18dez2025
14:00 Master's Defense fully remotely
Topic on
Intelligent, explainable, and adaptable routing in Software-Defined Networks.
Student
Yeison Stiven Murcia Calvo
Advisor / Teacher
Nelson Luis Saldanha da Fonseca - Co-advisor: Oscar Mauricio Caicedo Rendon
Brief summary
The management of communication networks has entered a new era, driven by paradigms such as Software-Defined Networks (SDN) and the application of Artificial Intelligence (AI). In particular, Reinforcement Learning (RL) and Deep Reinforcement Learning (DRL) have been adopted as powerful solutions for automating traffic routing. However, the "black box" nature of these models represents a critical obstacle to their deployment in production environments, since the lack of transparency hinders operator trust, debugging of anomalous behaviors, and informed system optimization. This dissertation addresses this challenge in two stages, progressing from interpretability to dynamic and intelligent adaptation.
The first stage introduces eXplaNet, a systematic pipeline that uses Explainable Artificial Intelligence (XAI) not only to interpret but also to actively improve routing policies. Rigorously applied to two intelligent routing models, RSIR and DRSIR, eXplaNet uses surrogate models and feature relevance techniques to identify the most influential network metrics. This process generated the enhanced static variants, XRSIR and XDRSIR, which demonstrated significant and quantifiable improvements in critical metrics such as throughput, delay, stretch, and packet loss.
Although static optimization represents a significant advance, real-world networks are inherently dynamic, which makes any fixed configuration suboptimal. To overcome this limitation, the second stage of the research introduces LARO (LLM-driven Adaptive Reward Optimizer), an innovative architecture that integrates a Large Language Model (LLM) as a second-level control module. LARO converts the static reward function into an adaptive mechanism by analyzing the network state in real time and continuously adjusting the weights of the DRL agent's reward function.
Experimental evaluation demonstrated that the LARO adaptive system consistently outperforms not only the baseline model but also the best optimized static configurations. The key to its success lies in its ability to make a "smart trade-off": the system accepts marginally longer routes to avoid congested connections, which translates into significant improvements in the most critical Quality of Service (QoS) metrics, including reduced latency and packet loss, and increased throughput.
Taken together, this dissertation traces a path from static optimization to dynamic adaptation, demonstrating how the synergy between DRL, XAI, and LLMs can solve complex problems in network management, transforming opaque models into transparent, optimized, and adaptive routing systems.
Examination Board
Headlines:
| Nelson Luis Saldanha da Fonseca | IC / UNICAMP |
| Jussara Marques de Almeida | DCC / UFMG |
| Marcelo da Silva Reis | IC / UNICAMP |
Substitutes:
| Julio Cesar dos Reis | IC / UNICAMP |
| Juliana de Santi | DAINF / UTFPR |