skopt_obj_plot_experiment.py

from skopt import gp_minimize
from skopt.space import Real

from skopt.plots import plot_convergence
from skopt.plots import plot_objective, plot_evaluations

import numpy as np
np.random.seed(2)

# a dummy objective with 6 dimensions
dims = [Real(-1.0, 1.0, name=name) for name in ['a', 'b', 'c', 'd', 'e', 'f']]
N = len(dims)

# weights of nn that serves as objective function
a = 2.0
M = np.random.randn(N, N) * a
b = np.random.randn(N) * a
s = np.random.randn(N) * a

def fnc(x):
    # correlate the inputs
    h = np.dot(M, x) + b
    # apply non - linear function
    h = np.tanh(h)
    # final output
    h = np.dot(h, s)
    return h

# run the minimization
search_result = gp_minimize(fnc, dims, n_calls=200)

# get the search space and best x, y, model
space = search_result.space
model = search_result.models[-1]
best_x = search_result.x
best_y = search_result.fun

# sample points from search space
random_samples = np.array(space.rvs(n_samples=1000000))
best_x_sample = np.array([best_x])

# make estimations with surrogate
y_random = model.predict(space.transform(random_samples))
best_y_estimated = model.predict(space.transform(best_x_sample))[0]

# print best found actual objective value
print("Best objective value found:")
print(search_result.fun)

# print the best estimated objective
print("Surrogate value at best x:")
print(best_y_estimated)

# print the minimum objective found by random sampling from parameter space
print("Best objective value found by sampling 1000000 random points:")
print(np.min(y_random))

# visualize plots for comparison
import matplotlib.pyplot as plt
plot_objective(result=search_result)
plt.show()