Modern digital services are typically composed of multiple interconnected components,

each dedicated to a specific functionality. These components can be dynamically instanti-

ated, terminated, and scaled on demand across the available network nodes. Data traffic

is intelligently directed to interconnect these components, thereby forming diverse end-

to-end (E2E) service chains. While such service function chaining enables flexible and

efficient service deployment, it also presents a range of emerging challenges. These in-

clude determining the optimal placement of service functions, efficiently routing traffic to

avoid new bottlenecks—particularly in scenarios where traditional IP traffic coexists—and

selecting alternative components or paths in the event of failures. This study addresses

several of these challenges by proposing innovative solutions across a variety of network

scenarios.

ˆ A Centralised Solution for SFC in Hybrid Networks: Unlike existing state-

of-the-art studies that typically handle Service Function Chaining (SFC) flows in-

dividually, this study proposes a centralised solution that jointly considers all SFC

flows along with background traffic. By addressing resource competition in a holis-

tic manner, the proposed approach reduces Maximum Link Utilisation (MLU) and

Virtual Network Function (VNF) Maximum Utilisation (VMU) by 30–40% com-

pared with traditional shortest path algorithms. Compared with optimal solutions

obtained through Mixed Integer Linear Programming, this method reduces compu-

tation time from several days to tens of seconds while achieving performance that

closely approximates the optimal results.

ˆ A Unified Model for Diverse Network Failures: This study introduces a

novel unified model that integrates Graph Neural Network with Deep Reinforcement

Learning to enhance resilience against unpredictable link and VNF node failures.

The proposed algorithm is capable of rapidly selecting alternative VNF nodes and

routing paths while maintaining performance levels comparable to those achieved by

the centralized solution prior to failures. This unified model eliminates the need to

train tens or hundreds of separate models for different failure scenarios; instead, a

single trained model can generalize across all unpredictable failure cases.

ˆ A Cluster-Based Distributed Approach for Balancing Performance and

Privacy: Existing solutions are typically either fully distributed or centralized, mak-

ing them inflexible for deployment across different scenarios. To address this limi-

tation, this study proposes an innovative cluster-based framework with user-defined

clustering schemes, introduces information-sharing mechanisms among clusters, al-

lowing it to be adapted to fully distributed, centralized, or hybrid cluster-based

deployments. The clustering approach provides a balanced trade-off between net-

work information privacy and overall system performance. Compared with SOTA

benchmarks, this approach achieves approximately 20% better performance in the

same scenario.