Trust-Based Adaptive LQR Control for Autonomous Multi-Tenant Network Management

Abstract

In multi-tenant network environments, deploying Linear Quadratic Regulators at scale is impeded not by theory but by a practical bottleneck: cost matrices require per-instance manual tuning, yet optimal parameters vary across tenants and evolve as traffic dynamics change, making per-instance configuration infeasible at large scale. Prior adaptive and gain-scheduling approaches either require offline enumeration of operating regimes [1] or designer-specified tuning parameters [2], and prior work has not addressed autonomous Q-matrix discovery from online statistical observation of plant behavior. We present a trust-based adaptation mechanism that continuously monitors traffic predictability via Coefficient of Variation analysis and autonomously maps observed statistics to LQR cost matrix parameters, selecting among provably stable controllers without any prior knowledge of tenant traffic profiles. Using NS-3 simulations with multi-phase dynamic traffic and scalability experiments across varying tenant populations, we show the system autonomously discovers the full control spectrum from uniform initialization, significantly reduces queue occupancy and latency compared to manually-tuned fixed LQR, and maintains equivalent fairness at scale. This work demonstrates that statistical plant characterization can drive zero-touch controller synthesis with formal stability guarantees, offering a practical path to autonomous LQR deployment where per-instance expert tuning is infeasible.