def get_GP_optimum(obj):
"""
Finds the optimal design by maximising the mean of the surrogate
probabilistic GP model.
Parameters
----------
obj: GPyOpt object
The GPyOpt object with a surrogate probabilistic model.
"""
# Define space
space = Design_space(obj.domain, obj.constraints)
bounds = space.get_bounds()
# Specify Optimizer --- L-BFGS
optimizer = OptLbfgs(space.get_bounds(), maxiter=1000)
# Do the optimisation
x, _ = optimizer.optimize(
x0=obj.x_opt,
f=lambda d: fun_dfun(obj, space, d)[0],
f_df=lambda d: fun_dfun(obj, space, d),
)
# TODO: MULTIPLE RE-STARTS FROM PREVIOUS BEST POINTS
# Round values if space is discrete
xtest = space.round_optimum(x)[0]
if space.indicator_constraints(xtest):
opt = xtest
else:
# Rounding mixed things up, so need to look at neighbours
# Compute neighbours to optimum
idx_comb = np.array(
list(itertools.product([-1, 0, 1], repeat=len(bounds))))
opt_combs = idx_comb + xtest
# Evaluate
GP_evals = list()
combs = list()
for idx, d in enumerate(opt_combs):
cons_check = space.indicator_constraints(d)[0][0]
bounds_check = indicator_boundaries(bounds, d)[0][0]
if cons_check * bounds_check == 1:
pred = obj.model.predict(d)[0][0][0]
GP_evals.append(pred)
combs.append(d)
else:
pass
idx_opt = np.where(GP_evals == np.min(GP_evals))[0][0]
opt = combs[idx_opt]
return opt
def test_invalid_constraint(self):
space = [{'name': 'var_1', 'type': 'continuous', 'domain':(-5, 5), 'dimensionality': 1}]
constraints = [{'name': 'const_1', 'constraint': 'x[:,20]**2 - 1'}]
x = np.array([[0]])
design_space = Design_space(space, constraints=constraints)
with self.assertRaises(Exception):
design_space.indicator_constraints(x)
def test_invalid_constraint(self):
space = [{'name': 'var_1', 'type': 'continuous', 'domain':(-5, 5), 'dimensionality': 1}]
constraints = [{'name': 'const_1', 'constrain': 'x[:,20]**2 - 1'}]
x = np.array([[0]])
design_space = Design_space(space, constraints=constraints)
with self.assertRaises(Exception):
design_space.indicator_constraints(x)
def test_indicator_constraints(self):
space = [{'name': 'var_1', 'type': 'continuous', 'domain':(-5, 5), 'dimensionality': 1}]
constraints = [ {'name': 'const_1', 'constraint': 'x[:,0]**2 - 1'}]
x = np.array([[0], [0.5], [4], [-0.2], [-5]])
expected_indicies = np.array([[1], [1], [0], [1], [0]])
design_space = Design_space(space, constraints=constraints)
I_x = design_space.indicator_constraints(x)
self.assertTrue(np.array_equal(expected_indicies, I_x))
def test_indicator_constraints(self):
space = [{'name': 'var_1', 'type': 'continuous', 'domain':(-5, 5), 'dimensionality': 1}]
constraints = [ {'name': 'const_1', 'constrain': 'x[:,0]**2 - 1'}]
x = np.array([[0], [0.5], [4], [-0.2], [-5]])
expected_indicies = np.array([[1], [1], [0], [1], [0]])
design_space = Design_space(space, constraints=constraints)
I_x = design_space.indicator_constraints(x)
self.assertTrue(np.array_equal(expected_indicies, I_x))