This is the official code for the paper "A Value-Function-based Interior-point Method for Non-convex Bi-level Optimization" (ICML2021) and its improved version "Value-Function-based Sequential Minimization for Bi-level Optimization"
""
This is an efficient algorithm for solving bilevel optimization problems.
Our code runs on the windows platform based on Python 3.6. You can simply execute the following command to automatically configure the environment
pip install -r requirement.txt
We provide examples for optimistic BLO.
You can adjust the algorithm through parameter setting. We will give the default setting in the following example.
The results will be saved in ./result_{time.strftime("%Y_%m_%d_%H_%M_%S")}.csv
python demo.py
--y_size 2 #Lower level problem dimension
--z_loop 50
--y_loop 25
--x_loop 500
--z_lr 0.01
--y_lr 0.01
--x_lr 0.01
--z_L2_reg 0.01 #\mu
--y_L2_reg 0.01 #\theta
--y_ln_reg 0.001 #\sigma^(1)
--x0 0.
--y0 0.
In addition, you can also manually change the lower objective f def lf(x,y) and upper objective F def uF(x,y) in the code to test the functions you need
def lf(x, y):
out = 0
for i in range(args.y_size):
out = out + torch.sin((x + y[i] - C[i]))
return out
def uF(x, y):
return torch.norm(x - a) ** 2 + torch.norm(y - a - C) ** 2
If you find our work useful in your research please consider citing our paper:
@inproceedings{liu2021value,
title={A Value-Function-based Interior-point Method for Non-convex Bi-level Optimization},
author={Liu, Risheng and
Liu, Xuan and
Yuan, Xiaoming and
Zeng, Shangzhi and
Zhang, Jin},
journal={arXiv preprint arXiv:2106.07991},
booktitle = {ICML},
year={2021}
}
@misc{liu2021valuefunctionbased,
title={Value-Function-based Sequential Minimization for Bi-level Optimization},
author={Risheng Liu and Xuan Liu and Shangzhi Zeng and Jin Zhang and Yixuan Zhang},
year={2021},
eprint={2110.04974},
archivePrefix={arXiv},
primaryClass={cs.LG}
}