def __init__(self, endog, exog, bw, var_type, fform, estimator, nboot=100):
self.endog = endog
self.exog = exog
self.var_type = var_type
self.fform = fform
self.estimator = estimator
self.nboot = nboot
self.bw = KDEMultivariate(exog, bw=bw, var_type=var_type).bw
self.sig = self._compute_sig()
def get_suggested_next_parameter(self, goods, bads):
good_model, bad_model = self.last_models
if good_model is None or bad_model is None:
good_model = KDEMultivariate(data=[
self.kde_data_encoder(self.get_trial_params(t))
for t, _ in goods
],
var_type=self.kde_vartypes,
bw=self.bandwidth_estimation)
bad_model = KDEMultivariate(data=[
self.kde_data_encoder(self.get_trial_params(t))
for t, _ in bads
],
var_type=self.kde_vartypes,
bw=self.bandwidth_estimation)
good_model.bw = np.clip(good_model.bw, self.min_bandwidth, None)
bad_model.bw = np.clip(bad_model.bw, self.min_bandwidth, None)
self.last_models = good_model, bad_model
best_score = float('-inf')
best_candidate = None
use_guided = self.guided_dice()
for _ in range(self.guided_sample_size if use_guided else self.
random_sample_size):
if use_guided:
next_param = self.guided_modify_parameter(
random.choice(goods)[0], good_model)
else:
next_param = super().get_next_parameter()
good_score = np.log(
np.clip(good_model.pdf(self.kde_data_encoder(next_param)),
1e-32, None))
bad_score = np.log(
np.clip(bad_model.pdf(self.kde_data_encoder(next_param)),
1e-32, None))
score = good_score - bad_score
if score > best_score:
best_score = score
best_candidate = next_param
log_print(Fore.LIGHTBLACK_EX + 'proposing',
'guided' if use_guided else 'sieved', 'parameter with score',
best_score)
return best_candidate