def first(self, e):
self.mb = set(population_icp.markov_blanket(self.target, e))
self.candidates = self.mb.copy()
if self.mb == set():
var = np.random.choice(utils.all_but(self.target, self.p))
else:
var = np.random.choice(list(self.mb))
self.interventions.append(var)
return var
def markov_blanket(i, dist, tol=1e-10):
"""Return the Markov blanket of variable i wrt. the given
distribution. HOW IT WORKS: In the population setting, the
regression coefficients of all variables outside the Markov
Blanket are 0. Taking into account numerical issues (that's what
tol is for), use this to compute the Markov blanket.
"""
(coefs, _) = dist.regress(i, utils.all_but(i, dist.p))
return utils.nonzero(coefs, tol)
def first(self, e):
mb = population_icp.markov_blanket(self.target, e)
if len(
mb
) == 0: # If the estimate of the MB is empty resort to the random strategy
self.mb = np.random.permutation(utils.all_but(self.target, self.p))
else:
self.mb = np.random.permutation(mb)
self.i = 0
return self.pick_intervention()
def markov_blanket(sample, target, tol=1e-3, debug=False):
"""Use the Lasso estimator to return an estimate of the Markov Blanket"""
with warnings.catch_warnings():
warnings.simplefilter("ignore")
p = sample.shape[1]
predictors = utils.all_but(target, p)
X = sample[:, predictors]
Y = sample[:, target]
coefs = np.zeros(p)
coefs[predictors] = linear_model.LassoCV(cv=10,
normalize=True,
max_iter=1000,
verbose=debug).fit(X, Y).coef_
return utils.nonzero(coefs, tol)
def intervention_targets(target, response, p, max_off_targets):
"""Return intervention targets. If no off target effects are allowed
(i.e. max_off_targets = 0), the intervention is at the given
target. Otherwise, a number (at random, up to max_off_targets) of
additional targets are selected at random, excluding the response
(i.e. Y)."""
if max_off_targets == 0:
return [target]
else:
max_off_targets = min(max_off_targets, p-2)
num_off_targets = np.random.choice(max_off_targets+1)
off_targets = list(np.random.choice(utils.all_but([target, response], p),
size=num_off_targets,
replace=False))
return [target] + off_targets
def first(self, _):
self.idx = np.random.permutation(utils.all_but(self.target, self.p))
self.i = 0
return self.random_intervention()
def first(self, sample):
self.obs_sample = sample
self.candidates = set(utils.all_but(self.target, self.p))
var = self.pick_intervention()
self.interventions.append(var)
return var
def pick_intervention(self):
return np.random.choice(utils.all_but(self.target, self.p))
def test_all_but(self):
self.assertTrue([0, 1, 3, 4] == all_but(2,5))
self.assertTrue([0, 3, 4] == all_but([1,2],5))