VABO: Violation-Aware Bayesian Optimization

This repo implements a version of the violation-aware Bayesian optimization. Violation-aware Bayesian optimization is a derivative-free and sample-efficient method that automatically tunes parameters of a system to optimize a performance/cost metric while keeping the cost of constraint violation under prescribed budgets.

If you use the repo for your research, we appreciate it that you can cite the paper:

@article{xu2021vabo, title={VABO: Violation-Aware Bayesian Optimization for Closed-Loop Control Performance Optimization with Unmodeled Constraints}, author={Xu, Wenjie and Jones, Colin N and Svetozarevic, Bratislav and Laughman, Christopher R and Chakrabarty, Ankush}, journal={arXiv preprint arXiv:2110.07479}, year={2021} }

Install

• git clone https://github.com/JackieXuw/VABO_violation_aware_Bayesian_optimization.git
• Run pip install . in this directory.

Usage

• The source code is under ./vabo.
• You can run the interactive notebook under ./examples to learn how to use VABO.
• To construct an optimization problem. You need to provide a config dictionary to initialize an object of the class vabo.optimization_problem.OptimizationProblem.
  • For example.

    config = {
    'problem_name': 'my_problem',   # the name of your problem
    'var_dim': 5,   # the dimension of input variables
    'discretize_num_list': [5, 5, 5, 5, 5],  # the discretizing number you want along each direction
    'num_constrs': 1,   # the number of constraints
    'bounds': [(100, 300), (300, 500), (300, 500), (300, 500), (100, 200)],  # the bounds for the input variables
    'obj': <function>,  # the optimization objective functions
    'constrs_list': <function>,  # the optimization constraints functions
    'vio_cost_funcs_list': <function>,  # the function mapping violations to the cost
    'vio_cost_funcs_inv_list': <function>,  # the inverse mapping from cost to violation
    'init_safe_points': array([[300, 500, 500, 500, 200]]),  # an array of initial safe points
    'train_X': array([[300, 500, 500, 500, 200]]),  # an array of training points to learn the kernel
    'kernel': [<GPy.kern.src.rbf.RBF>,
               <GPy.kern.src.rbf.RBF>],  # the array of kernel functions to model the unknown functions
    'single_max_budget': 1600  # the maximum allowed violation cost in one single step
    }