def test_fixed_agent_simulation_runs_successfully(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
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 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_raises_invalid_observation_error(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
with self.assertRaises(core.InvalidObservationError):
agent.act(observation={0: 'Invalid Observation'}, done=False)
def test_agent_produces_zero_no_epsilon_greedy(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7,
epsilon_greedy=False)
epsilon_probs = [
agent.initial_action()['epsilon_prob'] for _ in range(10)
]
self.assertEqual(epsilon_probs, [0] * 10)
def test_agent_produces_different_epsilon_with_epsilon_greedy(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7,
epsilon_greedy=True)
obs, _, done, _ = env.step(agent.initial_action())
epsilon_probs = [float(agent.initial_action()['epsilon_prob'])]
epsilon_probs.extend(
[float(agent.act(obs, done)['epsilon_prob']) for _ in range(10)])
self.assertGreater(len(set(epsilon_probs)), 1)
def test_agent_raises_episode_done_error(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7)
with self.assertRaises(core.EpisodeDoneError):
agent.act(observation={
'threshold': np.array(0.5),
'epsilon_prob': np.array(0)
},
done=True)
def test_epsilon_prob_decays_as_expected(self):
env = college_admission.CollegeAdmissionsEnv()
agent = college_admission_jury.FixedJury(
action_space=env.action_space,
observation_space=env.observation_space,
reward_fn=(lambda x: 0),
threshold=0.7,
epsilon_greedy=True,
initial_epsilon_prob=0.3,
decay_steps=5,
epsilon_prob_decay_rate=0.001)
obs, _, done, _ = env.step(agent.initial_action())
epsilon_probs = [float(agent.initial_action()['epsilon_prob'])]
epsilon_probs.extend(
[float(agent.act(obs, done)['epsilon_prob']) for _ in range(2)])
self.assertTrue(
np.all(np.isclose(epsilon_probs, [0.3, 0.0753, 0.0189],
atol=1e-2)))