Code for Double Check Your State Before Trusting It: Confidence-Aware Bidirectional Offline Model-Based Imagination (NeurIPS 2022). The code is highly based on the offlineRL repository.
git clone https://agit.ai/Polixir/neorl.git
cd neorl
pip install -e .pip install git+https://github.com/rail-berkeley/d4rl@master#egg=d4rlFor more details on use, please see d4rl.
pip install -e .CUDA_VISIBLE_DEVICES=5 nohup python examples/train_d4rl.py --algo_name=td3_bc --task d4rl-halfcheetah-medium-v0 --seed 45 --real-data-ratio 0.7 --horizon 3 --log-dir ./logs/td3-bc-cabi/halfcheetah-medium-v0/r45 > out.log 2>&1 &In this part, we present the results of CABI+TD3_BC on MuJoCo "-v2" datasets. The results are averaged over 4 random seeds.
| Task Name | TD3_BC | TD3_BC+CABI |
|---|---|---|
| halfcheetah-random-v2 | 11.0 |
14.3 |
| hopper-random-v2 | 8.5 |
15.7 |
| walker2d-random-v2 | 1.6 |
6.0 |
| halfcheetah-medium-v2 | 48.3 |
50.1 |
| hopper-medium-v2 | 59.3 |
60.1 |
| walker2d-medium-v2 | 83.7 |
84.3 |
| halfcheetah-medium-replay-v2 | 44.6 |
46.2 |
| hopper-medium-replay-v2 | 60.9 |
70.1 |
| walker2d-medium-replay-v2 | 81.8 |
83.5 |
| halfcheetah-medium-expert-v2 | 90.7 |
94.8 |
| hopper-medium-expert-v2 | 98.0 |
111.4 |
| walker2d-medium-expert-v2 | 110.1 |
110.5 |
| halfcheetah-expert-v2 | 96.7 |
96.9 |
| hopper-expert-v2 | 107.8 |
109.3 |
| walker2d-expert-v2 | 110.2 |
110.5 |
If you use our method or code in your research, please consider citing the paper as follows:
@inproceedings{lyu2022doublecheck,
title={Double Check Your State Before Trusting It: Confidence-Aware Bidirectional Offline Model-Based Imagination},
author={Jiafei Lyu and Xiu Li and Zongqing Lu},
booktitle={Thirty-sixth Conference on Neural Information Processing Systems},
year={2022}
}
I do have noticed that the reported results in our original paper maybe out-of-date. I am now working on the experimental results on MuJoCo-v2 datasets, and Adroit-v1 datasets. I will append the results in this repo when the experiments are done.
2023/6/8, update results on mujoco-v2 datasets