1. Background and motivation
Modern mobile networks are increasingly managed by AI agents rather than fixed rule-based controllers, for example, agents that read live network telemetry (throughput, latency, energy consumption) and issue configuration commands (transmit power, scheduling weights, cell sleep states) to meet operational goals. This is the paradigm behind O-RAN’s RIC (RAN Intelligent Controller) xApps/rApps.
A single network almost never has a single objective. A realistic deployment will have:
- an energy-saving agent, trying to put cells/carriers to sleep, reduce transmit power, and consolidate traffic to save power, and
- a throughput/QoS agent, trying to maximise user data rates and minimise latency, which usually means keeping capacity on and power up.
These two agents will regularly issue conflicting commands on the same physical resources (e.g. one wants to shut down a carrier, the other wants it kept alive to serve a burst of traffic). If conflicts are resolved badly, the network can oscillate (repeatedly switching a cell on and off, flipping between configurations) which is worse for both objectives than a stable compromise.
The main research questions is:
When two AI agents with conflicting objectives share the same network, what mechanism should decide who “wins” a given conflict, and does that mechanism actually produce good, stable behaviour, or does it make things worse?
2. Objective
Build a small, controlled emulated network environment in which two AI agents, one optimising for energy, one optimising for throughput, issue commands to the same network resources. Implement two to three different conflict-resolution policies, and empirically determine which policy produces the best combined (joint) outcome and the most stable behaviour over time.
For the purpose of this project, the student is expected to use the AI Native Network Controller
3. Research questions
- RQ1 (Joint outcome): Which conflict-resolution policy (e.g., fixed priority, negotiation/bargaining, or a meta-agent/mediator) yields the best combined outcome across both objectives (energy saved and throughput/QoS delivered), compared to each agent acting alone with no coordination?
- RQ2 (Stability): Which policy minimises oscillation (i.e., avoids repeated flip-flopping between conflicting configurations over time)?
- RQ3 (Fairness / dominance): Does any policy consistently let one agent “win” at the expense of the other, and is that trade-off acceptable, tunable, or a design flaw?
- RQ4 (Robustness — stretch goal): How does the best-performing policy hold up when traffic load is bursty/unpredictable, or when one agent is given a higher-priority weight (e.g. “energy savings matter more at night”)?