def run_hartmann6_benchmarks(D, rep, random_subspace=False):
if D == 100:
problem = hartmann6_100
elif D == 1000 and not random_subspace:
problem = hartmann6_1000
elif D == 1000 and random_subspace:
problem = hartmann6_random_subspace_1000
experiment, f = benchmark_minimize_callable(
problem=problem,
num_trials=200,
method_name='turbo',
replication_index=rep,
)
turbo1 = turbo.Turbo1(
f=f,
lb=np.zeros(D),
ub=np.ones(D),
n_init=10,
max_evals=200,
)
turbo1.optimize()
rs_str = 'random_subspace_' if random_subspace else ''
with open(f'results/hartmann6_{rs_str}{D}_turbo_rep_{rep}.json',
"w") as fout:
json.dump(object_to_json(experiment), fout)
def run_hartmann6_benchmarks(D, rep, random_subspace=False):
if D == 100:
problem = hartmann6_100
elif D == 1000 and not random_subspace:
problem = hartmann6_1000
elif D == 1000 and random_subspace:
problem = hartmann6_random_subspace_1000
experiment, f = benchmark_minimize_callable(
problem=problem,
num_trials=200,
method_name='smac',
replication_index=rep,
)
fmin_smac_nopynisher(
func=f,
x0=[0.5] * D,
bounds=[[0, 1]] * D,
maxfun=200,
rng=rep + 1,
)
rs_str = 'random_subspace_' if random_subspace else ''
with open(f'results/hartmann6_{rs_str}{D}_smac_rep_{rep}.json',
"w") as fout:
json.dump(object_to_json(experiment), fout)
def test_minimize_callable(self):
problem = BenchmarkProblem(
name="Branin",
search_space=get_branin_search_space(),
optimization_config=get_branin_optimization_config(),
)
experiment, f = benchmark_minimize_callable(problem=problem,
num_trials=20,
method_name="scipy",
replication_index=2)
res = minimize(
fun=f,
x0=np.zeros(2),
bounds=[(-5, 10), (0, 15)],
options={"maxiter": 3},
method="Nelder-Mead",
)
self.assertTrue(res.fun < 0) # maximization problem
self.assertEqual(len(experiment.trials), res.nfev)
self.assertEqual(len(experiment.fetch_data().df), res.nfev)
self.assertEqual(experiment.name, "scipy_on_Branin__v2")
with self.assertRaises(ValueError):
minimize(fun=f,
x0=np.zeros(2),
bounds=[(-5, 10), (0, 15)],
method="Nelder-Mead")
def run_hartmann6_benchmarks(D, rep, random_subspace=False):
if D == 100:
problem = hartmann6_100
elif D == 1000 and not random_subspace:
problem = hartmann6_1000
elif D == 1000 and random_subspace:
problem = hartmann6_random_subspace_1000
experiment, f = benchmark_minimize_callable(
problem=problem,
num_trials=200,
method_name='cmaes',
replication_index=rep,
)
try:
cma.fmin(
objective_function=f,
x0=[0.5] * D,
sigma0=0.25,
options={
'bounds': [[0] * D, [1] * D],
'maxfevals': 200
},
)
except ValueError:
pass # CMA-ES doesn't always terminate at exactly maxfevals
rs_str = 'random_subspace_' if random_subspace else ''
with open(f'results/hartmann6_{rs_str}{D}_cmaes_rep_{rep}.json',
"w") as fout:
json.dump(object_to_json(experiment), fout)
def run_hartmann6_benchmarks(D, rep):
if D == 100:
problem = hartmann6_100
elif D == 1000:
problem = hartmann6_1000
experiment, f = benchmark_minimize_callable(
problem=problem,
num_trials=200,
method_name='ebo',
replication_index=rep,
)
options = {
'x_range': np.vstack((np.zeros(D), np.ones(D))),
'dx': D,
'max_value': 3.32237, # Let it cheat and know the true max value
'T': 200,
'gp_sigma': 1e-7,
}
options.update(core_options)
f_max = lambda x: -f(x) # since EBO maximizes
e = ebo(f_max, options)
try:
e.run()
except ValueError:
pass # EBO can ask for more than T function evaluations
with open(f'results/hartmann6_{D}_ebo_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_branin_benchmarks(rep):
experiment, f = benchmark_minimize_callable(
problem=branin_100,
num_trials=50,
method_name='add_gp_ucb',
replication_index=rep,
)
options = Namespace(acq="add_ucb")
res = minimise_function(
f,
domain=[[-5, 10]] * 50 + [[0, 15]] * 50,
max_capital=49,
options=options,
)
with open(f'results/branin_100_addgpucb_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_branin_benchmarks(rep):
experiment, f = benchmark_minimize_callable(
problem=branin_100,
num_trials=50,
method_name='smac',
replication_index=rep,
)
fmin_smac_nopynisher(
func=f,
x0=[2.5] * 50 + [7.5] * 50,
bounds=[[-5, 10]] * 50 + [[0, 15]] * 50,
maxfun=50,
rng=rep + 1,
)
with open(f'results/branin_100_smac_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_hartmann6_benchmarks(D, rep):
if D == 100:
problem = hartmann6_100
elif D == 1000:
problem = hartmann6_1000
experiment, f = benchmark_minimize_callable(
problem=problem,
num_trials=200,
method_name='add_gp_ucb',
replication_index=rep,
)
options = Namespace(acq="add_ucb")
res = minimise_function(f,
domain=[[0, 1]] * D,
max_capital=199,
options=options)
with open(f'results/hartmann6_{D}_addgpucb_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_branin_benchmarks(rep):
experiment, f = benchmark_minimize_callable(
problem=branin_100,
num_trials=50,
method_name='turbo',
replication_index=rep,
)
turbo1 = turbo.Turbo1(
f=f,
lb=np.hstack((-5 * np.ones(50), np.zeros(50))),
ub=np.hstack((10 * np.ones(50), 15 * np.ones(50))),
n_init=10,
max_evals=50,
)
turbo1.optimize()
with open(f'results/branin_100_turbo_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_branin_benchmarks(rep):
experiment, f = benchmark_minimize_callable(
problem=branin_100,
num_trials=50,
method_name='cmaes',
replication_index=rep,
)
try:
cma.fmin(
objective_function=f,
x0=[2.5] * 50 + [7.5] * 50,
sigma0=3.75,
options={
'bounds': [[-5] * 50 + [0] * 50, [10] * 50 + [15] * 50],
'maxfevals': 50,
},
)
except ValueError:
pass # CMA-ES doesn't always terminate at exactly maxfevals
with open(f'results/branin_100_cmaes_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)
def run_branin_benchmarks(rep):
experiment, f = benchmark_minimize_callable(
problem=branin_100,
num_trials=50,
method_name='ebo',
replication_index=rep,
)
options = {
'x_range':
np.vstack((
np.hstack((-5 * np.ones(50), np.zeros(50))),
np.hstack((10 * np.ones(50), 15 * np.ones(50))),
)),
'dx':
100,
'max_value':
-0.397887, # Let it cheat and know the true max value
'T':
50,
'gp_sigma':
1e-7,
}
options.update(core_options)
f_max = lambda x: -f(x) # since EBO maximizes
e = ebo(f_max, options)
try:
e.run()
except ValueError:
pass # EBO can ask for more than T function evaluations
with open(f'results/branin_100_ebo_rep_{rep}.json', "w") as fout:
json.dump(object_to_json(experiment), fout)