def deduplicate(self, select_best=False):
"""De-duplicates Samples with identical structures.
Args:
select_best: Whether to select the sample with the highest reward among
samples with the same structure. Otherwise, the sample that occurs
first will be selected.
Returns:
A Population with de-duplicated samples.
"""
return Population(utils.deduplicate_samples(self, select_best=select_best))
def best_n(self, n=1, q=None, discard_duplicates=False, blacklist=None):
"""Returns the best n samples.
Note that ties are broken deterministically.
Args:
n: Max number to return
q: A float in (0, 1) corresponding to the minimum quantile for selecting
samples. If provided, `n` is ignored and samples with a reward >=
this quantile are selected.
discard_duplicates: If True, when several samples have the same structure,
return only one of them (the selected one is unspecified).
blacklist: Iterable of structures that should be excluded.
Returns:
Population containing the best n Samples, sorted in decreasing order of
reward (output[0] is the best). Returns less than n if there are fewer
than n Samples in the population.
"""
if self.empty:
raise ValueError('Population empty.')
samples = self.samples
if blacklist:
samples = self._filter(samples, blacklist)
# are unique.
if discard_duplicates and len(samples) > 1:
samples = utils.deduplicate_samples(samples)
samples = sorted(samples,
key=lambda sample: sample.reward,
reverse=True)
if q is not None:
q_value = np.quantile([sample.reward for sample in samples], q)
return Population(
[sample for sample in samples if sample.reward >= q_value])
else:
return Population(samples[:n])