BOHF Project

Description

This project evaluates the BOHF proposed algorithm (referred to as MR-LPF in the associated paper) compared to the baseline MaxMinLCB.

---

Installation

To install the required libraries, use the following command:

pip install -r requirements.txt

---

Algorithm Options (--algo)

• BOHF
• Max_Min_LCB

Preference Function Options (--preference_function)

• RKHS
• ackley
• yelp

Base Kernel Options (--kernel)

• RBF
• Matern

---

Parameters

• --beta: Confidence interval width.
• --n_runs: Number of sequential runs.
• --alpha_gp: Regularization hyperparameter for Gaussian Process (GP) regression.
• --length_scale: Length scale of the kernel
• --lambda_reg: L2 loss penalty
• --learning_rate: Learning rate for gradient descent.
• --n_iterations: Time horizon (T).
• --smoothness: Smoothness parameter of the kernel.

---

Examples

Synthetic Functions (RKHS and Ackley)

CPU Version (the one used in the paper with scikit learn library):

python clean_test_algos_synthetic_cpu.py --alpha_gp 0.05 --length_scale 0.1 --lambda_reg 0.05 --beta 1 --learning_rate 0.01 --n_iterations 2000 --n_runs 1 --algo Max_Min_LCB --kernel Matern --smoothness 1.5 --seed 0 --preference_function ackley

GPU Version:

python clean_test_algos_full_gpu_optim_fix_synthetic.py --alpha_gp 0.05 --length_scale 0.1 --lambda_reg 0.05 --beta 1 --learning_rate 0.005 --n_iterations 300 --n_runs 1 --algo BOHF --kernel RBF --seed 0 --preference_function RKHS

YELP Dataset (GPU version with botorch library used in the paper)

python clean_test_algos_yelp_full_gpu_optim_fix.py --alpha_gp 0.1 --length_scale 0.1  --lambda_reg 0.05 --beta 2 --learning_rate 0.01 --n_iterations 2000 --n_runs 1 --algo Max_Min_LCB --lr_decay 0 --kernel RBF --smoothness 1 --seed 24 --preference_function yelp