Projects

Transforming Network Service Delivery: Intelligent Reconfiguration, Optimization, and Generative AI-Driven Monitoring in Edge-Cloud Computing Ecosystem

Principle Investigator: Hojjat Baghban

Project Number: 114-2221-E-182-060, National Science and Technology Council (NSTC), Taiwan.

This research project has established an integrated research foundation for transforming conventional network and microservice management into intelligent, intent-driven, resource-aware, and increasingly autonomous service management. The research addresses a fundamental limitation of existing cloud-native and edge computing infrastructures: although platforms such as Kubernetes and Prometheus provide sophisticated mechanisms for orchestration and observability, they still rely heavily on manually defined rules, expert knowledge, predefined queries, and reactive operational procedures. Therefore, we investigated how Generative AI and Small Language Models (SLMs) can be systematically integrated with existing network services orchestration technologies to provide more adaptive and intelligent network service management. 

A central impact of this research is the development of a progression from natural-language interaction to reliable edge intelligence and ultimately to intelligent operational analysis. the proposed LEMON architecture, which we have designed and developed at ICPS research lab,  establishes the foundation of this progression by introducing an LLM-enabled monitoring paradigm in which infrastructure operators can express monitoring requirements and operational intentions using natural language. Rather than requiring operators to manually construct complex monitoring queries or understand the detailed semantics of heterogeneous telemetry, the proposed approach enables language models to interpret high-level operational intent and translate it into structured monitoring requirements. This represents a shift from conventional telemetry-centric monitoring toward intent-driven observability, providing an important foundation for future autonomous network services orchestration systems. 

This research subsequently addresses a critical challenge that emerges when such intelligence is moved from centralized cloud environments to resource-constrained edge infrastructure. Although large language models provide powerful reasoning capabilities, their computational and memory requirements can make direct deployment impractical at the edge. We designed and deployed the InverseTune methodology, which addresses this challenge by demonstrating that model reliability does not necessarily have to depend on increasing model size. Through inverse synthetic fine-tuning and schema-driven data generation, the research improves the ability of small language models to produce globally coherent and machine-interpretable structured outputs. This is particularly important for autonomous infrastructure management because AI-generated responses must not only be semantically correct but also sufficiently reliable to interact with software tools, monitoring systems, and network services orchestration platforms. The demonstrated performance of lightweight locally deployed models approaching or exceeding the performance of considerably larger cloud-based models provides evidence that specialized SLMs can serve as practical AI agents for edge environments. 

Building upon these foundations, the we designed and deployed the LLM-NSMA framework, at ICPS research group testbed, which integrates the developed concepts with Kubernetes-based edge network services and Prometheus observability infrastructure. The resulting approach combines real-time telemetry, retrieval-augmented generation (RAG), function calling, localized language models, quantization, and heterogeneous model deployment to provide intelligent monitoring and analysis under strict computational constraints. A particularly important impact is the introduction of a Mixture-of-Sizes inference strategy, in which smaller models handle frequent and computationally lightweight tasks while larger expert models are selectively activated for more complex reasoning. This demonstrates a practical approach to balancing AI capability, latency, memory consumption, and computational efficiency, which is essential for real-world edge deployments. The trajectory-based retrieval mechanism further enables the system to reuse successful historical tool-execution experiences, improving the reliability and consistency of interactions between language models and operational tools. 

Collectively, these developments represent a significant transition from reactive and expert-dependent infrastructure management toward intelligent, context-aware, and potentially autonomous service management. The research does not treat Generative AI as a replacement for existing network services monitoring technologies; rather, it establishes an approach in which LLMs and SLMs augment existing infrastructure with semantic understanding, contextual reasoning, and adaptive interaction. This integration creates a pathway in which a network operator can express a high-level operational intent, the AI system can identify and retrieve the relevant telemetry, reason about the observed system state, generate structured and machine-interpretable results, and ultimately provide the information required for intelligent orchestration and remediation.

Intelligent Service Chain Orchestration in Multi-Domain Edge Intelligence Ecosystem: Bridging Boundaries for Seamless Service Management

Principle Investigator: Hojjat Baghban

Project Number: 113-2222-E-182-001, National Science and Technology Council (NSTC), Taiwan.

The proliferation of latency-sensitive applications has driven the adoption of edge computing, where efficient and resilient service delivery relies on the integration of Service Chain Orchestration (SCO) and Network Service Lifecycle Management (NSLM). While SCO dynamically arranges virtualized network functions (VNFs) into service chains, NSLM ensures their seamless operation throughout the design, deployment, monitoring, scaling, and termination phases. Their convergence enables agile and context-aware service management but introduces challenges of scalability, coordination, and real-time decision-making across geo-distributed and resource-constrained edge environments.

This research project investigates the interplay of SCO and NSLM, with a particular focus on service chain graph (SCG) embedding, workload placement, and dynamic resource reconfiguration. First, we formulate the SCG embedding problem and develop a data-driven solution using an Actor–Critic Multi-Armed Bandit framework (A2C-SCGE) to achieve reliable, ultra-low latency, and energy-efficient service provisioning. Experimental results demonstrate that our approach outperforms baseline algorithms in terms of reward and service acceptance rates, particularly for large SCG topologies. Second, we propose a mixed-integer linear programming model for SCG resource configuration that ensures low computational overhead while achieving optimal values across networks of different scales. Building on this, we design a two-stage reconfiguration mechanism that minimizes end-to-end latency and VNF migration costs while introducing novel latency constraints between sequential functions and achieving load balancing across substrate networks. Finally, we design and validate an large language models-driven monitoring and analysis system to enable real-time service functions performance insights and adaptive orchestration decisions. The outcomes of this project advance lifecycle-aware service chain orchestration in multi-domain edgeintelligence ecosystems, contributing novel methods for workload placement, resource optimization, and configuration. Together, the contributions of this research project enhance the reliability, scalability, and adaptability of next-generation network service management.