def setup_algorithms(dist, training_data, n_x, n_a, n_y, delta):
start = time.time()
statistical_approximation = StatisticalApproximator(n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
function_approximation = FunctionApproximation(n_x, n_a, n_y, training_data)
doctor_approximation = DoctorApproximator(n_x, n_a, n_y, training_data)
print("Initializing Constraint")
start = time.time()
constraintStatUpper = Constraint(training_data, n_a, n_y, approximator=statistical_approximation, delta=delta, bound='upper')
constraintFuncApprox = Constraint(training_data, n_a, n_y, approximator=function_approximation, delta=delta)
constraint_exact_func = TrueConstraint(dist, approximator=function_approximation, delta=delta)
print("Initializing the constraint took {:.3f} seconds".format(time.time() - start))
print("Initializing algorithms")
algorithms = [
#ConstrainedGreedy(n_x, n_a, n_y, training_data, constraintStatUpper, statistical_approximation,
# name='Constrained Greedy', label='CG'),
# ConstrainedGreedy(n_x, n_a, n_y, split_training_data, constraintStatLower, statistical_approximation,
# name='Constrained Greedy Lower', label='CG_L'),
ConstrainedGreedy(n_x, n_a, n_y, training_data, constraintFuncApprox, function_approximation,
name="Constrained Greedy FuncApprox", label="CG_F"),
#ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraintStatUpper,
# statistical_approximation),
ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraintFuncApprox,
function_approximation, name="Constrained Dynamic Programming FuncApprox", label="CDP_F"),
#NaiveGreedy(n_x, n_a, n_y, function_approximation, max_steps=n_a),
#NaiveGreedy(n_x, n_a, n_y, function_approximation, max_steps=n_a),
NaiveDynamicProgramming(n_x, n_a, n_y, training_data, constraintStatUpper, reward=-0.35),
Doctor(),
EmulatedDoctor(n_x, n_a, n_y, training_data, approximator=doctor_approximation)
]
return algorithms
def setup_algorithms(training_data, dist, delta):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
constraint_upper = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta,
bound='upper')
constraint_lower = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta,
bound='lower')
constraint_exact = TrueConstraint(
dist, approximator=statistical_approximation_prior, delta=delta)
algorithms = [
#ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraint_upper, statistical_approximation_prior, name="Dynamic Programming Upper Bound", label="CDP_U"),
#ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraint_lower, statistical_approximation_prior, name="Dynamic Programming Lower bound", label="CDP_L"),
#ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraint_exact, statistical_approximation_prior, name="Dynamic Programming Exact Bound", label="CDP_E"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_upper,
statistical_approximation_prior,
name="Greedy Upper Bound",
label="CG_U"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_lower,
statistical_approximation_prior,
name="Greedy Lower Bound",
label="CG_L"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_exact,
statistical_approximation_prior,
name="Greedy Exact Bound",
label="CG_E"),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
def setup_algorithms(training_data, dist, delta, train=True):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
constraint_prior = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraint_prior = TrueConstraint(
dist, approximator=statistical_approximation_prior, delta=delta)
algorithms = [
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Dynamic Programming",
label="CDP"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Greedy",
label="CG"),
NaiveGreedy(n_x,
n_a,
n_y,
statistical_approximation_prior,
round(delta * (n_a - 1)) + 1,
name='Naive Greedy',
label='NG'),
NaiveDynamicProgramming(n_x,
n_a,
n_y,
training_data,
statistical_approximation_prior,
reward=-(delta + 0.0001),
name='Naive Dynamic Programming',
label='NDP'),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
def setup_algorithms(training_data, dist, delta, train=True):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
algorithms = [
NaiveDynamicProgramming(n_x, n_a, n_y, training_data, statistical_approximation_prior, reward=-(delta/2+0.0001), name='Naive Dynamic Programming', label='NDP'),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() - start))
return algorithms
def setup_algorithms(training_data, dist, delta, train=True):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
true_approximation = ExactApproximator(dist)
constraint_prior = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraint_true = Constraint(training_data,
n_a,
n_y,
approximator=true_approximation,
delta=delta)
algorithms = [
ConstrainedDynamicProgramming(
n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Dynamic Programming Historical Prior",
label="CDP_H"),
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_true,
true_approximation,
name="Dynamic Programming True",
label="CDP_T"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Greedy Historical Prior",
label="CG_H"),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
def setup_algorithms(training_data, dist, delta):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_none = StatisticalApproximator(n_x, n_a, n_y, training_data, smoothing_mode='none')
statistical_approximation_prior = StatisticalApproximator(n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
function_approximation = FunctionApproximation(n_x, n_a, n_y, training_data)
true_approximation = ExactApproximator(dist)
constraint_none = Constraint(training_data, n_a, n_y, approximator=statistical_approximation_none, delta=delta)
constraint_prior = Constraint(training_data, n_a, n_y, approximator=statistical_approximation_prior, delta=delta)
constraint_func = Constraint(training_data, n_a, n_y, approximator=function_approximation, delta=delta)
constraint_true = Constraint(training_data, n_a, n_y, approximator=true_approximation, delta=delta)
algorithms = [
ConstrainedGreedy(n_x, n_a, n_y, training_data, constraint_none, statistical_approximation_none, name="Greedy Uniform Prior", label="CG_U"),
ConstrainedGreedy(n_x, n_a, n_y, training_data, constraint_prior, statistical_approximation_prior, name="Greedy Historical Prior", label="CG_H"),
ConstrainedGreedy(n_x, n_a, n_y, training_data, constraint_func, function_approximation, name="Greedy Function Approximation", label="CG_F"),
ConstrainedGreedy(n_x, n_a, n_y, training_data, constraint_true, true_approximation, name="Greedy True Approximation", label="CG_T"),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() - start))
return algorithms
def setup_algorithms(training_data, dist, delta):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
statistical_approximation_none = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='none')
true_approximation = ExactApproximator(dist)
constraint_prior = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraint_none = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_none,
delta=delta)
constraint_true = Constraint(training_data,
n_a,
n_y,
approximator=true_approximation,
delta=delta)
algorithms = [
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_true,
true_approximation,
name="Dynamic Programming True",
label="CDP_T"),
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Dynamic Programming",
label="CDP_H"),
ConstrainedDynamicProgramming(
n_x,
n_a,
n_y,
training_data,
constraint_none,
statistical_approximation_none,
name="Constrained Dynamic Programming Uninformed",
label="CDP_U"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Greedy",
label="CG"),
#NaiveGreedy(n_x, n_a, n_y, statistical_approximation_prior, round(delta * (n_a-1))+1, name='Naive Greedy', label='NG'),
#NaiveDynamicProgramming(n_x, n_a, n_y, training_data, statistical_approximation_prior, reward=-(delta+0.0001), name='Naive Dynamic Programming', label='NDP'),
]
print("Setting up algosrithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
def setup_algorithms(training_data, dist, delta, train=True):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
function_approximation = FunctionApproximation(n_x, n_a, n_y,
training_data)
constraint_prior = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraintFuncApprox = Constraint(training_data,
n_a,
n_y,
approximator=function_approximation,
delta=delta)
algorithms = [
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Dynamic Programming",
label="CDP"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Constrained Greedy",
label="CG"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraintFuncApprox,
function_approximation,
name="Constrained Greedy FuncApprox",
label="CG_F"),
# ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data, constraintStatUpper,
# statistical_approximation),
ConstrainedDynamicProgramming(
n_x,
n_a,
n_y,
training_data,
constraintFuncApprox,
function_approximation,
name="Constrained Dynamic Programming FuncApprox",
label="CDP_F"),
NaiveDynamicProgramming(n_x,
n_a,
n_y,
training_data,
statistical_approximation_prior,
reward=-(delta / 2 + 0.0001),
name='Naive Dynamic Programming',
label='NDP'),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
n_a = dist.n_a
n_y = dist.n_y
n_test_samples = len(test_data)
split_training_data = datasets['training']['data']
test_data = datasets['test']['data']
# print("Initializing function approximator")
# start = time.time()
function_approximation = FunctionApproximation(n_x, n_a, n_y,
split_training_data)
# print("Initializing {} took {:.3f} seconds".format(function_approximation.name, time.time()-start))
print("Initializing statistical approximator")
start = time.time()
statistical_approximation = StatisticalApproximator(n_x,
n_a,
n_y,
split_training_data,
smoothing_mode='none')
# print("Initializing {} took {:.3f} seconds".format(statistical_approximation.name, time.time() - start))
true_approximation = ExactApproximator(dist)
print("Initializing Constraint")
start = time.time()
constraintStatUpper = Constraint(split_training_data,
n_a,
n_y,
approximator=statistical_approximation,
delta=delta,
epsilon=epsilon,
def setup_algorithms(training_data, dist, delta, train=True):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
true_approximation = ExactApproximator(dist)
constraint_upper_stat = Constraint(
training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraint_upper_true = Constraint(training_data,
n_a,
n_y,
approximator=true_approximation,
delta=delta)
constraint_exact_true = TrueConstraint(dist,
approximator=true_approximation,
delta=delta)
constraint_exat_stat = TrueConstraint(
dist, approximator=statistical_approximation_prior, delta=delta)
algorithms = [
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_upper_stat,
statistical_approximation_prior,
name="Dynamic Programming Upper Stat",
label="CDP_US"),
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_upper_true,
true_approximation,
name="Dynamic Programming Upper True",
label="CDP_UT"),
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_exact_true,
true_approximation,
name="Dynamic Programming Exact True",
label="CDP_ET"),
ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
training_data,
constraint_exat_stat,
statistical_approximation_prior,
name="Dynamic Programming Exact Stat",
label="CDP_ES"),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms
n_training_samples = 500000
delta = 0.3
dist = DiscreteDistributionWithSmoothOutcomes(n_z,
n_x,
n_a,
n_y,
seed=seed,
outcome_sensitivity_x_z=1)
dist.print_treatment_statistics()
dist.print_detailed_treatment_statistics()
split_training_data = split_patients(generate_data(dist, n_training_samples))
sa = StatisticalApproximator(n_x,
n_a,
n_y,
split_training_data,
smoothing_mode='gaussian')
ta = ExactApproximator(dist)
print("Init constraints")
csa = Constraint(n_x, n_a, n_y, approximator=sa, delta=delta)
cta = Constraint(n_x, n_a, n_y, approximator=ta, delta=delta)
cdp = ConstrainedDynamicProgramming(n_x,
n_a,
n_y,
split_training_data,
csa,
sa,
name="Dynamic Programming",
label="CDP")
print("Generating training data set {}".format(n_training_samples))
main_data = generate_data(dist, n_training_samples)
for i in range(len(data_limits)):
d_tmp = {}
d_tmp['x'] = np.copy(main_data['x'][0:data_limits[i]])
d_tmp['h'] = np.copy(main_data['h'][0:data_limits[i]])
d_tmp['z'] = np.copy(main_data['z'][0:data_limits[i]])
data_sets.append(split_patients(d_tmp))
true_approximation = ExactApproximator(dist)
evaluations_data_amount = {}
time_name = 'time'
outcome_name = 'outcome'
for training_data_set in data_sets:
print("Initializing approximator")
statistical_approximationPrior = StatisticalApproximator(n_x, n_a, n_y, training_data_set, smoothing_mode='gaussian')
statistical_approximationNone = StatisticalApproximator(n_x, n_a, n_y, training_data_set, smoothing_mode='none')
function_approximation = FunctionApproximation(n_x, n_a, n_y, training_data_set)
print("Initializing Constraint")
start = time.time()
constraintNone = Constraint(training_data_set, n_a, n_y, approximator=statistical_approximationNone, delta=delta, epsilon=epsilon)
constraintPrior = Constraint(training_data_set, n_a, n_y, approximator=statistical_approximationPrior, delta=delta, epsilon=epsilon)
constraintFunc = Constraint(training_data_set, n_a, n_y, approximator=function_approximation, delta=delta, epsilon=epsilon)
constraintTrue = Constraint(training_data_set, n_a, n_y, approximator=true_approximation, delta=delta, epsilon=epsilon)
print("Initializing the constraint took {:.3f} seconds".format(time.time()-start))
print("Initializing algorithms")
algorithms = [
ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data_set, constraintNone, statistical_approximationNone, name="CDP_U", label="CDP_U"),
ConstrainedDynamicProgramming(n_x, n_a, n_y, training_data_set, constraintPrior, statistical_approximationPrior, name="CDP_H", label="CDP_H"),
def setup_algorithms(training_data, dist, delta):
start = time.time()
n_x = dist.n_x
n_a = dist.n_a
n_y = dist.n_y
statistical_approximation_none = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='none')
statistical_approximation_prior = StatisticalApproximator(
n_x, n_a, n_y, training_data, smoothing_mode='gaussian')
function_approximation = FunctionApproximation(n_x, n_a, n_y,
training_data)
true_approximation = ExactApproximator(dist)
constraint_none = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_none,
delta=delta)
constraint_prior = Constraint(training_data,
n_a,
n_y,
approximator=statistical_approximation_prior,
delta=delta)
constraint_func = Constraint(training_data,
n_a,
n_y,
approximator=function_approximation,
delta=delta)
constraint_true = Constraint(training_data,
n_a,
n_y,
approximator=true_approximation,
delta=delta)
algorithms = [
ConstrainedDynamicProgramming(
n_x,
n_a,
n_y,
training_data,
constraint_prior,
statistical_approximation_prior,
name="Dynamic Programming Historical Prior",
label="CDP_H"),
ConstrainedGreedy(n_x,
n_a,
n_y,
training_data,
constraint_func,
function_approximation,
name="Constrained Greedy Historical Prior",
label="CG_H"),
NaiveDynamicProgramming(
n_x,
n_a,
n_y,
training_data,
statistical_approximation_prior,
reward=-(delta + 0.0001),
name='Naive Dynamic Programming Historical Prior',
label='NDP_H'),
NaiveGreedy(n_x,
n_a,
n_y,
approximator=statistical_approximation_prior,
max_steps=4),
]
print("Setting up algorithms took {:.3f} seconds".format(time.time() -
start))
return algorithms