def add_columns(parent, dk, almost_zero):
child = copy.deepcopy(parent)
G = child.variables[0]
S = child.variables[1]
m = S.shape[0]
n = G.shape[0]
k = child.objectives[1]
# create larger arrays and store G,S in them
newG = np.zeros((n, k + dk))
newS = np.zeros((m, k + dk, k + dk))
newG[:, :k] = G
newS[:, :k, :k] = S
# add random columns to newG
newG[:, k:] = almost_zero * np.random.rand(n, dk)
# create side and corner block of S in symmetric way ...
S_side = almost_zero * np.random.rand(m, k, dk)
S_corner = 0.5 * almost_zero * np.random.rand(m, dk, dk)
S_corner = S_corner + np.transpose(S_corner, (0, 2, 1))
# ... and add those blocks to newS
newS[:, :k, k:] = S_side
newS[:, k:, :k] = np.transpose(S_side, (0, 2, 1))
newS[:, k:, k:] = S_corner
child.variables = [newG, newS]
child.evaluated = False
return child
def evolve(self, parents):
result = copy.deepcopy(parents[0])
G1 = np.array(parents[0].variables)[0:self.n * self.k].reshape(
(self.n, self.k))
Ss1 = np.array(parents[0].variables)[(self.n * self.k):].reshape(
(self.m, self.k, self.k))
G2 = np.array(parents[1].variables)[0:self.n * self.k].reshape(
(self.n, self.k))
Ss2 = np.array(parents[1].variables)[(self.n * self.k):].reshape(
(self.m, self.k, self.k))
newG = np.zeros(G1.shape)
from_G1 = 0
for i in range(G1.shape[0]):
bit = bool(random.getrandbits(1))
from_G1 += int(bit)
newG[i, :] = (G1[i, :] if bit else G2[i, :])
if from_G1 < G1.shape[0] / 2:
result.variables = list(
np.concatenate((newG.flatten(), Ss1.flatten()), axis=0))
else:
result.variables = list(
np.concatenate((newG.flatten(), Ss2.flatten()), axis=0))
return [result]
def mutate(self, parent, dk):
if dk > 0:
return add_columns(parent, dk, self.almost_zero)
elif dk < 0:
child = copy.deepcopy(parent)
for i in range(-dk):
child = delete_lest_significant_column(child)
return child
else:
raise PlatypusError('Algorithm wanted to delete 0 columns!')
def delete_random_column(parent):
child = copy.deepcopy(parent)
k = child.objectives[1]
G = child.variables[0]
S = child.variables[1]
i = np.random.randint(k)
G = np.delete(G, i, 1)
S = np.delete(S, i, 1)
S = np.delete(S, i, 2)
child.variables = [G, S]
child.evaluated = False
return child
def mutate(self,
parent,
worsening_factor=5.,
convergence_steps=150,
convergence_factor=3.):
child = copy.deepcopy(parent)
G = child.variables[0]
S = child.variables[1]
k = S.shape[1]
c, newG, newS, dnfe = self.gd.adam(G, S, self.steps, worsening_factor,
convergence_steps,
convergence_factor)
child.variables = [newG, newS]
child.objectives = [c, k]
child.evaluated = True
return child, dnfe
def mutate(self, parent):
child = copy.deepcopy(parent)
problem = child.problem
probability = self.probability
if isinstance(probability, int):
probability /= float(
len([t for t in problem.types if isinstance(t, Real)]))
if random.uniform(0.0, 1.0) <= self.probability:
G, Ss = self.reshaper.vec2mat(child.variables)
c, newG, newS = self.p.adam(G, Ss, self.steps)
con = self.reshaper.mat2vec(newG, newS)
child.variables = con.tolist()
child.objectives = [c]
child.evaluated = True
return child
def evolve(self, parents):
result = copy.deepcopy(parents[0])
# Join G matrices side by side
G1 = parents[0].variables[0]
G2 = parents[1].variables[0]
newG = self.factorG * np.concatenate((G1, G2), axis=1)
# Join S tensors as a direct sum
S1 = parents[0].variables[1]
S2 = parents[1].variables[1]
k1 = parents[0].objectives[1]
k2 = parents[1].objectives[1]
m = S1.shape[0]
newS = np.empty((m, k1 + k2, k1 + k2))
newS[:, :k1, :k1] = self.factorS * S1
newS[:, k1:, k1:] = self.factorS * S2
almost_zero = self.almost_zero * np.random.rand(m, k1, k2)
newS[:, :k1, k1:] = almost_zero
newS[:, k1:, :k1] = np.transpose(almost_zero, (0, 2, 1))
# Save to result
result.variables = [newG, newS]
result.evaluated = False
return [result]
algorithm.run(int(nfe))
elapsed_time = time.time() - start_time
cumulative_time = cumulative_time + elapsed_time
cumulative_nfe += algorithm.nfe
print('time needed=' + str(elapsed_time))
population = algorithm.result
new_p = constraint.evolve(population)
new_new_p = []
for sub in new_p:
G = sub.variables[0]
S = sub.variables[1]
p_copy = copy.deepcopy(sub)
S_new = np.zeros((Ri.shape[0], G.shape[1], G.shape[1]))
for i in range(Ri.shape[0]):
Rii = Ri[i, :, :]
S_new[i] = _gradientdescent(G, Rii)
p_copy.variables = [G, S_new]
new_new_p.append(p_copy)
algorithm.evaluate_all(new_new_p)
# Save values of objectives and solutions
write_population_objectives(algorithm.result, out_file_name + run_str)
write_population_objectives(new_new_p,
out_file_name + "_ortoghonal_" + run_str)