def test_incorporating_feedback_multiupdate_with_lambda_decay_diff_value(self):
legal_actions = get_legal_actions()
discount = 1
feature_extractor = feature_extractors.IdentityFeatureExtractor()
exploration_prob = 0
step_size = 1
decay = .2
maxGradient = 10
threshold = .001
maxGradient = 5
num_consecutive_random_actions = 0
ql = learning_agents.SARSALambdaLearningAlgorithm(legal_actions, discount,
feature_extractor, exploration_prob, step_size, threshold, decay, maxGradient,
num_consecutive_random_actions)
state_1 = {'test_feature_1' : 1}
state_2 = {'test_feature_2' : 1}
state_3 = {'test_feature_3' : 1}
state_4 = {'test_feature_4' : 1}
action = 0
reward = 1
ql.incorporateFeedback(state_1, action, reward, state_2)
ql.incorporateFeedback(state_2, action, reward, state_3)
ql.incorporateFeedback(state_3, action, reward, state_4)
actual = ql.weights
expected = collections.Counter({'test_feature_1': 1.24,
'test_feature_2': 1.2, 'test_feature_3': 1})
self.assertEquals(actual, expected)
def test_incorporating_feedback_overlapping_multiupdate_with_decay_negative(self):
legal_actions = get_legal_actions()
discount = 1
feature_extractor = feature_extractors.IdentityFeatureExtractor()
exploration_prob = 0
step_size = 1
decay = .5
maxGradient = 10
threshold = .001
maxGradient = 5
num_consecutive_random_actions = 0
ql = learning_agents.SARSALambdaLearningAlgorithm(legal_actions, discount,
feature_extractor, exploration_prob, step_size, threshold, decay, maxGradient,
num_consecutive_random_actions)
state_1 = {'test_feature_1' : 1}
# ql.weights['test_feature_2'] = 1
state_2 = {'test_feature_2' : -1}
state_3 = {'test_feature_1' : 1}
action = 0
reward_1 = 1
reward_2 = -1
ql.incorporateFeedback(state_1, action, reward_1, state_2)
ql.incorporateFeedback(state_2, action, reward_2, state_3)
actual = ql.weights
expected = collections.Counter({'test_feature_1': 1,
'test_feature_2': 0})
self.assertEquals(actual, expected)