def __call__(self, g):
"""return `list` of AL penalties for constraints values in `g`.
Penalties are zero in the optimum and can be negative down to
``-lam**2 / mu / 2``.
"""
if self.lam is None:
return [0.] * len(g)
assert len(self.lam) == len(self.mu)
if len(g) != len(self.lam):
raise ValueError("len(g) = %d != %d = # of Lagrange coefficients"
% (len(g), len(self.lam)))
assert self._equality.dtype == 'bool'
idx = _and(_not(self._equality), self.mu * g < -1 * self.lam)
if any(idx):
g = np.array(g, copy=True)
g[idx] = -self.lam[idx] / self.mu[idx]
return [self.lam[i] * g[i] + 0.5 * self.mu[i] * g[i]**2
for i in range(len(g))]
def set_coefficients(self, F, G):
"""compute initial coefficients based on some f- and g-values.
The formulas to set the coefficients::
lam = iqr(f) / (sqrt(n) * iqr(g))
mu = 2 * iqr(f) / (5 * n * (iqr(g) + iqr(g**2)))
are taken out of thin air and not thoroughly tested. They are
additionally protected against division by zero.
Each row of `G` represents the constraints of one sample measure.
Set lam and mu until a population contains more than 10% infeasible
and more than 10% feasible at the same time. Afterwards, this at least...?...
"""
self.F, self.G = F, G # versatile storage
if self.mu is not None and all(self._initialized * (self.mu > 0)):
return # we're all set
G = np.asarray(G).T # now row G[i] contains all values of constraint i
sign_average = np.mean(np.sign(G),
axis=1) # caveat: equality vs inequality
if self.lam is None: # destroys all coefficients
self.set_m(len(G))
elif not self._initialized.shape:
self.lam_old, self.mu_old = self.lam[:], self.mu[:] # in case the user didn't mean to
self._initialized = np.array(self.m * [self._initialized])
self._check_dtypes()
idx = _and(_or(_not(self._initialized), self.mu == 0),
_or(sign_average > -0.8, self.isequality))
# print(len(F), len(self.G))
if np.any(idx):
df = _Mh.iqr(F)
dG = np.asarray([_Mh.iqr(g)
for g in G]) # only needed for G[idx], but like
dG2 = np.asarray([_Mh.iqr(g**2)
for g in G]) # this simpler in later indexing
if np.any(df == 0) or np.any(dG == 0) or np.any(dG2 == 0):
_warnings.warn("iqr(f), iqr(G), iqr(G**2)) == %s, %s, %s" %
(str(df), str(dG), str(dG2)))
assert np.all(df >= 0) and np.all(dG >= 0) and np.all(dG2 >= 0)
# 1 * dG2 leads to much too small values for Himmelblau
mu_new = 2. / 5 * df / self.dimension / (
dG + 1e-6 * dG2 + 1e-11 *
(df + 1)) # TODO: totally out of thin air
idx_inequ = _and(idx, _not(self.isequality))
if np.any(idx_inequ):
self.lam[idx_inequ] = df / (self.dimension * dG[idx_inequ] +
1e-11 * (df + 1))
# take min or max with existing value depending on the sign average
# we don't know whether this is necessary
isclose = np.abs(sign_average) <= 0.2
idx1 = _and(
_and(
_or(isclose,
_and(_not(self.isequality), sign_average <= 0.2)),
_or(self.mu == 0, self.mu > mu_new)), idx) # only decrease
idx2 = _and(
_and(
_and(_not(isclose), _or(self.isequality,
sign_average > 0.2)),
self.mu < mu_new), idx) # only increase
idx3 = _and(idx, _not(_or(idx1, idx2))) # others
assert np.sum(_and(idx1, idx2)) == 0
if np.any(idx1): # decrease
iidx1 = self.mu[idx1] > 0
if np.any(iidx1):
self.lam[idx1][
iidx1] *= mu_new[idx1][iidx1] / self.mu[idx1][iidx1]
self.mu[idx1] = mu_new[idx1]
if np.any(idx2): # increase
self.mu[idx2] = mu_new[idx2]
if np.any(idx3):
self.mu[idx3] = mu_new[idx3]
if 11 < 3: # simple version, this may be good enough
self.mu[idx] = mu_new[idx]
self._initialized[_and(
idx,
_or(
self.count >
2 + self.dimension, # in case sign average remains 1
np.abs(sign_average) < 0.8))] = True
elif all(self._initialized) and all(self.mu > 0):
_warnings.warn(
"Coefficients are already fully initialized. This can (only?) happen if\n"
"the coefficients are set before the `_initialized` array.")
