Graph Neural Networks for Dynamic Network Topology Optimization

Thesis Abstract

The rapid evolution of communication and sensor networks has intensified the need for adaptive network topologies that can respond to dynamic traffic patterns, failures, and changing QoS requirements. Traditional static topology optimization fails to sustain performance in heterogeneous environments, leading to higher latency, reduced throughput, and energy inefficiencies. This study addresses the problem of dynamically optimizing network topology in real time by leveraging graph neural networks (GNNs) to learn predictive control policies that reconfigure connections and routing in response to observed network states. The aim is to develop a scalable GNN-based framework that predicts optimal topology adjustments to meet predefined performance criteria, such as latency thresholds, bandwidth utilization, and fault tolerance. The research objectives are (1) to design a heterogeneous graph representation capturing nodes, links, traffic demands, and failure signals; (2) to develop a GNN model, specifically a spatio-temporal graph neural network (ST-GNN), that outputs topology modification actions and routing decisions; (3) to implement a reinforcement learning (RL) component that optimizes a composite reward function balancing delay, reliability, and energy efficiency; (4) to evaluate the framework on realistic synthetic and real-world datasets with varying scales; and (5) to compare the proposed approach against baseline heuristic and model-based optimization methods. A mixed-methods methodology is adopted. The population comprises simulated data from sparse to large-scale enterprise and metropolitan-area networks, augmented with real traffic traces from publicly available datasets such as CAIDA and MAWI. A stratified sampling approach creates three cohorts representing small, medium, and large networks, with 50, 150, and 350 networks respectively. Data collection uses a combination of synthetic network generators (random geometric graphs with power-law degree distributions) and trace-based traffic patterns to ensure diverse topologies and dynamic conditions. Instrumentation includes network simulators (ns-3 and Mininet) integrated with a custom GNN-RL pipeline, plus monitoring dashboards to capture performance metrics. Validity is addressed through cross-validation on temporal folds and ablation studies. The analytical framework comprises (i) graph construction protocols embedding node features (processing capacity, buffer occupancy), edge features (link speed, utilization, failure likelihood), and dynamic attributes (packet arrival rates, delay distributions); (ii) a spatio-temporal GNN architecture (ST-GCN with GRU/LSTM temporal units) to learn state representations and propose topology and routing actions; (iii) a proximal policy optimization (PPO) RL algorithm to optimize the action policy with a multi-objective reward incorporating latency, packet loss, energy consumption, and reconfiguration cost; (iv) an offline training phase followed by online fine-tuning in a simulated testbed; and (v) performance analysis using ANOVA and non-parametric tests to assess significance of improvements across network scales. Model specification includes loss functions combining cross-entropy for action selection with a temporal smoothness regularizer to prevent oscillatory reconfigurations; detailed hyperparameter tuning (learning rate 1e-4, batch size 32, ST-GCN layers 3–4) is reported. Expected findings indicate that the ST-GNN with RL will achieve statistically significant reductions in end-to-end latency (up to 28%), improved throughput (up to 22%), and lower energy per delivered bit (up to 18%) compared with static and heuristic baselines, particularly in networks with high variability and frequent topology changes. The study anticipates that the learned policies generalize well to unseen topologies with transfer learning enabled by graph-level embeddings and domain-adaptive training. Contributions to knowledge include demonstrating the feasibility of end-to-end learned dynamic topology optimization in ICT networks using a unified ST-GNN–RL framework, providing a reproducible benchmark suite for dynamic topology tasks, and extending graph-based control paradigms to practical, scalable network management. The study concludes that data-driven topology adaptation can outperform traditional methods in dynamic environments when paired with robust state representations and stable RL training. Recommendations emphasize extending the framework to multi-domain networks, investigating safety constraints for live deployments, and integrating with programmable data planes (P4-enabled) to realize real-time reconfigurations.