A newly proposed framework called Adaptive Cooperative Graph-Attention Service Replacement, or ACGSR, significantly enhances mobile edge computing performance by uniting spatial demand forecasting, reinforcement learning, and cooperative neighbor discovery into a unified system control loop. Developed to support latency-sensitive applications such as augmented reality, virtual reality, intelligent transportation, and industrial automation, the architecture addresses the chronic operational bottlenecks that arise when computation and storage resources are distributed across base stations close to end users. Unlike conventional content caching, which handles static files, mobile edge service caching requires hosting executable application container images, runtime dependencies, libraries, and synchronization states directly on edge servers. In extensive simulations modeled on real-world edge server topologies, the researchers found that this tightly coupled architecture reduced average request latencies by up to forty percent while simultaneously boosting cache hit rates and curbing energy consumption compared to established edge management baselines.
The fundamental challenge confronting mobile edge computing lies in the sharp trade-offs between local execution, cooperative inter-edge offloading, and remote cloud fallback under severe hardware restrictions. When edge nodes run out of local storage capacity or experience sudden surges in user requests, improper replacement decisions can trigger cache thrashing, recurrent container pulling, excessive migration penalties, and severe service-level agreement breaches. Existing edge management methods frequently partition these operational dimensions, applying isolated machine learning models or heuristic policies to task offloading, static content caching, or isolated server clustering. In contrast, the research team designed an optimization formulation that simultaneously balances multi-objective performance criteria including overall service latency, cache hit ratio, dynamic edge computation energy, inter-node transmission overhead, container migration costs, queue stability, and deadline violations across discrete operational time intervals.
To anticipate where computational requests will arise across an urban or regional edge infrastructure, the ACGSR framework employs a Graph Attention Network that treats edge base stations and their inter-server communication links as a dynamic graph. By evaluating node feature vectors comprising historical demand rates, cached application states, queue lengths, and residual cache capacities, the graph attention model computes dynamic attention coefficients between neighboring edge nodes. This enables the system to capture spatial correlations driven by regional popularity and user movement across cellular base stations rather than relying solely on temporal demand histories. During controlled evaluations, the graph attention forecasting module attained a root mean square error of 0.087, a mean absolute error of 0.066, and a mean absolute percentage error of 7.9 percent, outperforming conventional temporal predictors such as long short-term memory networks, gated recurrent units, and autoregressive integrated moving average models by substantial margins.
Operational decision-making within the framework is driven by a Dueling Double Deep Q-Network that governs service caching, retention, replacement, and routing across six categorical actions. The dueling structure separates the evaluation of underlying edge system states from the relative advantage of specific caching or offloading actions, preventing optimistic value overestimation when workloads experience bursty fluctuations. Rather than deploying static reward weights, the reinforcement learning agent incorporates an adaptive reward weighting mechanism that updates individual penalty and reward terms based on recent deviations from predefined operational targets for latency, migration overhead, energy usage, service-level agreement compliance, control traffic, and cache hit ratio. In addition, the policy enforces strict system feasibility boundaries, ensuring that execution traffic intensity remains within queue stability thresholds and restricting the total volume of service replacements executed at any given edge node per control slot.
When an incoming computational request misses the local cache of an edge node, ACGSR initiates a cooperative edge discovery mechanism to search for the required service image among adjacent base stations before triggering cloud fallback. The architecture restricts its active neighbor exploration set to a latency-bounded group of up to five neighboring edge nodes, avoiding the exponential communication overhead associated with full-mesh inter-edge broadcasting. Candidate neighbor servers exchange lightweight metadata snapshots containing cache indicators, queue lengths, delay estimates, and timestamps. Requests are cooperatively forwarded to neighboring nodes only when the metadata validity threshold confirms that advertised information remains fresh, balancing the latency benefits of localized cooperative execution against the serialization delay and potential misdirection associated with stale state information.
To prevent continuous and unnecessary container swapping, ACGSR couples the reinforcement learning policy with a utility-guided service replacement and eviction function. When cache capacity is exhausted and a new service is selected for admission, the utility model evaluates candidate services according to historical popularity, graph-predicted future arrivals, dependency-sharing benefits, container image transfer costs, cold-start delays, and state synchronization overhead. Cached services that offer high library-sharing utility or that incur substantial migration and cold-start penalties are shielded from eviction, while services with the lowest composite utility are selected for removal. Furthermore, an incoming application is admitted only if its calculated utility exceeds that of the eviction candidate without exceeding the node replacement budget, directly mitigating cache thrashing under volatile request arrivals.
The performance of the proposed architecture was evaluated across thirty independent simulation runs using an extended EdgeCloudSim environment integrated with Python-based learning modules. The simulation mapped base station coordinates from the publicly available EUA Edge Computing Edge Servers Dataset, constructing realistic network topologies that featured twenty edge nodes and one hundred distinct service categories operating under Poisson arrival processes with Zipf-distributed popularity skews and burst multipliers reaching up to three times base traffic. The experimental results demonstrated that ACGSR achieved an average service latency of 75.76 milliseconds, outperforming established baseline frameworks including CSC-noLM, CDOSC, conventional deep reinforcement learning caching, and TOSC by 40.57 percent, 32.35 percent, 24.28 percent, and 15.80 percent, respectively, with non-overlapping confidence intervals confirming statistical significance.
Beyond average response times, the architecture demonstrated pronounced advantages in tail latency and service reliability, which represent vital metrics for industrial and real-time edge workloads. In tail-latency evaluations, ACGSR lowered ninety-ninth percentile response times to 141 milliseconds, representing reductions of 42.2 percent relative to CSC-noLM and 19.9 percent compared to TOSC. Under extreme traffic loads reaching fifty requests per second per region, ACGSR maintained an overall service-level agreement violation ratio of only 11.2 percent, whereas baseline schemes suffered breach rates ranging from 17.9 percent up to 31.2 percent. These gains were mirrored in system efficiency, where ACGSR recorded a cache hit ratio of 75.16 percent and curbed cloud fallback to 13.9 percent, yielding total normalized energy savings between 26.92 percent and 47.22 percent primarily through the curtailment of heavy backhaul data transmissions.
Component ablation experiments and sensitivity analyses confirmed that each design element contributes crucially to system stability and responsiveness. Disabling the cooperative discovery component caused mean latency to surge to 97.1 milliseconds due to heavy cloud offloading, while eliminating queue awareness or the Graph Attention Network predictor raised average latencies to 90.5 milliseconds and 88.4 milliseconds, respectively. Removing migration cost penalties caused a sharp spike in replacement frequency to roughly fifteen replacements per epoch, showing that aggressive caching without migration awareness produces temporary latency gains at the cost of network thrashing. Furthermore, testing across unseen service catalogues and modified base station topologies revealed modest latency increases of 3.2 percent to 6.7 percent compared to target-trained models, demonstrating practical generalization capability across dynamic edge environments without extensive retraining.
The authors noted that practical deployment in real-world infrastructure would entail operating the Graph Attention Network predictor at regional edge controllers while executing the lightweight reinforcement learning agent on local Kubernetes or KubeEdge clusters. However, the study acknowledged several operating boundaries, pointing out that current evaluations were restricted to topologies of up to one hundred edge nodes with candidate neighbor degrees capped at five. Future investigations will evaluate ultra-dense networks, microservice dependency graphs, and secure authentication mechanisms to identify the precise threshold where cooperative metadata synchronization costs intersect with cloud-fallback latency savings.
Subject of Research: Cooperative edge service caching and task offloading in mobile edge computing
Article Title: SLA-aware cooperative service caching in MEC via graph-attention demand prediction and migration-constrained reinforcement learning
Article References: Adhikari, J. D., & Mishra, K. (2026). SLA-aware cooperative service caching in MEC via graph-attention demand prediction and migration-constrained reinforcement learning. Discover Internet of Things, 6(1), Article 258. https://doi.org/10.1007/s43926-026-00550-0
Image Credits: AI Generated
DOI: 10.1007/s43926-026-00550-0
Keywords: mobile edge computing, service caching, graph attention networks, deep reinforcement learning, cooperative service discovery, cloud fallback, service level agreements, edge server topology, task offloading, cache replacement, latency reduction, energy efficiency, container migration, dueling double dqn, workload forecasting, queue dynamics
News Source: Marilyn Langley. (November 13, 2026). Cooperative Edge Computing Framework Cuts Delay and Energy Using Graph Prediction. Scienmag.
