Individual MEng project / 2026 / Reinforcement learning
Learning to share the load.
LLM-assisted multi-agent learning for edge-network resource management
- Built
- A five-node packet simulator and MARL training environment
- Owned
- Simulation, bridge, environment and reward-shaping experiments
- Hard part
- A stable reset/step contract across Simulink and EPyMARL
- Evidence
- Award-winning MEng project with public source and setup
Follow a task through the system.
Select an edge node to inspect its processing and offload paths, or explore how the agents learn.
Lines: neighbour links · Highlight: selected node’s links
Labels: configured processing rates in the base network (Table 3.1).
Edge A / task routing
Newly generated tasks, or tasks from E and B
Applies the agent’s offload probability and routing safety rules.
Tasks wait in arrival order, then process at 69,000 bit/s.
A heuristic selects between both neighbours using processing rates and link delays.
The controller drops an arriving task when the overload safety condition is met.
One-hop limit: an already-offloaded task must be processed locally, unless the overload rule drops it.
The problem
Each edge node can process incoming packets or offload them to its neighbour. Moving work out of one queue can overload the next. I modelled this as a cooperative, partially observable decision problem: five agents, local observations, and a shared reward for reducing latency and packet drops.
What I found
IQL was more consistent on the simpler network; QMIX became more useful when heterogeneous processing rates and queue capacities made coordination matter. The dissertation’s detailed comparisons focus on IQL, QMIX and a random baseline.
The LLM rates network states and contributes an auxiliary learning signal. Its effect depended on prompting, weighting and agent masking. The complex-network results were inconclusive: the best shaped run matched the best QMIX return, without establishing a general improvement.
Implementation & experimental limits
The Python environment exposes reset and step operations through the MATLAB Engine API. Each agent chooses from five offload levels: 0%, 25%, 50%, 75% or 100%. EPyMARL also has VDN, IPPO and MAPPO configurations in the public repository.
This is an abstract packet simulation, with a small number of experimental seeds. Training time constrained the search over LLM prompts and settings. Findings are from dissertation §§4.2–4.5; they do not establish performance on a physical edge network.
