BayesOpt OpenMDAO Driver

Installation (Windows)

pip install bayesopt
Install this package with pip install -e . (or python -m pip install -e . if using the Tonka Python).

Usage

The BayesOpt driver can be instantiated like any OpenMDAO driver:

from bayesopt_openmdao.bayesopt_optimizer import BayesoptOptimizer
# ...
top = Problem()
root = top.root = Group()
# ...
top.driver = BayesoptOptimizer()
top.driver.options["n_iterations"] = 200
top.driver.add_desvar('p1.x', lower=-40, upper=50)
top.driver.add_desvar('p2.y', lower=-20, upper=50)
top.driver.add_objective('p.f_xy')
top.setup()

Supported options will be passed through to BayesOpt. These currently include:

n_iterations - Number of iterations of BayesOpt
n_inner_iterations - maximum number of iterations to optimize the acquisition function
n_iter_relearn - Number of iterations between re-learning kernel parameters (iterations where relearning happens take longer to compute)
n_init_samples - Number of initial samples when learning the preliminary model of the target function. Each sample requires a target function evaluation.
surr_name - Name of the surrogate function

For more details, see the BayesOpt documentation.

Examples

Example code is located in the examples subdirectory.

optimizer.py - simple paraboloid optimization problem, with two independent variables.
cobyla_opt.py - same problem as optimizer.py, using COBYLA instead of BayesOpt.
rosenbrock_multidim.py - optimization problem using the rosenbrock test function, with a configurable number of independent variables (change dimensions in its main function).
rosenbrock_multidem_cobyla - same as above, using COBYLA instead of BayesOpt.
comparison.py - automated comparison test between COBYLA and BayesOpt, using a set of parameters that seems to work reasonably well. Varies the number of independent variables and maximum number of samples to evaluate (BayesOpt will always evaluate this many samples; COBYLA can complete with fewer samples if its tolerance is reached).