def test_empty_belief_udpate(self):
prior = IndependentNormals(features=['pink'])
empty_df = pd.DataFrame()
posterior = prior.literal_update(empty_df)
self.assertDictEqual(prior.beliefs, posterior.beliefs)
def test_pragmatic_update(self):
prior = IndependentNormals(features=['pink', 'circle'],
precision=PRIOR_PRECISION)
obs_mean = 2
obs_precision = PRIOR_PRECISION
obs = pd.DataFrame.from_records([{
"feature": "pink",
"mean": obs_mean,
"precision": obs_precision
}])
literal_posterior = prior.literal_update(obs)
pragmatic_precision = 2
pragmatic_posterior = prior.pseudopragmatic_gaussian_update(
obs, pragmatic_precision)
# Literal and pragmatic should be the same for means of both, and variance of pink
self.assertEqual(literal_posterior["pink"]["mean"],
pragmatic_posterior["pink"]["mean"])
self.assertEqual(literal_posterior["pink"]["precision"],
pragmatic_posterior["pink"]["precision"])
self.assertEqual(literal_posterior["circle"]["mean"],
pragmatic_posterior["circle"]["mean"])
# Should *differ* for precision of unmentioned (circle) feature
self.assertNotEqual(literal_posterior["circle"]["precision"],
pragmatic_posterior["circle"]["precision"])
self.assertEqual(pragmatic_posterior["circle"]["precision"],
PRIOR_PRECISION + pragmatic_precision)
def test_belief_samples(self):
mean = 0
var = 2
prior = IndependentNormals(features=['pink'],
mean=0,
precision=1 / var)
samples = prior.sample_beliefs(n=100000)
self.assertAlmostEqual(mean, abs(samples['pink'].mean()), 1)
self.assertAlmostEqual(var, samples['pink'].var(), 1)
def test_generate_gaussian_priors(self):
prior = IndependentNormals(features=['pink'])
self.assertDictEqual(
{"pink": {
'mean': PRIOR_MEAN,
"precision": PRIOR_PRECISION
}}, prior.beliefs)
def test_literal_update_precision(self):
prior = IndependentNormals(features=['pink', 'circle'],
precision=PRIOR_PRECISION)
obs_precision = 1
obs = pd.DataFrame.from_records([{
"feature": "pink",
"mean": 0,
"precision": obs_precision
}])
posterior = prior.literal_update(obs)
self.assertEqual(posterior["pink"]["precision"],
PRIOR_PRECISION + obs_precision)
self.assertEqual(posterior["circle"]["precision"], PRIOR_PRECISION)
def __init__(self, name="Ensemble Feedforward NN", precision=2):
priors = IndependentNormals(features=CONJUNCTIONS)
super().__init__(CONJUNCTIONS, name=name + " (P{})".format(precision), prior=priors)
# Choose to load a specific model from a training fold
self.models = [TrajectoryFeedbackRewardPredictor.from_cv_fold(fold) for fold in range(0, 10)]
self.precision = precision
def __init__(self, features=CONJUNCTIONS, name="Stateful Feedforward Neural Net", precision=2, fold=4):
priors = IndependentNormals(features=features)
super().__init__(features, name=name + " (P{})".format(precision), prior=priors)
# Choose to load a specific model from a training fold
self.nn = TrajectoryFeedbackRewardPredictor.from_cv_fold(fold)
self.precision = precision
def __init__(self, features=PILOT_FEATURES, prior=None, name=None):
self._name = name
self.features = sorted(features)
if prior is None:
prior = IndependentNormals(self.features)
self.initial_beliefs = prior
self.belief_state = copy.deepcopy(self.initial_beliefs)
def test_optimal_action_selection(self):
env = Environment.from_pilot_config(PILOT_FIRST_LEVEL_CONFIG)
# Generate a belief that all objects except pink circles are terrible
belief = IndependentNormals(mean=-10, precision=50)
belief['pink|circle']['mean'] = 150
belief_samples = belief.sample_beliefs(n=10)
object_choices = optimal_choices_under_hypotheses(env, belief_samples)
# Validate that we only ever chose pink circles
for object_locations in object_choices.locations.values:
self.assertSetEqual(PINK_CIRCLE_LOCATIONS, set(object_locations),
msg="Should only have chosen pink circles.")
# Reset beliefs, resample, validate we only choose yellow squares.
belief = IndependentNormals(mean=-10, precision=50)
belief['yellow|square']['mean'] = 150
belief_samples = belief.sample_beliefs(n=10)
object_choices = optimal_choices_under_hypotheses(env, belief_samples)
for object_locations in object_choices.locations.values:
self.assertSetEqual(YELLOW_SQUARE_LOCATIONS, set(object_locations),
msg="Should only have chosen yellow square.")
def test_multiple_literal_updates(self):
prior = IndependentNormals(features=['pink', 'circle'],
precision=PRIOR_PRECISION)
obs_mean = 2
obs_precision = PRIOR_PRECISION
obs = pd.DataFrame.from_records([
{
"feature": "pink",
"mean": obs_mean,
"precision": obs_precision
},
{
"feature": "circle",
"mean": obs_mean,
"precision": obs_precision
},
])
posterior = prior.literal_update(obs)
self.assertAlmostEqual(posterior["pink"]["mean"],
(PRIOR_MEAN + obs_mean) / 2)
self.assertAlmostEqual(posterior["circle"]["mean"],
(PRIOR_MEAN + obs_mean) / 2)
self.assertAlmostEqual(posterior["circle"]["precision"],
PRIOR_PRECISION + obs_precision)
posterior = posterior.literal_update(obs)
self.assertAlmostEqual(posterior["pink"]["mean"],
(PRIOR_MEAN + 2 * obs_mean) / 3)
self.assertAlmostEqual(posterior["circle"]["mean"],
(PRIOR_MEAN + 2 * obs_mean) / 3)
self.assertAlmostEqual(posterior["circle"]["precision"],
PRIOR_PRECISION + 2 * obs_precision)
def from_json(cls, data):
new_agent = cls()
new_agent.belief_state = IndependentNormals.from_json(data)
return new_agent