Synthetic
收藏资源简介:
Synthetic数据集由伦斯勒理工学院和加州大学洛杉矶分校的研究团队创建,用于离线强化学习研究。该数据集包含150条数据,通过数据蒸馏技术从专家政策生成的轨迹数据中提炼而来。数据集的创建旨在提高样本效率和泛化能力,适用于训练强化学习模型,特别是在需要高质量数据集以减少随机性和过拟合的场景中。
The Synthetic Dataset was created by research teams from Rensselaer Polytechnic Institute (RPI) and the University of California, Los Angeles (UCLA) for offline reinforcement learning research. This dataset contains 150 data samples, which are derived from the trajectory data generated by expert policies via data distillation technology. The dataset was developed to improve sample efficiency and generalization ability, and is suitable for training reinforcement learning models, especially in scenarios where high-quality datasets are required to reduce randomness and overfitting.
DDRL 数据集概述
数据集环境
环境设置
- 使用
conda创建和激活环境: bash conda env create -f environment.yml conda activate DDRL
环境配置
-
使用 Procgen 基准测试中的三个环境:bigfish、starpilot 和 jumper,均为 easy 模式,200 个种子。
-
环境名称在配置文件中表示为:
env=bigfish200seeds # bigfish env=jumper200seeds # jumper env=starpilot200seeds # starpilot
数据集生成步骤
1. 生成离线数据集
- 运行命令: bash python src/offline_data.py model_save_path=bf200seeds.pt env=bigfish200seeds model_class=teacher.PPOTeacherProcgen
2. 生成蒸馏数据集
- 运行命令: bash python src/train_distill_data_and_save.py env=jumper200seeds distill=exp10 distill.teacher_path=jp200seeds.pt distill.teacher_class=teacher.PPOTeacherProcgen loss=MSE student=policy_student_CNN trainablebuffer.buffer_kwargs.synthetic_buffer_size=15 trainablebuffer.buffer_kwargs.synthetic_init_threshold_size=1000 student_train_epochs=10 student_update_freq=1 distill.distill_kwargs.batch_size=75
3. 使用蒸馏数据集训练学生网络
- 运行命令: bash python src/run_offline_datadistill.py env=bigfish200seeds distill=exp100 distill.teacher_path=bf200seeds.pt distill.teacher_class=teacher.PPOTeacherProcgen loss=MAE student=policy_student_CNN trainablebuffer.buffer_kwargs.synthetic_buffer_size=5000 trainablebuffer.buffer_kwargs.synthetic_init_threshold_size=1000 student_train_epochs=50 student_update_freq=1 distill.distill_kwargs.batch_size=256
4. 评估
- 运行命令: bash python src/run_evaluation.py model_save_path=student-model__offline__starpilot200seeds__sp200seeds.pt__policy-student-CNN__CNN0__MAE__adam0__batch_buffer100__batch_trainablebuffer100__exp100.pt env=starpilot200seeds evaluator_class=evaluation.VecEvaluator
引用
-
数据集的引用信息:
@misc{light2024datasetdistillationofflinereinforcement, title={Dataset Distillation for Offline Reinforcement Learning}, author={Jonathan Light and Yuanzhe Liu and Ziniu Hu}, year={2024}, eprint={2407.20299}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2407.20299}, }




