def _optimize(self, acquisition: Acquisition , context_manager: ContextManager) -> Tuple[np.ndarray, np.ndarray]:
"""
See AcquisitionOptimizerBase._optimizer for parameter descriptions.
Optimize an acqusition function using a GA
"""
# initialize population of strings
random_design = RandomDesign(self.space)
population = random_design.get_samples(self.population_size)
# clac fitness for current population
fitness_pop = acquisition.evaluate(population)
standardized_fitness_pop = fitness_pop / sum(fitness_pop)
# initialize best location and score so far
X_max = population[np.argmax(fitness_pop)].reshape(-1,1)
acq_max = np.max(fitness_pop).reshape(-1,1)
iteration_bests=[]
_log.info("Starting local optimization of acquisition function {}".format(type(acquisition)))
for step in range(self.num_evolutions):
_log.info("Performing evolution step {}".format(step))
# evolve populations
population = self._evolve(population,standardized_fitness_pop)
# recalc fitness
fitness_pop = acquisition.evaluate(population)
standardized_fitness_pop = fitness_pop / sum(fitness_pop)
# update best location and score (if found better solution)
acq_pop_max = np.max(fitness_pop)
iteration_bests.append(acq_pop_max)
_log.info("best acqusition score in the new population".format(acq_pop_max))
if acq_pop_max > acq_max[0][0]:
acq_max[0][0] = acq_pop_max
X_max[0] = population[np.argmax(fitness_pop)]
# if dynamic then keep running (stop when no improvement over most recent 10 iterations)
stop = False
i=self.num_evolutions
while not stop:
_log.info("Performing evolution step {}".format(step))
# evolve populations
population = self._evolve(population,standardized_fitness_pop)
# recalc fitness
fitness_pop = acquisition.evaluate(population)
standardized_fitness_pop = fitness_pop / sum(fitness_pop)
# update best location and score (if found better solution)
acq_pop_max = np.max(fitness_pop)
iteration_bests.append(acq_pop_max)
_log.info("best acqusition score in the new population".format(acq_pop_max))
if acq_pop_max > acq_max[0][0]:
acq_max[0][0] = acq_pop_max
X_max[0] = population[np.argmax(fitness_pop)]
if acq_max[0][0]==max(iteration_bests[:-10]):
stop=True
# also stop if ran for 100 evolutions in total
if i==100:
stop=True
i+=1
# return best solution from the whole optimization
return X_max, acq_max
def _optimize(self, acquisition: Acquisition, context_manager: ContextManager) -> Tuple[np.ndarray, np.ndarray]:
"""
Implementation of abstract method.
Taking into account gradients if acquisition supports them.
See AcquisitionOptimizerBase._optimizer for parameter descriptions.
See class docstring for implementation details.
"""
x_min, fx_min = None, None
for i in range(self.batch_size):
acquisition.reset(acquisition.model)
x_min_i, fx_min_i = super()._optimize(acquisition, context_manager)
if x_min is None:
x_min, fx_min = x_min_i, fx_min_i
else:
x_min, fx_min = np.concatenate((x_min, x_min_i), axis=1), fx_min + fx_min_i
return x_min, fx_min
def _optimize(self, acquisition: Acquisition, context_manager: ContextManager) -> Tuple[np.ndarray, np.ndarray]:
"""
Implementation of abstract method.
Taking into account gradients if acquisition supports them.
See AcquisitionOptimizerBase._optimizer for parameter descriptions.
See class docstring for implementation details.
"""
# Take negative of acquisition function because they are to be maximised and the optimizers minimise
f = lambda x: -acquisition.evaluate(x)
# Context validation
if len(context_manager.contextfree_space.parameters) == 0:
_log.warning("All parameters are fixed through context")
x = np.array(context_manager.context_values)[None, :]
return x, f(x)
if acquisition.has_gradients:
def f_df(x):
f_value, df_value = acquisition.evaluate_with_gradients(x)
return -f_value, -df_value
else:
f_df = None
optimizer = self._get_optimizer(context_manager)
anchor_points_generator = ObjectiveAnchorPointsGenerator(self.space, acquisition, self.num_anchor_points)
# Select the anchor points (with context)
anchor_points = anchor_points_generator.get(num_anchor=1, context_manager=context_manager)
_log.info("Starting gradient-based optimization of acquisition function {}".format(type(acquisition)))
optimized_points = []
for a in anchor_points:
optimized_point = apply_optimizer(optimizer, a, space=self.space, f=f, df=None, f_df=f_df,
context_manager=context_manager)
optimized_points.append(optimized_point)
x_min, fx_min = min(optimized_points, key=lambda t: t[1])
return x_min, -fx_min