LLM-Driven Regime-Adaptive strategy synthesis for Polymorphic Network routing.

Area of Science:

• Computer Science
• Network Engineering
• Artificial Intelligence

Background:

• Modern networks face challenges in dynamic routing due to polymorphic traffic and unpredictable topology shifts.
• Traditional Deep Reinforcement Learning and static heuristics struggle with adaptability and generalization in complex network environments.

Purpose of the Study:

• To develop a regime-adaptive control framework for dynamic network routing in polymorphic settings.
• To address the limitations of existing methods in handling unpredictable structural shifts and distributional heterogeneity in network traffic.

Main Methods:

• Proposes RASS-Evo (Regime-Adaptive Strategy Synthesis via Evolution), reformulating adaptive control as event-triggered heuristic code generation.
• Integrates a Regime Perception Module using Kullback-Leibler divergence for detecting epistemic regime shifts.
• Employs a Verifier-Guided Heuristic Repair mechanism with Large Language Models for refining routing logic.

Main Results:

• RASS-Evo demonstrates effective adaptation to unseen network regimes on Abilene and GEANT topologies.
• The framework significantly reduces computational inference overhead by approximately 60% compared to continuous synthesis baselines.
• Successfully distinguishes between epistemic regime shifts and aleatoric noise, triggering synthesis only when necessary.

Conclusions:

• RASS-Evo offers a robust and computationally efficient solution for dynamic network routing in complex, adaptive environments.
• The event-triggered heuristic code generation approach enhances plasticity and generalization capabilities.
• This method provides a promising direction for intelligent network management and control.