Edge Computing and Multi-Agent Optimization Boost Intelligent Customer Service Allocation

A research team led by Xuanang Li and colleagues has introduced an intelligent customer service task allocation framework that integrates edge computing and multi-agent collaborative optimization within a hierarchical cloud-edge-end computing architecture. As enterprise customer interactions expand across web chats, emails, telephone assistance, mobile applications, and social media platforms, conventional centralized customer support architectures increasingly encounter severe operational bottlenecks, including elevated transmission latency, inefficient resource allocation, limited computational scalability, and poor adaptability under volatile operational conditions. Under classical administrative regimes, customer care operations frequently depended on manual planning and rigid centralized control, which introduced administrative friction, sluggish queue throughput, and inconsistent service delivery when processing large volumes of concurrent customer requests. To overcome the systemic limitations of centralized routing, the investigators formulated a distributed paradigm wherein service requests originating on end devices are evaluated and coordinated at proximal network edge nodes, while a centralized cloud layer oversees global resource optimization, historical analytical evaluation, and master model training routines.

To rigorously substantiate the operational viability of this distributed paradigm, the investigators established and analyzed the Edge-Aware Customer Service Allocation Dataset, comprising 6,770 empirical interaction records capturing operational customer service activity, agent attributes, real-time system performance statuses, and allocation outcomes. The foundational dataset reflects multifaceted operational dynamics across heterogeneous communication channels, task complexity grades, dynamic agent capabilities, and fluctuating runtime constraints. Supervision labels were established through rule-based simulations incorporating agent workload thresholds, service request priorities, and edge resource capacity profiles, ensuring consistent and reproducible operational constraints independent of circular model feedback. In configuring the experimental pipeline, the researchers partitioned the complete 6,770-record dataset into an 80 percent training partition and a 20 percent testing partition. Systematic data preprocessing was applied across all raw attributes, employing min-max feature scaling to normalize heterogeneous continuous metrics into standardized intervals, algorithmic outlier detection routines to expunge aberrant noise and inconsistent observation points, and k-nearest neighbor imputation to reconstruct missing categorical and numeric values from contextually similar interaction profiles.

Because raw customer service inputs encompass both structured metadata and unstructured natural language narratives—such as customer problem statements, incoming inquiry transcripts, and historical agent interaction logs—the architecture implements a dual-stream feature engineering pipeline combining principal component analysis and term frequency-inverse document frequency vectorization. Unstructured problem descriptions contain decisive semantic and contextual indicators that structured numeric tables cannot capture, such as issue urgency or domain specificity. The term frequency-inverse document frequency mechanism converts these textual problem statements into meaningful numerical embeddings by measuring the relative occurrence of lexical terms within specific requests against their broader distribution across the corpus, assigning greater diagnostic weight to informative, domain-specific terminology. Concurrently, principal component analysis reduces the dimensionality of correlated numerical features, stripping out redundant statistical noise while retaining principal covariance structures to minimize runtime computational overhead. Through this integrated pipeline, the unified tabular and semantic features are translated into clean, dense, structured representations ready for rapid downstream evaluation.

At the algorithmic core of the decision-making framework is a novel hybrid model designated as the Transformer-Intelligent Moth-Flame Optimization architecture, which pairs deep sequential feature representation with nature-inspired heuristic optimization. In this structure, the transformer component uses multi-head self-attention mechanisms, feedforward networks, layer normalization steps, and positional encodings to capture intricate semantic relationships and long-range contextual dependencies between incoming service requests and the specialized competencies of distributed service agents. By processing sequential service characteristics in parallel rather than through traditional recurrent operations, the transformer module constructs enriched task-agent alignment representations that serve as real-time decision signals. An agent in this framework is defined as an artificial-intelligence-enabled service-processing entity deployed directly at an edge node, tasked with addressing or coordinating assigned requests. The transformer network systematically analyzes request importance, task priority, and agent domain expertise to generate highly accurate initial mapping candidates.

To refine and balance these initial assignment candidates, the Intelligent Moth-Flame Optimization algorithm executes iterative search operations that optimize resource consumption and workload distribution across distributed network clusters. Conventional metaheuristics, such as adaptive ant colony optimization, often exhibit acute sensitivity to initial parameter choices and suffer from excessive computational overhead, while standard reinforcement learning frameworks require continuous environment interactions and extensive retraining intervals. The Intelligent Moth-Flame Optimization component addresses these operational bottlenecks by utilizing a multi-objective fitness formulation that simultaneously factors in task priority metrics, available agent resource capacity, and dynamic workload coefficients. To prevent optimization searches from becoming trapped in sub-optimal local minima, the algorithm incorporates a mathematical Lévy flight mechanism that injects long-distance stochastic jumps through heavy-tailed probability distributions governed by specific scaling factors and stability indices, thereby balancing wide-ranging global exploration with meticulous local solution exploitation.

Within the operational edge computing tier, multi-agent collaboration is formalized through a distributed Markov decision framework wherein each autonomous edge agent continuously observes its local operational environment and state variables. These state descriptions quantify real-time operating conditions, including incoming request priorities, current queue depths, local hardware resource availability, and measurable communication latency across edge interfaces. Rather than routing all telemetry back to an overburdened central controller, agents communicate through local edge-based consensus mechanisms and peer-to-peer data exchanges confined to localized network clusters. This decentralized interaction structure allows agents to dynamically evaluate local workloads, assign incoming inquiries, perform collaborative task reassignments, and balance system loads cooperatively without introducing prohibitive inter-agent communication overhead. The governing reward functions encourage rapid response times, balanced utilization across all edge processing resources, and stable queue execution even during sudden spikes in inbound inquiry volume.

Extensive computational experiments were executed in a controlled high-performance testing environment utilizing high-tier multicore processing units, dedicated graphics acceleration hardware including Nvidia RTX 4090 and A100 platforms, 64 to 128 gigabytes of system memory, high-speed solid-state storage, and a Python 3.10 programming environment. Empirical evaluation demonstrated that the hybrid model accomplished superior task allocation accuracy and operational efficiency across the designated Edge-Aware Customer Service Allocation Dataset. The proposed framework attained an overall task allocation accuracy of 95.4 percent, an intent classification accuracy of 98.7 percent, and a slot-filling F1-score of 98.9 percent. In operational systems metrics, the architecture recorded an average task assignment latency of 28.0 milliseconds, an average computational energy consumption of 13.4 joules, a load balancing index of 95.2 percent, and an overarching system stability measurement of 96.7 percent, confirming that the collaborative edge framework reliably stabilizes computational demand without degrading assignment accuracy.

The research team conducted extensive comparative benchmarks against established semantic understanding and resource scheduling architectures to evaluate the framework across diverse operational domains. Evaluated against the Air Travel Information System benchmark, the proposed model attained a slot F1-score of 98.7 percent, an intent accuracy of 97.9 percent, and an overall accuracy of 92.4 percent, outperforming baseline dual-structure models combining bidirectional gated recurrent units, convolutional neural networks, and conditional random fields with and without attention schemes. On the electric power domain dialogue framework, known as the PSCF dataset, the proposed architecture registered an 89.5 percent slot F1-score, 93.8 percent intent accuracy, and 90.4 percent overall accuracy. In scheduling evaluations utilizing customized cloud manufacturing data and server real-time performance reinforcement learning baselines, the model maintained lower load variance scores ranging from 0.046 to 0.051 compared to competing scores of 0.052 to 0.057, underscoring stable workload smoothing across varying task length conditions.

To isolate and verify the individual algorithmic contributions of the system components, the authors performed detailed ablation experiments, multi-fold cross-validation, and sensitivity tests across varying hyperparameter configurations. Ablation evaluations demonstrated that the full integration of both the transformer feature extraction pipeline and the moth-flame optimization module was essential to achieving top-tier performance, as removing either module degraded overall accuracy, intent recognition, and slot prediction metrics. Across k-fold cross-validation routines, the model exhibited consistent operational stability across distinct data folds, achieving mean metrics of 98.7 percent for slot F1-score, 98.4 percent for intent accuracy, and 98.1 percent for overall accuracy. Statistical significance was verified through paired t-test analyses comparing the integrated model against baseline implementations, which yielded p-values of 0.012 for slot F1-score, 0.015 for intent accuracy, and 0.021 for overall accuracy—all well below the standard 0.05 significance threshold—with corresponding 95 percent confidence intervals demonstrating consistent and statistically robust performance gains.

Despite these favorable experimental findings, the researchers acknowledged specific technical limitations and computational complexities that warrant consideration prior to industrial deployment. Although the moth-flame component utilizes lightweight iterative updates designed for resource-constrained edge hardware, the dual-structure system still requires careful hyperparameter tuning and incurs computational complexity during the initial transformer training phases. Furthermore, real-world deployment across large-scale, heterogeneous enterprise systems requires robust data synchronization under fluctuating network bandwidths and intermittent physical connectivity. Looking toward future developments, the authors identified several promising investigative avenues, including the integration of federated learning paradigms to preserve localized data privacy across edge clients, the implementation of adaptive real-time online learning mechanisms to accommodate changing customer behaviors dynamically, and the deployment of distributed edge intelligence frameworks to support resilient, large-scale autonomous customer care infrastructure.

Subject of Research: Intelligent customer service task allocation using edge computing and multi-agent optimization

Article Title: Edge computing and multi-agent collaboration-driven optimization model for intelligent customer service task allocation

Article References: Li, X., Wu, S., Zhang, P., Feng, Q., Li, Y., & Zhuang, S. (2026). Edge computing and multi-agent collaboration-driven optimization model for intelligent customer service task allocation. Discover Computing, 29(1), Article 792. https://doi.org/10.1007/s10791-026-10606-y

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10606-y

Keywords: edge computing, multi-agent systems, task allocation, customer service automation, Transformer models, Moth-Flame Optimization, cloud-edge-end architecture, computational latency, workload balancing, feature extraction, principal component analysis, TF-IDF embeddings, Markov decision process, Lévy flight exploration, intent accuracy, slot filling, distributed intelligence