In today’s digital era, networks underpin critical communication infrastructures, yet their configuration and management remain laborious, error-prone, and heavily dependent on vendor-specific command syntaxes. This thesis introduces two complementary, AI-driven approaches that address these challenges through the integration of Natural Language Processing (NLP), Large Language Models (LLMs), and machine learning techniques.

The first part of the work presents Text2Net, a transformative framework that converts plain-text descriptions of network topologies into fully functional simulation configurations. The initial proof-of-concept demonstrates that natural language inputs can be systematically translated into structured command blueprints, thereby reducing configuration complexity and cognitive overhead. Building on this foundation, an extended version of Text2Net leverages advanced NLP methods including BERT-based classification and Retrieval-Augmented Generation (RAG) to handle a broader array of network scenarios encompassing dynamic routing protocols, switching, and security configurations. This modular system seamlessly integrates with network emulation environments, offering significant benefits for both academic instruction and professional rapid prototyping.

The second part of the thesis addresses Network Traffic Classification (NTC) by proposing a novel, hybrid machine learning framework. This framework employs a four-phased architecture that combines packet-based, payload-based, statistical, and behavioral features with lightweight classifiers such as Naive Bayes Multinomial, Random Forest, and Decision Trees. Comprehensive evaluations on a software-defined wireless network testbed reveal that the system achieves high classification accuracy even under minimal data conditions and offers detailed insights into feature importance through SHAP analysis. Notably, the framework supports fine-grained in-app service classification and robust performance on both unencrypted and encrypted traffic.

Towards natural language networking, these contributions demonstrate the transformative potential of merging NLP, LLMs, and machine learning with traditional networking challenges. By bridging the gap between human-centric language and machine-executable commands, and by enabling intelligent, interpretable network traffic analysis, this thesis lays a robust foundation for more adaptive, efficient, and user-friendly network systems.