def set_evaluation_feedback(self, feedbacks):
"""Inform optimizer of outcome of a rollout with current weights."""
return_ = check_feedback(feedbacks, compute_sum=True)
# Transform reward (e.g. to a log-scale)
return_ = self.value_transform(return_)
self.bayes_opt.update(self.parameters, return_)
def set_evaluation_feedback(self, feedbacks):
"""Inform optimizer of outcome of a rollout with current weights."""
return_ = check_feedback(feedbacks, compute_sum=True)
# Transform reward (e.g. to a log-scale)
return_ = self.value_transform(return_)
self.bayes_opt.update(self.parameters, return_)
def set_evaluation_feedback(self, feedbacks):
"""Inform optimizer of outcome of a rollout with current weights."""
return_ = check_feedback(feedbacks, compute_sum=True)
# Transform reward (e.g., to a log-scale)
return_ = self.value_transform(return_)
self.bayes_opt.update(np.hstack((self.context, self.parameters)),
return_)
if self.policy is not None:
# Policy derived from internal model is no longer valid as the data
# has changed
self.policy_fitted = False
def set_evaluation_feedback(self, feedbacks):
"""Inform optimizer of outcome of a rollout with current weights."""
return_ = check_feedback(feedbacks, compute_sum=True)
# Transform reward (e.g., to a log-scale)
return_ = self.value_transform(return_)
self.bayes_opt.update(np.hstack((self.context, self.parameters)),
return_)
if self.policy is not None:
# Policy derived from internal model is no longer valid as the data
# has changed
self.policy_fitted = False
def set_evaluation_feedback(self, feedbacks):
"""Inform optimizer of outcome of a rollout with current weights."""
self.reward = check_feedback(feedbacks, compute_sum=True)
self.history_theta.append(self.params)
self.history_R.append(self.reward)
self.it += 1
if self.it % self.train_freq == 0:
theta = np.asarray(self.history_theta)
R = np.asarray(self.history_R)
d = solve_dual_reps(R, self.epsilon, self.min_eta)[0]
self.policy_.fit(None, theta, d)
self.logger.info("Reward %.6f" % self.reward)
if self.reward > self.max_return:
self.max_return = self.reward
self.best_params = self.params
def set_evaluation_feedback(self, feedback):
"""Set feedbacks for the parameter vector.
Parameters
----------
feedback : list of float
feedbacks for each step or for the episode, depends on the problem
"""
k = self.it % self.n_samples_per_update
self.fitness[k] = check_feedback(feedback, compute_sum=True)
if self.maximize:
self.fitness[k] *= -1
if self.fitness[k] <= self.best_fitness:
self.best_fitness = self.fitness[k]
self.best_fitness_it = self.it
self.best_params[:] = self.samples[k]
self.it += 1
if (self.it - self.initial_it) % self.n_samples_per_update == 0:
self._update(self.samples, self.fitness)
def test_check_feedback_inf():
feedbacks = [0, 1, np.inf]
assert_true(
np.isinf(check_feedback(feedbacks, compute_sum=True, check_inf=False)))
assert_raises(ValueError, check_feedback, feedbacks)
def test_check_feedback_nan():
feedbacks = [0, 1, np.nan]
assert_true(
np.isnan(check_feedback(feedbacks, compute_sum=True, check_nan=False)))
assert_raises(ValueError, check_feedback, feedbacks)