def test_MLE_rate_estimation(self):
env_params = attention_allocation.Params()
env_params.prior_incident_counts = (500, 500)
env_params.n_attention_units = 5
# pylint: disable=g-long-lambda
agent_params = allocation_agents.MLEProbabilityMatchingAgentParams()
agent_params.feature_selection_fn = lambda obs: allocation_agents._get_added_vector_features(
obs, env_params.n_locations, keys=['incidents_seen'])
agent_params.interval = 200
agent_params.epsilon = 0
env = attention_allocation.LocationAllocationEnv(env_params)
agent = allocation_agents.MLEProbabilityMatchingAgent(
action_space=env.action_space,
reward_fn=lambda x: None,
observation_space=env.observation_space,
params=agent_params)
seed = 0
agent.rng.seed(seed)
env.seed(seed)
observation = env.reset()
done = False
steps = 200
for _ in range(steps):
action = agent.act(observation, done)
observation, _, done, _ = env.step(action)
self.assertTrue(
np.all(
np.isclose(list(agent.beliefs),
list(env_params.incident_rates),
atol=0.5)))
def mle_agent_epsilon_5_resource_all_dynamics():
"""Run experiments on a greedy-epsilon mle agent, epsilon=0.6, across dynamics."""
dynamic_values_to_test = [0.0, 0.01, 0.05, 0.1, 0.15]
experiment = _setup_experiment()
experiment.agent_class = allocation_agents.MLEProbabilityMatchingAgent
experiment.agent_params = allocation_agents.MLEProbabilityMatchingAgentParams(
)
experiment.agent_params.burn_steps = 25
experiment.agent_params.epsilon = 0.5
experiment.agent_params.window = 100
reports_dict = {}
for value in dynamic_values_to_test:
experiment.env_params.dynamic_rate = value
json_report = attention_allocation_experiment.run(experiment)
report = json.loads(json_report)
print('\n\nMLE Agent, 6 attention units, epsilon=0.5')
_print_discovered_missed_incidents_report(value, report)
output_filename = 'mle_epsilon.5_6units_%f.json' % value
with open(os.path.join(FLAGS.output_dir, output_filename), 'w') as f:
json.dump(report, f)
reports_dict[value] = json_report
return reports_dict
def test_MLEProbabilityMatchingAgent_works(self):
experiment = _setup_experiment()
experiment.agent_class = allocation_agents.MLEProbabilityMatchingAgent
experiment.agent_params = allocation_agents.MLEProbabilityMatchingAgentParams(
burn_steps=25, window=100)
result = attention_allocation_experiment.run(experiment)
# Tests that the result is a valid json string.
result = json.loads(result)
