maximize flag has no effect when f is None

[mhwalker]

Basically what title says. When using GPyOpt to optimize an externally evaluated function, the maximize flag has no effect.

I would expect this snippet to maximize based on the inputs: Snippet 1:

X_data = np.array([ [1.0, 1.5, 1.0], [1.5, 3.0, 1.0], 
                   [2.0, 5.0, 2.0]])
Y_data = np.array([ [0.7248], [0.4579],
                   [0.4192]])
bo_step = GPyOpt.methods.BayesianOptimization(f=None, domain=domain, X=X_data, Y=Y_data, maximize=True, acquisition_type='EI')
x_next = bo_step.suggest_next_locations()
x_next

array([[1.93984549, 4.90429719, 1.9519504 ]])

This snippet gives the result that I expect. Snippet 2:

X_data = np.array([ [1.0, 1.5, 1.0], [1.5, 3.0, 1.0], 
                   [2.0, 5.0, 2.0]])
Y_data = np.array([ [-0.7248], [-0.4579],
                   [-0.4192]])
bo_step = GPyOpt.methods.BayesianOptimization(f=None, domain=domain, X=X_data, Y=Y_data, maximize=False, acquisition_type='EI')
x_next = bo_step.suggest_next_locations()
x_next

array([[0.95508769, 1.42728523, 1.019358 ]])

Not sure if this is the designed behavior, but I think the naive expected behavior is that Y should be transformed to -Y internally if maximize=True

[mhwalker]

Also, I didn't follow the code fully, but I think that if Y_init is provided with maximize=True, the initialization is not correct (unless user is aware that -Y_init should be provided).

[gdikov]

GPyOpt version is 1.2.5

I also got confused by the maximize flag. The following example reproduces the unexpected behaviour:

import GPyOpt
import numpy as np
import matplotlib.pyplot as plt

def foo(val):
    return (np.sin(val * 8) * val - val ** 2 + val)

def plot_samples(samples_x, samples_y):
    xs = np.arange(-5.0, 5.0, 0.01)
    ys = foo(xs)
    plt.plot(xs, ys)
    plt.plot(samples_x, samples_y, "*r")

bounds = [{'name': 'x', 'type': 'continuous', 'domain': (-5.0, 5.0)}]
max_iter = 20

xs = np.array([[-5.0]])
ys = foo(xs)

for i in range(max_iter):
    bo_gp = GPyOpt.methods.BayesianOptimization(
        f=None, 
        domain=bounds, 
        model_type="GP", 
        acquisition_type ="EI", 
        maximize=True, 
        initial_design_numdata=len(xs),
        X=xs, 
        Y=ys,
        exact_feval=True)

    new_x = bo_gp.suggest_next_locations()
    xs = np.concatenate([xs, new_x])
    ys = np.concatenate([ys, foo(new_x)])

plot_samples(xs, ys)

resulting in:

However, if you change Y=ys to Y=-ys the optimisation produces results as expected and neither of the options maximize=False and maximize=True have any noticeable effect on the quality of the samples:

[dorian821]

I'm struggling with the same confusion. I've resorted to simply taking the negative of the objective value as mentioned above, but confused by the necessity for this.

[LarsHH]

I've also just discovered this issue after a while. Would be nice if this it was pointed out in the documentation since I've run quite a few things without being aware of this.

[F-A]

The problem is that GPyOpt changes a minimization problem into a maximization one by simply changing the sign of the evaluation function: https://github.com/SheffieldML/GPyOpt/blob/c9e214c49f57af4c7ee92d94ab859f46d65fd842/GPyOpt/methods/bayesian_optimization.py#L106

I guess there is no easy fix for this issue, but maybe one could issue a warning when maximize=True and f=None, so that users know the maximization will not be effective...