Source-linked AI summary
In-Edge AI: Intelligentizing Mobile Edge Computing, Caching and Communication by Federated Learning
Xiaofei Wang, Yiwen Han, Chenyang Wang, Qiyang Zhao, Xu Chen, Min Chen
TL;DR
The paper addresses intelligent joint resource management in mobile edge computing, caching, and communication. It couples Deep Reinforcement Learning with Federated Learning in the In-Edge AI framework, achieving near-optimal performance in edge caching and computation offloading while motivating further collaborative edge-AI research.
Problem
Mobile edge systems need intelligent joint management of communication, computation, and caching resources beyond existing approaches.
Method
In-Edge AI couples Deep Reinforcement Learning with Federated Learning, enabling edge nodes and devices to collaborate through learning updates for caching and computation tasks.
Results
The framework achieves near-optimal performance in experiments on edge caching and computation offloading, with federated DDQN approaching centralized DDQN after repeated model merging.
Takeaways & Limitations
In-Edge AI can support adaptive mobile-edge optimization while motivating collaborative scheduling of AI tasks across edge nodes and mobile devices.
Takeaways & Limitations
The privacy and security implications of Federated Learning are not the focus of this work.
Abstract
from arXiv · showhide
Recently, along with the rapid development of mobile communication technology, edge computing theory and techniques have been attracting more and more attentions from global researchers and engineers, which can significantly bridge the capacity of cloud and requirement of devices by the network edges, and thus can accelerate the content deliveries and improve the quality of mobile services. In order to bring more intelligence to the edge systems, compared to traditional optimization methodology, and driven by the current deep learning techniques, we propose to integrate the Deep Reinforcement Learning techniques and Federated Learning framework with the mobile edge systems, for optimizing the mobile edge computing, caching and communication. And thus, we design the "In-Edge AI" framework in order to intelligently utilize the collaboration among devices and edge nodes to exchange the learning parameters for a better training and inference of the models, and thus to carry out dynamic system-level optimization and application-level enhancement while reducing the unnecessary system communication load. "In-Edge AI" is evaluated and proved to have near-optimal performance but relatively low overhead of learning, while the system is cognitive and adaptive to the mobile communication systems. Finally, we discuss several related challenges and opportunities for unveiling a promising upcoming future of "In-Edge AI".
I. INTRODUCTION
Mobile edge computing reduces backbone burden and latency by placing resources near users, but jointly allocating communication and computation remains difficult under uncertain, dynamic, and incomplete information. The paper proposes combining DRL with FL in an In-Edge AI framework for adaptive MEC resource management.
- MEC pushes computation and storage toward user equipment, reducing backbone workload and propagation delay while supporting latency-critical applications.
- Jointly allocating wireless communication and edge computation resources is difficult because offloading can congest wireless channels.
- Existing optimization methods may rely on difficult-to-obtain inputs and often address only snapshots of highly time-varying MEC systems.
- The paper uses DRL to jointly manage communication and computation resources for edge caching and computation offloading, with FL training agents distributively.
- The In-Edge AI framework is evaluated through proof-of-concept experiments and reported to balance performance and cost while achieving near-optimal behavior.
II. OPTIMIZING THE EDGE BY DRL
The paper structures cognitive computing in MEC as a loop that collects system observations, fuses them into scheduling decisions, and handles user requests accordingly. DRL is positioned as the cognitive method, with caching and offloading as representative use cases.
- DRL is proposed as the cognitive-computing method for intelligentizing mobile edge computing, caching, and communication.
- Information Collecting: Information collecting gathers resource usage, wireless conditions, and user-request intensities from the MEC system.
- Cognitive Computing: Cognitive computing fuses observed system data and produces scheduling decisions.
- Request Handling: Request handling serves user-equipment requests according to the scheduling decision generated by cognitive computing.
- The section investigates edge caching and computation offloading as two representative MEC use cases.
A. DRL over the MEC System for Caching
The caching use case places popular content at edge nodes and uses DRL agents to decide which files to cache. The setup assumes slowly changing popularity and equal-sized contents to assess DRL efficiency.
- Caching content at intermediate servers serves repeated user demands without duplicate remote-cloud transmissions, reducing redundant traffic.
- The MEC caching model defines a library of F popular files and represents requests through a content-popularity distribution.
- Each edge-node DRL agent decides whether to cache requested content under slowly changing popularity and equal-content-size assumptions.
- The communication setting contains UEs covered by base stations, with UEs able to offload computation to edge nodes through wireless channels or execute locally.
- Wireless channel gains are modeled as finite-state discrete-time Markov transitions, and achievable rates are evaluated using the Shannon-Hartley theorem.
- Tasks are generated over time and queued for sequential FIFO execution on UEs or edge nodes.
2) Computation Model:
The computation-offloading model lets each UE jointly choose local or wireless execution and allocate energy. Deep Q-Learning and Double DQN train policies from network state to improve task-execution QoE.
- 2) Computation Model:: An offloaded task’s edge execution time is modeled as dE = ν/fE, with edge capability assumed much stronger than local UE capability.
- 3) Problem Formulation for Computation Offloading:: The UE’s control action jointly specifies the offloading decision c and allocated energy e for wireless communication and local computation.
- 3) Problem Formulation for Computation Offloading:: The action chooses local execution when c = 0 or wireless offloading otherwise, while e denotes allocated energy.
- 3) Problem Formulation for Computation Offloading:: Deep Q-Learning and Double DQN train a stationary policy from task queues, energy use, channel occupancy, and wireless-channel qualities.
- 3) Problem Formulation for Computation Offloading:: The immediate utility evaluates QoE through execution and queueing delays, UE energy consumption, and task dropping or failure.
III. IN-EDGE AI WITH FEDERATED LEARNING
The paper introduces Federated Learning to train distributed DRL agents across UEs and edge nodes, addressing heterogeneous, non-IID, unbalanced, privacy-sensitive, and intermittently connected environments. This supports In-Edge AI for collaborative edge caching and broader MEC resource management.
- Federated Learning trains DRL agents distributed across UEs and edge nodes while addressing heterogeneous and non-IID training data.FedAvg merges model updates from clients whose transition memories reflect their own wireless environments, computation capabilities, and energy consumption.
- FL reduces communication demands by uploading model updates rather than training data and by allowing only a subset of clients to upload in each round.Additional local computation can further reduce the number of communication rounds.
- FL addresses unbalanced client data and intermittent UE availability through model aggregation and partial client participation.The approach is designed for settings where UEs have varying amounts of training data or may be unpredictably offline.
- The framework uploads minimal DRL-agent updates, while secure aggregation and differential privacy are identified as applicable privacy mechanisms.Privacy and security are explicitly outside the paper’s main focus.
- In-Edge AI uses collaboration between edge nodes and UEs to support dynamic system-level optimization and balancing across the MEC system.The framework applies to edge caching and other mobile edge system tasks.
1) From Edge to Cloud:
The paper describes how FL connects UE-level learning with edge-side coordination for MEC applications. Devices train locally and share model parameters, enabling shared learning while retaining local training data, with a computation–communication trade-off.
- UEs use DRL-agent inference to decide task offloading, wireless-channel selection, and energy consumption.These decisions support computation-offloading scenarios in which device resources constrain large-scale DRL training.
- FL iteratively distributes model parameters, performs local training on client data, and aggregates uploaded updates at a server.Clients download the model, train locally, and upload only updated parameters.
- FL enables resource-constrained UEs and edge nodes to learn a shared model while keeping training data local.This distributes knowledge across many devices without centralizing their data.
- On-device data makes In-Edge AI more cognitive by exposing wireless-channel, battery, energy, computation-load, storage, and queue-state information.These data sources are more directly tied to MEC conditions than server-side proxy data.
- FL offers a computation–communication trade-off: more local computation can reduce communication rounds, but federated optimization cannot outperform centralized DL.The paper characterizes federated performance as near-best rather than superior to centralized training.
C. Practicability Discussion
In-Edge AI is presented as a future-oriented framework whose practicability depends on deployment constraints and training delay. Transfer learning is proposed to avoid training DRL models from scratch before distributing them to mobile UEs.
- In-Edge AI envisions mobile devices that can both infer and train Deep Learning models.
- Edge chips currently support only elementary Deep Learning training processes, so deployment must account for practical and delay requirements.
- Randomly initialized DRL models should not be deployed because preliminary exploration can make random decisions and paralyze the MEC system.
- Transfer learning can train an offline DRL agent in a simulated wireless environment before distributing the established model to mobile UEs.
2) Delay Requirement:
The paper addresses delay and computation concerns by using pre-established, lightweight DRL models and evaluating them in simulated caching and offloading settings. The simulations specify distributed clients, network channels, and Double DQN configurations.
- Transfer learning is expected to reduce training time or computation consumption compared with training from scratch.
- A pre-established DRL model can reach the desired accuracy after few mini-batches or directly infer, while adapting when environments or request patterns change.
- The DRL agents use an MLP with one hidden layer and 200 neurons, enabling quick inference after training because the MainNet has low complexity.
- The evaluation uses one-month Xender traces containing 9,514 active users, 188,447 content files, and 2,107,100 requests, with six cooperating edge nodes.
- For offloading, 10 UEs train DRL agents through FL over 10 wireless channels within a 5 MHz edge-node bandwidth, with channel quality represented by six levels.
- The experiments use Double DQN with tanh activation, Adam optimization, and fully connected networks containing 200 neurons.
B. Evaluation Results
Experiments evaluate In-Edge AI for edge caching and computation offloading. DDQN with FL approaches centralized DDQN and outperforms the listed heuristic baselines, while FL trades lower communication requirements for heavier client computation and merging delay.
- DDQN with FL improves caching hit rates and offloading utilities during training before stabilizing within a certain range.
- In edge caching, DDQN with FL achieves hit rates near centralized DDQN and outperforms LRU, LFU, and FIFO policies.
- In computation offloading, DDQN with FL performs near centralized DDQN and exceeds mobile execution, edge-node execution, and greedy execution baselines.
- Centralized DDQN performs better initially, but DDQN with FL approaches it after several model-merging rounds, requiring clients to wait for merging.
- FL shifts computation burden to clients because the coordinating server merges updates rather than conducting the whole training, increasing UE energy or UE and edge-node computation.
V. OPPORTUNITIES AND CHALLENGES
The paper identifies opportunities to use edge systems not only to optimize mobile communication, but also to support and coordinate learning computation itself. This reverse direction introduces resource-allocation and adaptive task-splitting challenges.
- Future work can use edge systems to optimize learning computation tasks as a specialized form of edge computing at the service level.
- This reverse direction of In-Edge AI requires selecting collaborating edge nodes and allocating resources for AI tasks with varied priorities, deadlines, scales, and hardware requirements.
- Potential research includes game-theoretic competition over multidimensional resources for AI acceleration at the edge.
- Dynamic and adaptive splitting of AI tasks is identified as a challenging research problem.
B. Efficiency of In-edge AI for Real-time Mobile Communication System towards 5G
In-Edge AI targets rapid, intelligent optimization for highly time-varying 5G mobile edge systems, where conventional deep-learning schemes may converge too slowly for millisecond-scale tasks. Experiments on caching and computation offloading report near-optimal performance.
- 5G URLLC requires very small delays and high reliability, but general deep-learning optimization and prediction can take too long to converge for millisecond-scale edge tasks.The paper identifies this convergence-time mismatch as a challenge for real-time mobile edge systems.
- Reliable and accurate incentives across mobile operators, service providers, content providers, and users remain an open challenge, including distributing proof-of-work computation and evaluating heterogeneous contributions.Blockchain integration is suggested, but these workload and contribution-assessment questions remain unexplored.
- In-Edge AI integrates Deep Reinforcement Learning and Federated Learning to optimize mobile edge computing, caching, and communication.The framework uses collaboration among mobile operators, service providers, content providers, users, edge nodes, and devices.
- Experiments on edge caching and computation offloading show that In-Edge AI achieves near-optimal performance in mobile edge systems.The authors identify future work on fine-grained collaborative scheduling of AI tasks across edge nodes and mobile devices.