Designing Network Algorithms via Large Language Models
Highlights
-
Using an automated LLM workflow to assist with the design and refinement of a network algorithm.
-
The problem is framed as an RL problem. The state and policy pieces are generated by an LLM, and the reward function is the "quality of experience" metric from the Pensieve paper. This paper builds strongly off Pensieve.
-
Easily approachable workshop paper. I would have given this paper a weak accept.
Summary
Network control algorithms such as adaptive bitrate (ABR) are traditionally hand-crafted by humans to improve network performance. However, with growing networking technologies, the need for specialized ABR has grown. This paper introduces NADA, a framework that utilizes LLMs to assist in the design and refinement of novel networking algorithms. This paper explores the use of NADA in specialized networking environments, such as Starlink and 5G. The authors show that the ABR algorithms designed by LLMs can consistently outperform established baselines in diverse networking environments.
Key Contributions
-
The authors prompt an LLM to design the state and model RL aspects of a network algorithm. They have automated filters that reject anything which isn't normalized or doesn't compile
- They find non-normalized data by fuzzing inputs and seeing if the output is within a threshold
-
The authors fix the state and model aspects of the RL problem and optimize for each. They select the top 30 performers from the state and model respectively (for a total of 900 combinations) and find the best state-model pair. This assumes that two weak performers won't generate the best design.
-
The authors show that the LLM designed algorithm outperforms SOTA on various datasets such as FCC, 4G, 5G, and Starlink.
Strengths
-
The authors use the Pensieve model as a fixed anchor to find an optimal state at first, and use the the pensieve state as a fixed anchor to find an optimal model. This avoids the need to do joint optimization right off the bat and gives an adequete baseline.
-
The model selection favors 0% FNR and maximizes the true negative rate as overlooking a performant design is much worse than optimizing a suboptimal design.
Weaknesses / Questions
-
What language was this all written in? It doesn't explicitly state though based on the syntax, I would assume python.
-
If python was used, a lot of zero-shot candidates would fail the compilation checker simply because of the syntax changing minorly in the libraries. I would be interested in giving the LLM one or a few shots to fix their candidates and see how that performs (or simply using a self-debugging LLM). This would obviously explode the computational time though.
-
I feel like the paper could have used more meat, but as a workshop paper, it's fine. Most my questions were anwsered by looking through the GitHub page.
Related Work
- Pensieve paper