Zeroth-Order Optimization with Trajectory-Informed Derivative Estimation

This is the official implementation for the ICLR paper Zeroth-Order Optimization with Trajectory-Informed Derivative Estimation.

Project Overview

This project provides a novel optimization method that uses trajectory information to estimate derivatives without any gradient information about the function. This method demonstrates superior performance in dealing with high-dimensional, non-convex, and noisy optimization problems.

Dependencies

Python 3.6+
PyTorch 1.4+
NumPy
Matplotlib
Jax
Optax
GPy
Scipy
scikit-learn
TuRBO
BayesianOptimization

Usage

Black-Box Adversarial Attack

PYTHONPATH=$PWD python run_exp.py --syn_func MNIST_Attack --dim 784 --lb -0.3 --ub 0.3 --bo_iters 2000 --gd_iters 0 --alg zord --trials 1 --save "eps(0.3)-" --gd_lr 0.5 --seed 0
PYTHONPATH=$PWD python run_exp.py --syn_func CIFAR10_Attack --dim 1024 --lb -0.2 --ub 0.2 --bo_iters 0 --gd_iters 1000 --alg zord --trials 1 --save "eps(0.2)-" --gd_lr 0.5 --seed 5

Non-Differentiable Metric Optimization

# Need to choose the corresponding metric function in the functions/metrics.py before runing the following commands

PYTHONPATH=$PWD python run_exp.py --syn_func MetricOpt --dim 2187 --lb -0.2 --ub 0.2 --alg zord --gd_iters 800 --bo_iters 0 --trials 5 --gd_lr 0.01 --save results/CovType/f1score-
PYTHONPATH=$PWD python run_exp.py --syn_func MetricOpt --dim 2187 --lb -0.2 --ub 0.2 --alg zord --gd_iters 800 --bo_iters 0 --trials 5 --gd_lr 0.01 --save results/CovType/jaccard-
PYTHONPATH=$PWD python run_exp.py --syn_func MetricOpt --dim 2187 --lb -0.2 --ub 0.2 --alg zord --gd_iters 800 --bo_iters 0 --trials 5 --gd_lr 0.2 --save results/CovType/precision-
PYTHONPATH=$PWD python run_exp.py --syn_func MetricOpt --dim 2187 --lb -0.2 --ub 0.2 --alg zord --gd_iters 800 --bo_iters 0 --trials 5 --gd_lr 0.2 --save results/CovType/recall-

Citation

@inproceedings{
    shu2023zerothorder,
    title={Zeroth-Order Optimization with Trajectory-Informed Derivative Estimation},
    author={Yao Shu and Zhongxiang Dai and Weicong Sng and Arun Verma and Patrick Jaillet and Bryan Kian Hsiang Low},
    booktitle={The Eleventh International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=n1bLgxHW6jW}
}