Optimizing Network Traffic Routing Using Deep Reinforcement Learning Algorithms

Thesis Abstract

The rapid expansion of digital communication networks has accentuated the critical need for efficient and adaptive traffic routing mechanisms to enhance network performance, reduce congestion, and ensure Quality of Service (QoS). Traditional routing algorithms often struggle to adapt to dynamic network conditions, leading to increased latency, packet loss, and suboptimal utilization of network resources. This study explores the potential of deep reinforcement learning (DRL) algorithms as a transformative approach to optimize network traffic routing. The primary aim is to develop a scalable DRL-based routing model that dynamically adapts to real-time network states to minimize packet delivery delays and improve overall throughput. The specific objectives include (1) to design and implement a DRL-based traffic routing framework suitable for large-scale networks; (2) to evaluate the performance of the proposed model against traditional routing algorithms such as OSPF and shortest path routing; (3) to analyze the impact of network load variability on route optimization; and (4) to identify the key parameters influencing the effectiveness of DRL in network routing contexts. To achieve these objectives, the research adopts a quantitative, experimental research design within a simulated network environment mirroring real-world conditions, specifically utilizing the NS-3 network simulator. The population comprises network nodes and traffic flows typical of internet backbone infrastructure, with a sample size of 200 nodes interconnected through simulated links subjected to variable traffic patterns. Data collection involves the use of synthetic traffic datasets generated under different load conditions, with the routing decisions and network performance metrics recorded at high temporal resolution. The DRL model is implemented through a Deep Q-Network (DQN) architecture, integrating convolutional neural networks to process network state information, with hyperparameters optimized using grid search procedures. The model’s training employs an epsilon-greedy policy, with reward functions defined by latency reduction, packet delivery ratio, and network throughput, and training continues until convergence over 100,000 episodes. Methodological analysis includes the use of descriptive statistics to summarize network performance metrics, followed by inferential techniques such as multivariate analysis of variance (MANOVA) to assess differences in performance among routing algorithms. To evaluate the robustness of the DRL model, sensitivity analyses are performed across varying traffic loads and network topologies. Furthermore, the study applies the framework of the Actor-Critic theory, which posits that adaptive decision-making in complex systems can be enhanced through learning-based mechanisms, providing the theoretical underpinning for the proposed DRL approach. Expected findings demonstrate that the DRL-based routing model significantly outperforms traditional algorithms in minimizing average latency, reducing packet loss, and optimizing bandwidth utilization under diverse traffic conditions, particularly in high-load scenarios. The model exhibits adaptability to fluctuating network states, maintaining stable performance and scalability in large networks. These results are anticipated to provide empirical evidence supporting the integration of DRL techniques into network management systems. The contribution of this study to knowledge lies in the development of a novel, scalable DRL-driven traffic routing framework validated through extensive simulation experiments, offering a viable solution for next-generation adaptive network management. It advances the understanding of applying machine learning algorithms to complex network optimization problems and provides a foundation for future research on real-time, self-optimizing routing protocols. In conclusion, the study recommends the deployment of DRL-based routing models in operational networks to enhance responsiveness and resilience. Future research should explore real-world implementation, evaluate the impact of different neural network architectures, and investigate cross-layer optimization strategies to further improve network performance and robustness.