def extract_training_observations(
self,
game_samples: SampleData,
test_fraction,
):
train_samples, test_samples = split_train_test(game_samples,
test_fraction)
# Make policy samples for each label in (features, labels) pairs
# - Note that this scrubbed the meta info
print("\nBuilding policy training observations. Sit tight.")
train_features, train_labels = self.extract_policy_observations(
train_samples)
train_samples = SampleData(
features=train_features,
labels=train_labels,
)
test_features, test_labels = self.extract_policy_observations(
test_samples)
test_samples = SampleData(
features=test_features,
labels=test_labels,
)
return train_samples, test_samples
def extract_training_observations(
self,
samples,
test_fraction,
) -> (SampleData, SampleData):
train_samples, test_samples = split_train_test(samples, test_fraction)
if self.weighting_strat:
train_samples.weights = calculate_weights(train_samples,
self.highest_generation,
self.weighting_strat)
test_samples.weights = calculate_weights(test_samples,
self.highest_generation,
self.weighting_strat)
return train_samples, test_samples
def train(self, samples, test_fraction=.2):
raise RuntimeError("Broken")
train_set, test_set = split_train_test(samples, test_fraction, "value")
# "Train"
self.state_visits = defaultdict(int)
self.state_wins = defaultdict(int)
for features, label in train_set:
self.state_visits[tuple(features)] += 1
self.state_wins[tuple(features)] += label
# Convert them to dicts to maintain consistency with load
self.state_visits = dict(self.state_visits)
self.state_wins = dict(self.state_wins)
# delete any keys that are too infrequent
to_delete = []
for k, v in self.state_visits.items():
if v <= 5:
to_delete.append(k)
for k in to_delete:
del self.state_visits[k]
del self.state_wins[k]
# "Test"
absolute_error = 0
absolute_error_random = 0
for features, label in test_set:
value = self.predict(features)
random_value = -1.0 + (2.0 * random.random())
absolute_error += abs(label - value)
absolute_error_random += abs(label - random_value)
mean_absolute_error = absolute_error / len(test_set)
mean_absolute_error_random = absolute_error_random / len(test_set)
print("MAE:", mean_absolute_error)
print("MAE (random):", mean_absolute_error_random)