LLM4Band

This repository contains the source code for the paper "LLM4Band: Enhancing Reinforcement Learning with Large Language Models for Accurate Bandwidth Estimation".

How to use?

Offline training

• Download the dataset from RL4BandwidthEstimationChallenge, download the pre-trained model from huggingface(gpt2, t5, qwen).
• Split the dataset and preprocess the data (pickle format).
• Replace the model path in the code, train model: run IQL.py.

Offline testing

• Prepare offline testing scenario in validation/prepare_scenario, evaluate the model in validation/evaluate.

Online application

• Test environment: AlphaRTC

• Download link for the docker image: alphartc4band

• Download link for the test media: testmedia

• Limit port traffic, run:

    modprobe sch_netem
  
    modprobe sch_htb
  
    docker run --rm -it -v $(pwd)/LLM4Band:/app -w /app -e PYTHONPATH=/usr/lib/python3/dist-packages --name alphartc4band --cap-add=NET_ADMIN alphartc4band
  

• Entering the container, run:

      sudo /root/go/bin/comcast --device lo --target-port 8000 --target-bw 200 --latency 50 --packet-loss 1
  
      peerconnection_serverless receiver_pyinfer.json
  

• Stop:

      comcast --device lo --stop
  

• Perform the test task in another terminal：

    docker exec alphartc4band peerconnection_serverless sender_pyinfer.json
  

• Calculate the score：

      docker run --rm -v `pwd`/LLM4Band:/app -w /app/metrics --name eval alphartc4band python3 eval_network.py --dst_network_log /app/logging/webrtc.log --output /app/result/out_eval_network.json --ground_recv_rate 500 --max_delay 500
  

Citation

@inproceedings{wang2025llm4band,
  title={LLM4Band: Enhancing Reinforcement Learning with Large Language Models for Accurate Bandwidth Estimation},
  author={Wang, Zhijian and Lu, Rongwei and Zhang, Zhiyang and Westphal, Cedric and He, Dongbiao and Jiang, Jingyan},
  booktitle={Proceedings of the 35th Workshop on Network and Operating System Support for Digital Audio and Video},
  pages={43--49},
  year={2025}
}

Acknowledgments

• RL4BandwidthEstimationChallenge - dataset
• AlphaRTC - simulation platform
• NAORL, CORL, HuggingFace - tools
• BoB, Schaferct, HRCC, Pioneer - baselines