def build_scenario(self):
"""Returns agent and env according to provided params."""
env = college_admission.CollegeAdmissionsEnv(
user_params=self.env_config)
if self.agent_type == 'robust':
agent = college_admission_jury.RobustJury(
action_space=env.action_space,
reward_fn=(lambda x: 0),
observation_space=env.observation_space,
group_cost=env.initial_params.group_cost,
subsidize=env.initial_params.subsidize,
subsidy_beta=env.initial_params.subsidy_beta,
gaming_control=env.initial_params.gaming_control,
epsilon_greedy=self.epsilon_greedy,
initial_epsilon_prob=self.initial_epsilon_prob,
decay_steps=self.decay_steps,
epsilon_prob_decay_rate=self.epsilon_prob_decay_rate,
burnin=self.burnin)
elif self.agent_type == 'static':
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
reward_fn=(lambda x: 0),
observation_space=env.observation_space,
threshold=0,
epsilon_greedy=self.epsilon_greedy,
initial_epsilon_prob=self.initial_epsilon_prob,
epsilon_prob_decay_rate=self.epsilon_prob_decay_rate,
decay_steps=self.decay_steps,
freeze_classifier_after_burnin=True,
burnin=self.burnin)
elif self.agent_type == 'continuous':
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
reward_fn=(lambda x: 0),
observation_space=env.observation_space,
threshold=0,
epsilon_greedy=self.epsilon_greedy,
initial_epsilon_prob=self.initial_epsilon_prob,
epsilon_prob_decay_rate=self.epsilon_prob_decay_rate,
freeze_classifier_after_burnin=False,
decay_steps=self.decay_steps,
burnin=self.burnin)
else:
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
reward_fn=(lambda x: 0),
observation_space=env.observation_space,
threshold=self.agent_threshold,
epsilon_greedy=self.epsilon_greedy,
decay_steps=self.decay_steps,
initial_epsilon_prob=self.initial_epsilon_prob,
epsilon_prob_decay_rate=self.epsilon_prob_decay_rate)
return env, agent
def test_agent_returns_correct_threshold(self):
env = college_admission.CollegeAdmissionsEnv(
user_params={
'gaming':
False,
'subsidize':
False,
'noise_params':
params.BoundedGaussian(max=0.3, min=0, sigma=0, mu=0.1),
'feature_params':
params.GMM(
mix_weight=[0.5, 0.5], mu=[0.5, 0.5], sigma=[0.1, 0.1])
})
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0,
burnin=9,
freeze_classifier_after_burnin=True)
test_util.run_test_simulation(env=env,
agent=agent,
num_steps=10,
stackelberg=True)
learned_threshold = env.history[-1].action['threshold']
self.assertTrue(np.isclose(learned_threshold, 0.55, atol=1e-2))
def test_simple_classifier_simulation_runs_successfully(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
test_util.run_test_simulation(env=env, agent=agent, stackelberg=True)
def test_jury_successfully_initializes(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
self.assertEqual(agent.initial_action()['threshold'], 0.7)
self.assertEqual(agent.initial_action()['epsilon_prob'], 0)
def test_label_fn_returns_correct_labels(self):
"""Checks that the label function works as expected."""
observations = {
'test_scores_y': [0.2, 0.3, 0.4, 0.5, 0.6],
'selected_ground_truth': [1, 0, 2, 1, 2],
'selected_applicants': [1, 1, 0, 1, 0]
}
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
labels = agent._label_fn(observations)
self.assertListEqual(labels, [1, 0, 1])
def test_get_default_features_returns_same_features(self):
"""Checks that the feature selection fn works as expected."""
observations = {
'test_scores_y': [0.2, 0.3, 0.4, 0.5, 0.6],
'selected_ground_truth': [1, 0, 2, 1, 2],
'selected_applicants': [1, 1, 0, 1, 0]
}
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
features = agent._get_default_features(observations)
self.assertListEqual(features, [0.2, 0.3, 0.5])
def test_agent_returns_same_threshold_till_burnin_learns_and_freezes(self):
"""Tests that agent returns same threshold till burnin and freezes after."""
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.NaiveJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.3,
burnin=4,
freeze_classifier_after_burnin=True)
test_util.run_test_simulation(env=env,
agent=agent,
num_steps=10,
stackelberg=True)
actions = [float(action['threshold']) for _, action in env.history]
self.assertEqual(set(actions[:4]), {0.3})
self.assertLen(set(actions), 3)