distributed traces Mobile workload dataset Simulated Real traces Simulated RIS dataset UAV traces Real synthetic Advantages Low complexity Efficient mathematical formulation Adaptive, high accuracy Energy-efficient, improved latency Faster convergence Privacy-preserving, scalable Scalable, distributed Balanced adaptability Improved forecasting accuracy Enhanced energy saving Flexible coverage Scalable, future-ready Table 1 contains a comparative synopsis of the most usable works in the field, indicating the methods used, the nature of the datasets, the merits and shortcoming.
This adaptability ensures that the system remains efficient even in fluctuating network environments.
Besides, scalability of the proposed framework will be able to apply it to large-scale MEC deployments and thus make it suitable for 5G and beyond-5G networks.
From an environmental perspective, this study aligns with green networking by reducing the overall energy footprint of edge computing infrastructure, thereby contributing to global efforts toward sustainable digital transformation 19.
First, this work has contributed significantly to the area of energy-efficient MEC in 5G networks, proposing an ML-based new optimization framework for joint task offloading and resource allocation, which makes it a very integrated solution for energy management .
While there are previous models discussing either the partial or separated features of the MEC, the model is addressing the balancing of the computation load Limitations Static, lacks adaptability Not scalable in dynamic MEC High training cost Large state/action space Computationally expensive Communication overhead Sensitive to delays No standardized dataset Limited scalability High deployment cost Mobility-induced delay Still simulation-based Ref.
energy-efficient MECs.
The key contribution of this work is to propose an MLbased framework for achieving optimally energy-efficient MEC in 5G networks.
This is achieved by developing a resource allocation mechanism that adaptively adjusts with changes in the network, while keeping power consumption as low as possible without affecting the quality of the services It also aims to try and bridge an important gap created by traditional techniques of optimization that are inflexible and non-adaptive to network conditions in real time .
Another objective is to critically assess the efficacy of the ML-based approach as compared to heuristic and mathematical methods of optimization.
The study tries to prove that the ML-driven decision-making of energy management is superior to that through an investigation into key performance metrics such as energy consumption, latency, and computation overhead.
Second, this research investigates how the intelligent learning model affects practical deployments of 5G regarding their feasibility.
Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 scalability, and adaptability.
Additionally, this work intends to develop practical recommendations on how best ML-based energy management solutions could be implemented within the MEC environment .
It also aims to contribute to the industry stakeholders, such as network operators, service providers, and researchers, by highlighting challenges and potential improvements.
The goal of this study is to contribute to the development of sustainable and intelligent MEC systems that will ensure 5G networks work efficiently with a minimal environmental impact.
The rapid expansion of 5G networks has brought up a surge in data traffic, computational demands, and wireless connected devices, which result in huge energy consumption challenges never experienced before.
Multi-Access Edge Computing has emerged to drive efficiency in such scenarios by offloading computation closer to users, thus reducing latency and bandwidth usage.
However.
MEC improves the performance but simultaneously increases the power consumption due to the deployment of multiple edge servers and continuous resource allocation in MECs .
Classic energy management techniques, such as rule-based policy and mathematical optimization, usually cannot adapt to dynamic MEC environments, leading to suboptimal energy utilization.
Therefore, new approaches are needed in intelligent energy efficiency optimization to sustain MEC for continued 5G operations .
Machine learning thus provides great avenues through adaptive and data-driven decision-making.
Contrary to previous methods, machine learning can enable the algorithms to learn from experience in history for workload predictions, thus taking resource adjustment actions for minimal consumption of energy without losing QoS.
Most of the existing research on ML-driven energy optimization in MEC is fragmented, with most studies focusing either on task offloading or resource allocation in Very few comprehensive frameworks integrate multiple ML techniques to provide holistic energy optimization in MEC environments.
Besides, many ML models require huge computational resources and time for training, which makes them impractical for real-time decision-making.
This work is, therefore, crucial in bridging these research gaps through the development of an integrated ML-based framework that will jointly optimize task offloading with resource allocation for energy-efficient MEC .
The developed approach will achieve better decisionmaking with reduced computational overhead and further enhance adaptability to real network variations by incorporating reinforcement learning along with predictive These findings will have important practical implications for network operators and service providers, as this research will help shed light on how real-world 5G networks implement ML-driven energy management In this way, this research work contributes to the advancement in green computing, ensuring future wireless communication systems have high performance while being energy-efficient, amidst growing demand for sustainable and intelligent solutions .
SYSTEM MODEL AND PROBLEM FORMULATION
Architecture of Multi-Access Edge Computing in 5G Networks The architecture of MEC in 5G consists of three tiers: user equipment, edge nodes, and the core network.
The UE includes mobile devices.
IoT sensors, and autonomous systems generating computation tasks that need to be Instead of relying solely on cloud computing, these tasks can be offloaded to nearby MEC servers deployed at 5G base stations (BS.
, reducing latency and bandwidth congestion .
MEC servers provide localized computing, storage, and caching capabilities, ensuring real-time data processing for latency-sensitive applications.
The BSs act as intermediaries, forwarding computational tasks to either MEC servers or the cloud, depending on the network conditions, workload, and energy efficiency considerations.
However, the architectural task of offloading and resource management require intelligent optimizations to strike the balance between power consumption, computation efficiency, and service quality.
Energy Consumption Model for MEC Nodes and 5G Base Stations The of energy MEC systems include multiple parts: the power for executing tasks, the communication power betwe en network components, and the idle energy of the MEC The total energy consumption yatotal in a MECenabled 5G network can be modeled as:
yatotal = yacomp yatrans yaidle .
yacomp represents the energy consumed for computation, yatrans denotes the energy required for data transmission, yaidle accounts for the power consumption of MEC servers and BSs in idle states.
The computation energy yacomp is modelled as:
yacomp = OcycA ycn=1 ycEycaycyyc y ycNyceycuyceyca .
where ycEycaycyyc is the CPU power consumption per unit cycle, and ycNexec .
is the execution time of task i.
The execution time is given by:
ya ycNexec .
Oe yceycn where yaycn is the computational complexity .
n CPU cycle.
and yceycn is the allocated CPU frequency for task ycn.
The transmission energy yatrans is given by:
yaycycycaycuyc = OcycA ycnOe1 ycEycycu y ycNycycu .
where ycEycyc is the transmission power, and ycNycyc .
is the transmission time of task ycn, given by:
Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ycNycycu .
ycIycn
ISSN: 2715-5072
where yaycn is the data size of task ycn, and ycIycn is the transmission The idle energy yaidle is modeled as:
yaidle = OcycA ycOe1 ycEidlle y ycNidle .
where ycEidlle is the idle power consumption, and ycNycnyccycoycoyce .
is the duration for which MEC node yc remains idle.
Problem Formulation: Optimization Objectives and Constraints The primary objective of this study is to minimize the total energy consumption while ensuring efficient resource utilization and maintaining the required Quality of Service (QOS).
The optimization problem can be formulated as min OcycA ycn=1 .
) The suggested algorithm has a polynomial complexity, which means that it is computationally viable in real-time 5G MEC Algorithm 1: Adaptive ML-based Resource Allocation in 5G MEC Input:
Set of tasks T = .
1, t2.
A, t.
MEC servers S = .
1, s2.
A, s.
System parameters: CPU cycles, bandwidth, transmission power Initialize:
State space representation .
orkload, channel state, energy level.
Action space .
ffloading decision, transmission power allocatio.
Deep Reinforcement Learning agent (Actor-Critic Predictive analytics model for channel and workload ya,ya subject to:
Computational Capacity Constraint:
yceycn O ycemax .
OAycn ensuring that the allocated CPU frequency does not exceed the MEC server's maximum processing capability.
Transmission Bandwidth Constraint:
ycIycn O ycIrax .
OAycn ensuring that the allocated transmission rate does not exceed the available network bandwidth.
Latency Constraint:
ycNerec .
ycNycycu .
O ycNmax .
OAycn ensuring that task execution and transmission do not exceed the delay tolerance of the application.
For each episode do For each task ti OO T do Observe current state .
Predict next workload and channel conditions Select action .
= .
ffloading decision, transmission powe.
using policy A Execute action at on MEC server sj Compute f(EnergySaving.
LatencyReduction.
Throughpu.
Update Actor and Critic networks using gradients Store .
t, at, rt, st .
in replay buffer End for Federated Learning step: aggregate updated local models across servers End for Output:
Optimized policy A* for adaptive task offloading and transmission power allocation Energy Budget Constraint:
yatotal O yabudget Ensuring that overall energy consumption remains within acceptable limits to sustainable MEC operations.
Three processes make up the computational complexity of the proposed framework, and they include:
Deep reinforcement Learning (DRL) training - its complexity is about O (E y T y N y A) where E is the number of episodes.
T is the number of tasks per episode.
N is the state dimensionality, and A is the size of the action space.
Predictive Analytics - this is O (N log N) in the case of workload prediction and is often more complex depending on the model.
Federated Learning updates Federated Learning model aggregation with M servers adds an overhead of O(M y P) to communication overheads where P is the number of parameters moved per update.
DRL Model Specification the Actor-Critic network uses a state space S = .
orkload distribution, channel SNR, battery levels, queue length, location dat.
, having the N=15 The action space A = .
ffloading decision .
, transmission power level .
, server selection .
has 8 dimensions.
The reward functionality has been mathematically expressed as .
ycI = yu y (Oeyuuya ) yu y (Oeyuuy.
yu y .
uuycN) Oe yu y .
cEycyyceycuycaycoycyc ) .
where yuuya is a normalized energy saving, yuuya is a reduction in latency, yuuycN is an improvement in throughput and ycEycyyceycuycaycoycyc is a penalty against policy oscillations.
QoS requirements are dynamically used to determine the weights yu, yu, yu using a meta-controller.
Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 generalizable to a variety of network conditions.
This synergistic combination enables the system to be proactive .
hrough predictio.
and adaptive .
hrough reinforcement learnin.
, and is scalable and private .
hrough federated learnin.
Mathematical Representation of Energy Efficiency in MEC Energy efficiency .
uC) is defined as the ratio of successfully processed tasks to the total energy consumed:
OcycA ya yuC = yaycnOe1 ycn Maximizing yuC ensures that the system processes more computational workload per unit of energy consumed, improving overall efficiency.
The challenge lies in balancing computational and transmission power consumption while meeting latency and QoS requirements.
Fig.
Flowchart of Methodology Figure 2 shows the connection between the main ML elements in the suggested system.
The process of work works as follows:
A Predictive Analytics Module: Since the predictive analytics can be used to predict future states in the shortterm, historical and real-time network data .
orkload, channel stat.
are used to make predictions.
These are the predictions which are fed at the state representation of the DRL agent.
A DRL Agent (Actor-Criti.
: The agent chooses a joint action, which includes decisions of offloading and transmission power levels, based on the enriched state .
urrent and predicte.
The policy update is guided by the reward function which aims at balancing a number of A Federated Learning Orchestrator: Local DRL models are trained on individual MEC servers and are aggregated every now and then.
The step guarantees protection of privacy and utilization of distributed knowledge without centralizing the raw data, making the model more Figure 2.
Adaptive ML Framework Proposed Flowchart.
The scheme depicts the workflow and interaction between the core components, namely Data Acquisition and Preprocessing, the Predictive Analytics Module, the DRLbased Joint Optimization Agent .
long with its state, action, reward cycl.
, the Federated Learning Orchestrator to merge the models, and the ultimate Resource Allocation decision that affects the MEC system.
Table 3 represents the definition of the model, energy consumption, transmission power, and the variations in data rate necessary to implement dynamic resource allocation.
For example.
Energy Efficiency Threshold EET = 50 Joules.
This work considers the total of 100 sensor nodes (SN).
each has transmitted data in power range between 0.
1 to 2.
0 Watt.
Data Rate.
DR total between 1 and 10 Mbps varies because it also shows the different forms of total consumed energy.
And Initial Energy consumption.
Ei is from 5 to 20 Joules depending on each network load condition and depending on transmission conditions.
The range of Optimum TP is 05 and 1.
5 Watts, while the reduction factor of adaptive transmission power varies from 0.
1 to 0.
maintain stability in operation, a threshold energy limit is imposed at 10 Joules to effectively adjust power.
This is a full parameter set that may enable the work to have the best TABLE II.
Parameters Simulation Variables Threshold of Energy Efficiency Nodes of the Sensor Power of the Transmission Ration of Data Consumption of Energy Consumption of Initial Energy Power of Optimized Transmission Energy Allocation Threshold Alpha (Factor of Power Reductio.
Symbols EET TP_opt Values 1 - 2.
1 - 10
Variable 5 - 20 05 - 1.
1 - 0.
Unit of the measure Joules (J) Count .
Watts (W) Mbps (Bits per secon.
Joules (J) Joules (J) Watts (W) Joules (J) Dimensionless Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 tradeoff between energy efficiency and network performance in MEC-based 5G systems.
TABLE i.
Key Parameters for Energy Optimization in MEC Networks Parameter Energy Efficiency Threshold
Sensor Nodes Symbol
EET
Value Unit Measure Joules (J) Transmission Power Data Rate 1 - 2.
1 - 10 Mbps Energy Consumption Initial Energy Consumption Optimized Transmission Power Threshold for energy allocation Alpha .
ower reduction facto.
Variable .
5 - 20 Count .
Watts (W) Bits per second .
Joules (J) TP_opt Joules (J) Watts (W) Dynamic .
- 1.
1 - 0.
Dimensionless Joules (J) The algorithm of this study as below:
Energy Efficiency Algorithm 1: Define: Energy Efficiency Threshold (EET) 2: Input: Sensor Nodes (SN) = .
1, s2, s3, .
, s.
Transmission Power (TP).
Data Rate (DR) 3: Initialization: Energy Consumption EC = 0.
Node Index I = 0 4:Output: Optimized Energy Allocation for Multi-hop Communication 5: for each sensor node si in SN do 6: Compute Initial Energy Consumption Ei based on TP and DR Ei = TP y DR 7: while EC O EET do if Ei O Threshold then Allocate Transmission Power Optimally TP_opt = TP y .
Ae ) Adjust Transmission Power and Recalculate EC EC = (TP y DR) end if Update EC IIaI 1 15: end while 16: end for which restricts practical deployment.
Regarding these challenges, the present research work proposes an ML-based energy optimization framework that dynamically updates task offloading and resource allocation strategies in real time.
In such a context, the proposed framework will contribute to enhancing energy efficiency without compromising network performance by effectively integrating reinforcement learning with predictive analytics.
The details of the proposed methodology, algorithm design for performance evaluation, and proof of the efficiency of ML-driven energy management in MEC-enabled 5G networks are presented in subsequent Simulation and Data Set-Up.
To test the suggested adaptive ML-based framework, synthetic data and real traffic traces were used.
The synthetic dataset was developed based on the simulation of heterogeneous workloads, arrival rate of the tasks, and different conditions of the channels between mobile users, and publicly available mobile edge computing workloads datasets were used to obtain real traces which guaranteed realistic user behavior and resource requirement As shown in Preprocessing of tasks included input feature normalization (CPU cycles, bandwidth requirement and task siz.
to make tasks easier to train.
The simulation environment was written in python 3.
10 containing 12 to train the ML model and NS-3 to simulate the network and was run on a workstation with Intel core i9 processor, 32GB of RAM and an NVIDIA RTX 3090 card.
To train the model, a Deep Reinforcement Learning (ActorCriti.
model was employed, complemented by predictive analytics to estimate workload and channel and affiliated with federated learning (FL).
TABLE IV.
Summarizes the Key Parameters Category Dataset Parameter Task types Task size CPU cycles per Dataset type Challenges in Achieving Optimal Energy Efficiency Preprocessing In MEC, several problems impede the path of optimum energy efficiency.
The first one is due to the dynamic nature of the 5G networks: fluctuating user demands, time-varying channel conditions, and mobility patterns cannot be handled with static optimization models.
Second is the tradeoff in energy efficiency with computational performance, without excessive delay in execution time or overutilization of available resources.
First, conventional mathematical optimization methods hardly scale with the growing complexity of real-world MEC deployments, which necessitates the need for advanced machine learning-based Second.
MEC nodes and 5G base stations work in heterogeneous environments.
thus, adaptive and decentralized energy management solutions are needed Finally, real-time decision-making is very critical for energy-efficient MEC, and most of the existing approaches suffer from fast convergence rates and generalizability.
Simulation Setup Training Setup MEC servers (M) Mobile users (N) Bandwidth (B) Transmission power (P) Channel model Simulation tool Algorithm Optimizer Learning rate (A) Discount factor Batch size Replay buffer size Number episodes (E) Convergence Value / Description Computation-intensive, delaysensitive, mixed workloads 5 Ae 10 MB 500 Ae 1500 cycles/bit Synthetic .
Real traffic traces Normalization of workload and channel features 50 Ae 200 10 Ae 20 MHz 1 Ae 1.
Rayleigh fading NS-3 integrated with Python
Actor-Critic DRL Predictive Analytics Federated Learning Adam
10,000 transitions Stable reward
A5%
Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 TABLE V.
Result Table Method
Proposed MLBased DRL
FL-RA
HBO
GT-ES
Energy Consumption (J) 0 (A0.
Latency Reduction (%) 2 (A1.
Throughput (%) Adaptability 1 (A1.
High Computational Overhead Moderate 3 (A1.
1 (A1.
2 (A0.
5 (A1.
5 (A2.
2 (A2.
0 (A1.
1 (A1.
4 (A2.
7 (A2.
3 (A1.
5 (A1.
Medium High Low Medium High Very High Low Moderate Experimental Protocol & Reproducibility Details In order to have reproducibility, the following protocol is given: Dataset Splits: The mixed dataset .
ynthetic and real trace.
is divided into 70 % training, 15 % validation and 15% A 5-fold cross-validation is used and the results averaged by the folds.
Random Seeds & Hyperparameters: There are three random seeds .
, 123, .
in all the experiments.
The hyperparameters .
ncluding learning rate, discount factor ) are determined by a grid search on the validation set.
Table 4 shows the final selected values.
Federated Learning Protocol FedAvg is implemented with aggregate frequency as 10 local epochs.
The communication cost is measured in terms of the sum of megabytes delivered across the globe in a round.
The total time of wall-clock training of the entire framework is observed.
Baseline Implementation & fair comparison: The implementation of all baseline algorithms (DRL.
FL-RA.
HBO.
GT-ES) is done on the same simulation environment (NS-3 Pytho.
They are given the same data sets, compute restrictions and adapted to their respective optimum settings so as to achieve a fair comparison.
Availability of Code and configuration: The code of the simulation, configuration files, and the samples of the processed datasets will publicly be available on publication in a GitHub repository .
ink to be provide.
IV.
RESULTS
This paper proposes a new machine learning-based solution for optimizing energy efficiency in MEC of 5G The proposed methodology based on dynamic transmission power and task allocation with an adaptive energy management strategy can significantly reduce energy consumption with the preservation of network performance.
In order to put the performance of the proposed framework into perspective, a comparative analysis with findings reported in studies that were done recently and which are directly related is provided.
As an example, the average energy saving was only about 22 % when the DRAbased method of .
was applied to task offloading only.
however, our joint optimization structure resulted in savings of more than 35 percent.
In the same way, the integrated method enhances throughput unlike the FL-RA method, which was demonstrated to have a throughput of about 88% in comparable conditions as discussed in .
Such an improvement can be attributed to the predictive aspect that reduces the latency of FL rounds and enables resource adjustment in a timelier manner.
The experimental results demonstrate an average energy efficiency of 35% and a corresponding reduction in power consumption from a total average over 20 Joules/node down to 13 Joules during dynamic workload conditions.
Besides these, the proposed model ensures latency-aware data communications with an overall average delay reduction of 28%, which highly supports real-time applications in the IoT and/or autonomous systems.
The adaptive power allocation mechanism allows runtime adjustments according to variations in network traffic, keeping throughput levels at their optimum, above 92%, beating static allocation.
These results depict the effectiveness of machine learning approaches toward addressing the energy consumption challenges of MEC-based 5G networks.
Table 5 and Figure 3 show the result of this study.
To validate the efficiency of the proposed method, a comparison was made against four recent energy optimization techniques in MEC:
Deep Reinforcement Learning (DRL) Ae Utilizes deep Qlearning for power optimization.
Federated Learning-based Resource Allocation (FL-RA) Ae Implements decentralized learning for adaptive energy Heuristic-based Optimization (HBO) Ae Uses rule-based algorithms for energy-efficient resource allocation.
Game Theory-based Energy Scheduling (GT-ES) Ae Employs Nash equilibrium strategies for workload Table 3 illustrates the simulation parameters employed in this research.
It shows energy thresholds, transmission power, data rate, and dynamic power controls that are important while evaluating the given ML-based optimization The simulation environment of this study as shown in table 4 was created using a computing system capable of supporting network simulations and machine learning The system consists of multi-core processor highperformance computing nodes with sufficient RAM and GPU acceleration for accelerating model training.
The platform software includes Python-based libraries such as TensorFlow and PyTorch for machine learning model implementation and MATLAB and NS-3 for network simulation.
These tools support the analysis of transmission power optimization, energy consumption modeling, and the impact of dynamic network conditions on energy efficiency.
The research Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 employs an adaptive energy management framework that is deployed using machine learning technologies.
The system can dynamically adjust resource allocation and power control based on the actual network conditions in real time.
The primary algorithm utilizes reinforcement learning methods to accomplish energy optimization with the best QoS.
The simulation process includes data preprocessing, model training, validation, and real-time inference, which allows for efficient testing of various optimization methods.
The data set used in this study is actual and synthetic data, simulating various network traffic behaviors and power consumption It includes:
A History of energy consumption patterns from MultiAccess Edge Computing (MEC) nodes.
A Task offloading data and their respective power A Network throughput and transmission power level measurements at different loads.
A Latency records for various task sizes and offloading Experimental Setting.
The DRL agent stabilized after around 650 episodes, and the policy stabilized between episode 600 and 700.
Every measure of performance .
uch as 92% ratio of packet deliver.
had been taken at the testing stage with invisible information.
Percentage of throughput and PDR are used uniformly in the entire paper and PDR is computed using the formula .
uccessfully delivered packets/ total packets sen.
The data was also normalized and cleaned before feeding into the machine learning algorithm to yield appropriate and interpretable results.
Various subsets of the data were trained and tested to assess its generalization capability in different MEC scenarios.
The results show that the proposed approach outperforms the state-of-the-art techniques by maintaining low energy consumption and high throughput with guaranteed real-time adaptability.
Unlike DRL and FL-RA, which are computationally expensive in terms of training, the proposed approach has a proper trade-off between energy efficiency and computational overhead.
The heuristic-based approach, although computationally lightweight, is poorly adapted to network efficiency.
Similarly, game theory-based scheduling gives average performance and lacks real-time adaptability in dynamic workloads.
Figure 3 and Table 5 shows the Performance comparison graph for energy consumed in Joule and throughput percentage for each technique of optimization.
Among the compared approaches, the proposed ML-based performs with the minimum energy consumption while offering the maximum throughput, becoming the most efficient technique.
Figure 3.
Comparison of results of energy optimization methods in MEC networks.
The bar chart shows the average energy .
n Joule.
used and the percentage throughput of the proposed method and four baseline algorithms .
eft and right Y-axis respectivel.
The error bars are the standard deviation provided in five independent runs.
The suggested ML-based approach is more efficient and effective, as this happens to have the lowest power consumption and the highest
DISCUSION
Ablation and Sensitivity Analysis: Ablation study is done to find out the contribution of each component.
The loss of the predictive analytics module leads to a 12 percent energy consumption increase and a 15 percent latency variance increase and reflects the ability to stabilize decision-making in relation to fluctuations.
model that has FL component turned off .
hrough centralized trainin.
converges faster by about 20 per cent at the cost of overfitting to individual server data and privacy.
A sensitivity analysis of the DRA discount factor () indicates that g values of 0.
9 to 0.
97 are stable, with the best results being obtained with () =0.
Statistical Redundancy: Paired t-test of the energy consumption data of all test episodes is carried out.
It is statistically significant when compared to all the baselines .
value < 0.
The 95% confidence intervals in the results table and abstract are obtained by computing the distribution of results of the 5 test folds.
Analysis of Failure Modes: The framework's performance degrades under extreme network congestion .
acket loss > 30%), where predictive accuracy drops and DRL exploration becomes less effective.
This scenario is identified as a boundary condition for the current model.
It is revealed in the experimental findings that the suggested adaptive ML-based framework, on average, outperforms the baseline techniques (DRL.
FL-RA.
HBO and GT-ES) in energy efficiency, latency, and throughput.
In particular, the framework demonstrated savings of up to 35 percent, a 28 percent latency reduction, and 92 percent of packet delivery ratio (PDR), which proves the efficiency of the framework to balance computational performance with quality of service (QoS).
The effectiveness of integrating DRL, predictive analytics, and FL into a single architecture to support dynamic 5G MEC environments is justified by these In a practical perspective, the findings demonstrate the relevance of the framework to real-life applications, including autonomous driving, telemedicine, and immersive applications (AR/VR), in which ultra-lowlatency and power efficiency is paramount.
The FL integration guarantees privacy and provides distributed learning over MEC servers, which is a highly scalable functionality that fits the 5G security and scalability needs.
Moreover, predictive analytics is important in stabilizing resource distribution when workloads and channel conditions are highly variant, which decreases the threat of service Although these results show that there are major improvements, various trade-offs and difficulties in implementation are observed.
It takes the DRA agent around 650 episodes to converge, and this might present an obstacle in the rapid deployment to highly volatile conditions.
Nonetheless, this expense can be spread throughout the working period and measures such as transfer learning are suggested to alleviate this expense.
The FL update communication overhead is estimated to cause an extra load of 15-20 percent on the backhaul links under the configuration tested.
The management of this overhead is by adaptive frequency of aggregation and compressing of The results obtained in a statistical analysis of the obtained data indicate that the energy saving reported to be Israa Fars Hassan.
An Adaptive DRL-FL-Predictive Analytics Framework for Joint Task Offloading and Power Allocation in 5G MEC: A Simulation-Based Evaluation Journal of Robotics and Control (JRC) ISSN: 2715-5072 achieved is within a 95% confidence level of A2.
1% .
% A
1%), which indicates the strength of the provided results.
This is a trade-off mainly between the complexity of the model .
nd therefore the quality of decisions thereo.
and the latency of the inference.
our model has inference times below 5 ms, which is adequately low to run on MEC in real-time.
Despite its advantages, the framework offered has several First, the DRA element necessitates a significant amount of training time until it converges, thus making it not as applicable to application in highly dynamic MEC situations in the short term.
Second, the federated learning updates provide extra communication overhead that can impact its performance in networks with limited backhaul Third, the analysis was majorly performed in a simulated setting, and actual implementation can expose more issues concerning the hardware limitations, diverse gadgets, and varying mobility trends.
In the future, several improvements can be made.
Transfer learning integration may also decrease the training cost of DRL by using trained models on similar MEC conditions.
Likewise, the use of lightweight communication in FL .
, compressed model update.
can be used to minimize communication delays.
Another prospective future application of the framework to 6G-enabled technologies, like Reconfigurable Intelligible Surfaces (RIS) and UAV-assisted MEC nodes, will provide scalability and resilience to next-generation networks.
general, the suggested framework provides a solid setting of sustainable and adaptive MEC resource management yet the need to overcome these limitations is critical to its implementation on a large scale.
Fig.
Performance Comparison of Energy Optimization Techniques in MEC Networks VI.
CONCLUSION
In this paper, a validated adaptive machine learning structure of the joint-offloading-transmission-powerallocation optimization in 5G MEC systems will be The essence of the work offered is a new architecture that entails the combination of Deep Reinforcement Learning, predictive analytics, and Federated Learning to make coordinated decisions in real-time under dynamic network conditions.
The proposed framework is conclusively proven to perform better with up to 35.
2 and 2 improvement in energy savings and reduction in latency respectively with a high ratio of packet delivery of 92.
1 in comparison to state-of-the-art baselines.
The main limitations are discovered to be the original DNR training convergence and the federated update communication overhead.
To this end, it is suggested that future work should center on incorporating meta-learning to quickly adapt models, creating lightweight and adaptive FL protocols that minimize the cost of communication and test their protocol on experimental 5G/6G testbeds that use the newly emerging technologies such as Reconfigurable Intelligent Surfaces (RIS).
Conflicts of Interest The authors declare no conflict of interest.
Acknowledgments The authors would like to rapid their sincere appreciation to Al-Iraqia University for their generous support and cooperation in this research.
REFERENCES