def _make_rvs(self):
r_min, r_max = self._get_return_ranges()
# Variables for expected return given T under the candidate policy
R1 = rvs.BoundedRealSampleSet(name='R1', lower=r_min, upper=r_max)
ER1 = R1.expected_value('E[R|T=1]', mode=mode)
# Variables for expected return given T under the reference policy
R_ref1 = rvs.BoundedRealSampleSet(name='R_ref1',
lower=r_min,
upper=r_max)
ER_ref1 = R_ref1.expected_value('E[R_ref|T=1]', mode=mode)
# Constants
e = rvs.constant(self.epsilons[0], name='e')
# BQF and Constraint Objectives
# g(theta) := (E[R_ref|T=1] - E[R|T=1]) - e
BQF = rvs.sum(ER_ref1, -ER1, name='BQF')
CO = rvs.sum(BQF, -e, name='CO')
SCO = rvs.sum(BQF, -e, name='SCO', scaling=scaling)
# Store the sample sets and variables
self._scheck_rvs = [CO]
self._ccheck_rvs = [SCO]
self._eval_rvs = {'bqf_0_mean': BQF, 'co_0_mean': CO}
# Add the sample sets and variables to the manager
self._vm = rvs.VariableManager(self._preprocessor)
self._vm.add_sample_set(R1, R_ref1)
self._vm.add(ER1, ER_ref1, BQF, CO, SCO)
def _make_rvs(self):
r_min, r_max = self._get_return_ranges()
# Variables for expected return given T under the candidate policy
R0 = rvs.BoundedRealSampleSet(name='R0', lower=r_min, upper=r_max)
R1 = rvs.BoundedRealSampleSet(name='R1', lower=r_min, upper=r_max)
ER0 = R0.expected_value('E[R|T=0]', mode=mode)
ER1 = R1.expected_value('E[R|T=1]', mode=mode)
# Constants
eM = rvs.constant(self.epsilons[0], name='eM')
eF = rvs.constant(self.epsilons[1], name='eF')
r_ref0 = rvs.constant(self._ref_return_T0, name='Avg[R_ref|T=0]')
r_ref1 = rvs.constant(self._ref_return_T1, name='Avg[R_ref|T=1]')
# BQF and Constraint Objectives
# g0(theta) := (E[R_ref|T=0] - Average(R|T=0,D) - eM
# g1(theta) := (E[R_ref|T=1] - Average(R|T=1,D) - eF
BQF0 = rvs.sum(r_ref0, -ER0, name='BQF0')
BQF1 = rvs.sum(r_ref1, -ER1, name='BQF1')
CO0 = rvs.sum(BQF0, -eM, name='CO0')
SCO0 = rvs.sum(BQF0, -eM, name='SCO0', scaling=scaling)
CO1 = rvs.sum(BQF1, -eF, name='CO1')
SCO1 = rvs.sum(BQF1, -eF, name='SCO1', scaling=scaling)
# Store the sample sets and variables
self._scheck_rvs = [CO0, CO1]
self._ccheck_rvs = [SCO0, SCO1]
self._eval_rvs = {
'bqf_0_mean': BQF0,
'co_0_mean': CO0,
'bqf_1_mean': BQF1,
'co_1_mean': CO1
}
# Add the sample sets and variables to the manager
self._vm = rvs.VariableManager(self._preprocessor)
self._vm.add_sample_set(R0, R1)
self._vm.add(ER0, ER1, BQF0, CO0, SCO0, BQF1, CO1, SCO1)
def __init__(self, epsilons, deltas, model_type, minimum_return, iw_type_corrections={}):
self.epsilons = epsilons
self.deltas = deltas
self.model_type = model_type
self.model_variables = {}
self.minimum_return = minimum_return
self._vm = rvs.VariableManager(self._preprocessor)
self._scheck_rvs = []
self._ccheck_rvs = []
self._eval_rvs = {}
self._iw_type_corrections = defaultdict(lambda : 1.0)
for k, v in iw_type_corrections.items():
self._iw_type_corrections[k] = v
self._make_rvs()
def _make_rvs(self):
# Sample sets for false-positive classifications conditioned on T
Pos0 = rvs.BoundedRealSampleSet(name='Pos0', lower=0, upper=1)
Pos1 = rvs.BoundedRealSampleSet(name='Pos1', lower=0, upper=1)
# Variables representing the conditional false-positive rates and the BQF
PrPos0 = Pos0.expected_value('Pr(Pos|T=0)', mode=mode)
PrPos1 = Pos1.expected_value('Pr(Pos|T=1)', mode=mode)
# Constants
pct = rvs.constant(self.epsilons[0], name='pct')
# BQF
BQF = rvs.maxrec(-(PrPos0 / PrPos1), name='BQF')
CO = rvs.sum(BQF, -pct, name='CO')
SCO = rvs.sum(BQF, -pct, name='SCO', scaling=scaling)
# Store the sample sets and variables
self._scheck_rvs = [CO]
self._ccheck_rvs = [SCO]
self._eval_rvs = {'bqf_0_mean': BQF, 'co_0_mean': CO}
# Add the sample sets and variables to the manager
self._vm = rvs.VariableManager(self._preprocessor)
self._vm.add_sample_set(Pos0, Pos1)
self._vm.add(PrPos0, PrPos1, BQF, CO, SCO)