← Research
02 / Research Project

AI Control Plane

Real-time decision-making for wireless physical-layer reconfiguration.

The physical layer of a 6G system will be far more reconfigurable than its predecessors—beams will steer, apertures will move, links will reform on millisecond timescales. Managing this reconfiguration in real time exceeds what traditional optimization algorithms can deliver: the decision space is too large and the time budget too tight. Our AI Control Plane research develops neural agents that orchestrate these physical-layer decisions continuously and autonomously. This is the software layer that makes the intelligent physical layer possible.

01

The Scale of the Problem

A room served by a PASS system with 100 dielectric waveguides, each with 10 potential pinch points, presents a discrete decision space of 10^100 configurations. Exhaustive search is impossible at any latency. Learned policies—trained offline against high-fidelity channel simulation, fine-tuned online from measurement—are the only tractable path to real-time control at this scale.

02

Neural Agent Architectures

We explore three classes of agents. Model-free reinforcement learning agents that learn from reward signals (throughput, latency, power). Model-based agents that maintain a differentiable world model of the channel and plan ahead. Hierarchical agents that separate fast beam-level decisions (sub-ms) from slower topology-level decisions (seconds to minutes). Each class makes different tradeoffs between sample efficiency and inference latency.

03

Latency-Constrained Deployment

Physical-layer decisions in 6G must complete in well under one millisecond to avoid acting on stale channel state. This hard constraint shapes every architectural choice: inference must run on edge hardware, not cloud servers. We co-design neural architectures with their inference targets, using quantization, pruning, and operator fusion to meet real-time budgets without sacrificing decision quality.

Key Concepts
Reinforcement Learning (RL)A machine learning paradigm where agents learn optimal behavior through interaction with an environment, maximizing cumulative reward.
Channel State Information (CSI)The instantaneous description of a wireless channel's amplitude and phase response across frequency, time, and space.
Beam ManagementThe process of selecting, maintaining, and switching beam directions in a directional antenna system to maintain link quality.
Domain RandomizationA sim-to-real transfer technique that exposes agents to diverse simulated conditions to improve generalization to real environments.
References
  1. [1]Luong, N. C., et al. (2019). Applications of Deep Reinforcement Learning in Communications and Networking: A Survey. IEEE Communications Surveys & Tutorials, 21(4), 3133–3174.
  2. [2]Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction. 2nd Edition, MIT Press.
  3. [3]3GPP TS 38.321. (2024). NR; Medium Access Control (MAC) Protocol Specification.
  4. [4]He, H., et al. (2020). Model-Driven Deep Learning for Physical Layer Communications. IEEE Signal Processing Magazine, 36(2), 11–22.
Get in touch →