SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning

July 16, 2026 · View on GitHub

Paper Dataset Python

中文 README

SABR (Stable Adaptive Bitrate) is a learning-based adaptive bitrate (ABR) framework designed for stable policy training and robust evaluation across diverse network traces. It combines behavior cloning pretraining with reinforcement learning fine-tuning, and provides a C++-accelerated ABR simulation environment, learning-based baselines, rule-based baselines, and ABRBench-compatible evaluation scripts.

The paper is available from IEEE Xplore: SABR: A Stable Adaptive Bitrate Framework Using Behavior Cloning Pretraining and Reinforcement Learning Fine-Tuning.

SABR adopts a two-stage training framework (Pretraining + Fine-tuning):

  1. Behavior Cloning (BC) Pretraining
    • Trained on demonstration data generated by expert policies (e.g., MPC with beam search), using DPO (Direct Preference Optimization) for efficient and stable imitation learning.
  2. Reinforcement Learning (RL) Fine-tuning
    • Starting from the pretrained model, applies PPO (Proximal Policy Optimization) for deeper exploration and policy improvement.

SABR overview

⚙️ Features

1️⃣ Standardized Gym Interface

  • The simulation environment is implemented following the OpenAI Gym (or Gymnasium) interface standard.
  • It is theoretically compatible with various RL algorithms in Stable Baselines3, including: ✅ PPO ✅ SAC ✅ DQN ✅ TD3 ✅ A2C, etc.
  • Supports VecEnv parallel simulation environments for improved training efficiency.

2️⃣ High-Performance C++ Simulation Core

  • The ABR simulation environment (for training) is implemented in C++ and wrapped as a Python library via pybind11.
  • Compared with a pure Python implementation, it offers higher simulation efficiency.

Prerequisites

  • OS: Ubuntu / CentOS
  • Python: 3.10
  • Install dependencies: 
    pip install -r requirements.txt

Dataset: ABRBench

  1. Download ABRBench.
  2. Place the unpacked folder in the project root and keep the name exactly ABRBench (the path must be ./ABRBench). Example: 
    SABR/
├── ABRBench/              # dataset folder name must be exactly this
├── build_env_c_plus/
├── config.py
├── train_sabr.py
└── ...

Train & Evaluate: SABR

Step 1 — Prepare the dataset

See the section above: Dataset: ABRBench.

Step 2 — Configure dataset options

  • In config.py, set _DATASET to one of:
    • ABRBench-3G, or
    • ABRBench-4G+
  • In build_env_c_plus/config.h, set DATASET_OPTION to:
    • 20 (for the ABRBench-3G mixed suite), or
    • 30 (for the ABRBench-4G+ mixed suite)

Step 3 — Rebuild the C++ environment

Recompile after every change to config.h:

cd build_env_c_plus
bash build_all.sh

Step 4 — Train and auto-evaluate

python train_sabr.py

After training, the script will automatically evaluate on each test set and OOD set, and write out the results.

Learning-based Baselines: Comyco & Pensieve

Steps 1/2/3 are the same as SABR (dataset, config, build).

Step 4 — Train:

# Comyco
python train_comyco.py

# Pensieve
python train_pensieve.py

After training, each baseline is automatically evaluated on all test and OOD sets.

Rule-based Baselines

Included: QUETRA / BOLA / BB / RobustMPC, etc.

  • Step 1 is the same as SABR (prepare the dataset).
  • Step 2 (important difference):
    Rule-based methods cannot select the mixed suites ABRBench-3G / ABRBench-4G+ directly.
    You must choose a single trace set and test them one by one.
    • Example for FCC-18:
      • In config.py: set _DATASET = 'FCC-18'
      • In build_env_c_plus/config.h: set DATASET_OPTION = 2
  • Step 3 is the same as SABR (rebuild after config.h changes): 
    cd build_env_c_plus
bash build_all.sh
  • Step 4 — Run all rule-based baselines: 
    bash ex_rule_baseline.sh

Optional: Lower-bound / Oracle-style analyses

These three use future bandwidth and serve as lower bounds or optimal solutions for case studies:

# 1) Beam-search based lower bound
python run_bs_mpc.py bs

# 2) MPC-based lower bound
python run_bs_mpc.py mfd

# 3) Dynamic programming (dp_my): per-case optimal solution (see Pensieve paper)
./dp_my

Note: dp_my may fail on some trace sets (you can skip it if needed).

QoE Results

After running rule-based (and learning) methods, use the plotting script to compare QoE across schemes. Select which schemes to display via SCHEMES in plot_results.py, then:

python plot_results.py

QoE results

FAQ

  • Changed config.h but nothing happened?
    You need to rebuild:

    cd build_env_c_plus && bash build_all.sh

    ⚠️ If you change config.h (C++ side) but do not update config.py accordingly, your training/evaluation and plotting may use inconsistent QoE parameters.

  • Does plot_results.py use QoE parameters from config.py?
    Yes.

    • plot_results.py reads the bitrate levels and rebuffer penalty for the current _DATASET.
    • These parameters come from the _DATASET_OPTION entry in config.py.
    • If you change VIDEO_BIT_RATE or REBUF_PENALTY in config.py, the QoE values computed by plot_results.py will change accordingly.

Citation

If SABR or ABRBench is useful for your research, please cite:

@inproceedings{luo2025sabr,
  title={Sabr: A stable adaptive bitrate framework using behavior cloning pretraining and reinforcement learning fine-tuning},
  author={Luo, Pengcheng and Zhao, Yunyang and Zhang, Bowen and Yang, Genke and Soong, Boon-Hee and Yuen, Chau},
  booktitle={2026 IEEE Wireless Communications and Networking Conference (WCNC)},
  pages={1--6},
  year={2026},
  organization={IEEE}
}

References & Acknowledgments