The Agentic Control Framework: Integrating Knowledge Graph Retrieval and Game-Theoretic Multi-Agent Learning for Robust Process Design
Year: 2026
Type: conference-paper
Abstract: The realization of resilient autonomous systems is currently impeded by the difficulty of translating abstract design requirements into verifiable control policies that can withstand significant model uncertainty. Here we present the Agentic Control Framework, a full-stack neuro-symbolic architecture that unifies semantic reasoning with game-theoretic hardening to maximize robustness. The workflow initiates with a Large Language Model (LLM)-based agent that leverages a domainspecific Knowledge Graph to comprehend unstructured design specifications and retrieve optimal control constraints. These formal specifications characterize a high-fidelity Digital Twin, which incorporates a Reduced Order Model (ROM) integrated via Runge-Kutta (RK4) methods. Within this adaptive environment, we solve the robust control problem as a multi-agent zero-sum differential game using Deep Reinforcement Learning (DRL), where a stabilizing Proximal Policy Optimization (PPO) policy competes against an adversary injecting worst-case stochastic disturbances. We benchmark this adversarial formulation against classical $H_{\infty}$ synthesis and standard Vanilla PPO, demonstrating that the system converges to an empirical Nash Equilibrium that rivals theoretical robust control baselines. Our results indicate that grounding agentic comprehension in rigorous game-theoretic training produces self-organizing controllers that maintain operational stability under severe uncertainty, bridging the gap between theoretical robust control and adaptive autonomy.
Institutions: DTX Digital Transformation CoLAB, University of Minho
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7165688982