def test_make_bandit_historical_info_from_params_variance_passed_through(self):
"""Test that the variance of a given sample arm got passed through."""
historical_info = self.three_arms_with_variance_no_unsampled_arm_test_case
T.assert_dicts_equal(
_make_bandit_historical_info_from_params(self.make_params_from_bandit_historical_info(historical_info)).json_payload(),
historical_info.json_payload()
)
def bandit_epsilon_view(self):
"""Endpoint for bandit_epsilon POST requests.
.. http:post:: /bandit/epsilon
Predict the optimal arm from a set of arms, given historical data.
:input: :class:`moe.views.schemas.rest.bandit_epsilon.BanditEpsilonRequest`
:output: :class:`moe.views.schemas.bandit_pretty_view.BanditResponse`
:status 200: returns a response
:status 500: server error
"""
params = self.get_params_from_request()
subtype = params.get('subtype')
historical_info = _make_bandit_historical_info_from_params(params)
bandit_class = EPSILON_SUBTYPES_TO_BANDIT_METHODS[subtype].bandit_class(historical_info=historical_info, **params.get('hyperparameter_info'))
arms_to_allocations = bandit_class.allocate_arms()
return self.form_response({
'endpoint': self._route_name,
'arm_allocations': arms_to_allocations,
'winner': bandit_class.choose_arm(arms_to_allocations),
})
def bandit_epsilon_view(self):
"""Endpoint for bandit_epsilon POST requests.
.. http:post:: /bandit/epsilon
Predict the optimal arm from a set of arms, given historical data.
:input: :class:`moe.views.schemas.rest.bandit_epsilon.BanditEpsilonRequest`
:output: :class:`moe.views.schemas.bandit_pretty_view.BanditResponse`
:status 200: returns a response
:status 500: server error
"""
params = self.get_params_from_request()
subtype = params.get('subtype')
historical_info = _make_bandit_historical_info_from_params(params)
bandit_class = EPSILON_SUBTYPES_TO_BANDIT_METHODS[
subtype].bandit_class(historical_info=historical_info,
**params.get('hyperparameter_info'))
arms_to_allocations = bandit_class.allocate_arms()
return self.form_response({
'endpoint':
self._route_name,
'arm_allocations':
arms_to_allocations,
'winner':
bandit_class.choose_arm(arms_to_allocations),
})
def test_make_bandit_historical_info_from_params_variance_passed_through(
self):
"""Test that the variance of a given sample arm got passed through."""
historical_info = self.three_arms_with_variance_no_unsampled_arm_test_case
assert _make_bandit_historical_info_from_params(
self.make_params_from_bandit_historical_info(historical_info)
).json_payload() == historical_info.json_payload()
def test_make_bandit_historical_info_from_params_make_bernoulli_arms(self):
"""Test that the function can make historical infos with Bernoulli arms."""
historical_info = self.three_arms_with_variance_no_unsampled_arm_test_case
for historical_info in self.bernoulli_historical_infos_to_test:
assert _make_bandit_historical_info_from_params(
self.make_params_from_bandit_historical_info(historical_info),
BernoulliArm).json_payload() == historical_info.json_payload()
def test_make_bandit_historical_info_from_params_make_bernoulli_arms(self):
"""Test that the function can make historical infos with Bernoulli arms."""
historical_info = self.three_arms_with_variance_no_unsampled_arm_test_case
for historical_info in self.bernoulli_historical_infos_to_test:
T.assert_dicts_equal(
_make_bandit_historical_info_from_params(self.make_params_from_bandit_historical_info(historical_info), BernoulliArm).json_payload(),
historical_info.json_payload()
)
def run_example(
verbose=True,
testapp=None,
bandit_bla_kwargs=None,
bandit_epsilon_kwargs=None,
bandit_ucb_kwargs=None,
**kwargs
):
"""Run the bandit example.
:param verbose: Whether to print information to the screen [True]
:type verbose: bool
:param testapp: Whether to use a supplied test pyramid application or a rest server [None]
:type testapp: Pyramid test application
:param bandit_bla_kwargs: Optional kwargs to pass to bandit_bla endpoint
:type bandit_bla_kwargs: dict
:param bandit_epsilon_kwargs: Optional kwargs to pass to bandit_epsilon endpoint
:type bandit_epsilon_kwargs: dict
:param bandit_ucb_kwargs: Optional kwargs to pass to bandit_ucb endpoint
:type bandit_ucb_kwargs: dict
:param kwargs: Optional kwargs to pass to all endpoints
:type kwargs: dict
"""
# Set and combine all optional kwargs
# Note that the more specific kwargs take precedence (and will override general kwargs)
bandit_kwargs = {}
if bandit_bla_kwargs is None:
bandit_bla_kwargs = {}
bandit_kwargs[BANDIT_BLA_ROUTE_NAME] = dict(kwargs.items() + bandit_bla_kwargs.items())
if bandit_epsilon_kwargs is None:
bandit_epsilon_kwargs = {}
bandit_kwargs[BANDIT_EPSILON_ROUTE_NAME] = dict(kwargs.items() + bandit_epsilon_kwargs.items())
if bandit_ucb_kwargs is None:
bandit_ucb_kwargs = {}
bandit_kwargs[BANDIT_UCB_ROUTE_NAME] = dict(kwargs.items() + bandit_ucb_kwargs.items())
# A BernoulliArm has payoff 1 for a success and 0 for a failure.
# See :class:`~moe.bandit.data_containers.BernoulliArm` for more details.
historical_info = _make_bandit_historical_info_from_params(
{
"historical_info": {
"arms_sampled": {
"arm1": {"win": 20, "loss": 0, "total": 25},
"arm2": {"win": 20, "loss": 0, "total": 30},
"arm3": {"win": 0, "loss": 0, "total": 0},
}
},
},
arm_type=BernoulliArm
)
# Run all multi-armed bandit strategies we have implemented. See :doc:`bandit` for more details on multi-armed bandits.
# We have implemented 3 bandit types: BLA (Bayesian Learning Optimization), Epsilon, and UCB (Upper Confidence Bound).
for type in BANDIT_ROUTE_NAMES:
if verbose:
print "Running Bandit: {0:s}...".format(type)
# Each bandit type has different subtypes. If a user does not specify a subtype, we use the default subtype.
# For example, the bandit type Epsilon has two subtypes: epsilon-first and epsilon-greedy.
# See :class:`~moe.bandit.epsilon.epsilon_first.EpsilonFirst` and :class:`~moe.bandit.epsilon.epsilon_greedy.EpsilonGreedy` for more details.
for subtype in BANDIT_ROUTE_NAMES_TO_SUBTYPES[type]:
if verbose:
print "Running subtype: {0:s}...".format(subtype)
bandit_kwargs[type]['subtype'] = subtype
# Compute and return arm allocations given the sample history of bandit arms.
# For example, the allocations {arm1: 0.3, arm2: 0.7} means
# if we have 10 arm pulls, we should pull arm1 3 times and arm2 7 times.
# See :func:`moe.bandit.bandit_interface.BanditInterface.allocate_arms` for more details.
arm_allocations = bandit(historical_info, type=type, testapp=testapp, **bandit_kwargs[type])
if verbose:
print "Arm allocations {0:s}".format(str(arm_allocations))
def run_example(verbose=True,
testapp=None,
bandit_bla_kwargs=None,
bandit_epsilon_kwargs=None,
bandit_ucb_kwargs=None,
**kwargs):
"""Run the bandit example.
:param verbose: Whether to print information to the screen [True]
:type verbose: bool
:param testapp: Whether to use a supplied test pyramid application or a rest server [None]
:type testapp: Pyramid test application
:param bandit_bla_kwargs: Optional kwargs to pass to bandit_bla endpoint
:type bandit_bla_kwargs: dict
:param bandit_epsilon_kwargs: Optional kwargs to pass to bandit_epsilon endpoint
:type bandit_epsilon_kwargs: dict
:param bandit_ucb_kwargs: Optional kwargs to pass to bandit_ucb endpoint
:type bandit_ucb_kwargs: dict
:param kwargs: Optional kwargs to pass to all endpoints
:type kwargs: dict
"""
# Set and combine all optional kwargs
# Note that the more specific kwargs take precedence (and will override general kwargs)
bandit_kwargs = {}
if bandit_bla_kwargs is None:
bandit_bla_kwargs = {}
bandit_kwargs[BANDIT_BLA_ROUTE_NAME] = dict(kwargs.items() +
bandit_bla_kwargs.items())
if bandit_epsilon_kwargs is None:
bandit_epsilon_kwargs = {}
bandit_kwargs[BANDIT_EPSILON_ROUTE_NAME] = dict(
kwargs.items() + bandit_epsilon_kwargs.items())
if bandit_ucb_kwargs is None:
bandit_ucb_kwargs = {}
bandit_kwargs[BANDIT_UCB_ROUTE_NAME] = dict(kwargs.items() +
bandit_ucb_kwargs.items())
# A BernoulliArm has payoff 1 for a success and 0 for a failure.
# See :class:`~moe.bandit.data_containers.BernoulliArm` for more details.
historical_info = _make_bandit_historical_info_from_params(
{
"historical_info": {
"arms_sampled": {
"arm1": {
"win": 20,
"loss": 0,
"total": 25
},
"arm2": {
"win": 20,
"loss": 0,
"total": 30
},
"arm3": {
"win": 0,
"loss": 0,
"total": 0
},
}
},
},
arm_type=BernoulliArm)
# Run all multi-armed bandit strategies we have implemented. See :doc:`bandit` for more details on multi-armed bandits.
# We have implemented 3 bandit types: BLA (Bayesian Learning Optimization), Epsilon, and UCB (Upper Confidence Bound).
for type in BANDIT_ROUTE_NAMES:
if verbose:
print "Running Bandit: {0:s}...".format(type)
# Each bandit type has different subtypes. If a user does not specify a subtype, we use the default subtype.
# For example, the bandit type Epsilon has two subtypes: epsilon-first and epsilon-greedy.
# See :class:`~moe.bandit.epsilon.epsilon_first.EpsilonFirst` and :class:`~moe.bandit.epsilon.epsilon_greedy.EpsilonGreedy` for more details.
for subtype in BANDIT_ROUTE_NAMES_TO_SUBTYPES[type]:
if verbose:
print "Running subtype: {0:s}...".format(subtype)
bandit_kwargs[type]['subtype'] = subtype
# Compute and return arm allocations given the sample history of bandit arms.
# For example, the allocations {arm1: 0.3, arm2: 0.7} means
# if we have 10 arm pulls, we should pull arm1 3 times and arm2 7 times.
# See :func:`moe.bandit.bandit_interface.BanditInterface.allocate_arms` for more details.
arm_allocations = bandit(historical_info,
type=type,
testapp=testapp,
**bandit_kwargs[type])
if verbose:
print "Arm allocations {0:s}".format(str(arm_allocations))
