def train(self):
"""Create a random population, evolve it and select a model based on selection."""
pop = Population(
self,
assign_type="EvoDN2",
pop_size=self.params["pop_size"],
recombination_type=self.params["recombination_type"],
crossover_type=self.params["crossover_type"],
mutation_type=self.params["mutation_type"],
plotting=self.params["plotting"],
)
pop.evolve(EA=self.params["training_algorithm"],
ea_parameters=self.ea_params)
non_dom_front = pop.non_dominated()
self.subnets, self.fitness = self.select(pop, non_dom_front,
self.params["selection"])
self.non_linear_layer, _ = self.activation(self.subnets)
self.linear_layer, *_ = self.calculate_linear(self.non_linear_layer)
def train(self):
"""Trains the networks and selects the best model from the non dominated front.
"""
pop = Population(
self,
assign_type="EvoNN",
pop_size=self.params["pop_size"],
plotting=self.params["plotting"],
recombination_type=self.params["recombination_type"],
crossover_type=self.params["crossover_type"],
mutation_type=self.params["mutation_type"],
)
pop.evolve(EA=self.params["training_algorithm"],
ea_parameters=self.ea_params)
non_dom_front = pop.non_dominated()
self.non_linear_layer, self.fitness = self.select(
pop, non_dom_front, self.params["selection"])
activated_layer, _ = self.activation(self.non_linear_layer)
self.linear_layer, *_ = self.calculate_linear(activated_layer)
def train(self):
"""Trains the networks and selects the best model from the non dominated front.
"""
# Minimize only error for first n generations before switching to bi-objective
ea_params = {
"generations_per_iteration": 1,
"iterations": self.params["single_obj_generations"],
}
self.minimize = [True, False]
print("Minimizing error for " +
str(self.params["single_obj_generations"]) + " generations...")
pop = Population(
self,
assign_type="BioGP",
pop_size=self.params["pop_size"],
plotting=self.params["plotting"],
recombination_type=self.params["recombination_type"],
crossover_type=self.params["crossover_type"],
mutation_type=self.params["mutation_type"],
)
pop.evolve(EA=TournamentEA, ea_parameters=ea_params)
# Switch to bi-objective (error, complexity)
self.minimize = [True, True]
pop.update_fitness()
print("Switching to bi-objective mode")
pop.evolve(EA=self.params["training_algorithm"],
ea_parameters=self.ea_params)
non_dom_front = pop.non_dominated()
self.linear_node, self.fitness = self.select(pop, non_dom_front,
self.params["selection"])
#k = 10
numobj = len(ANOVAMOPtest1SubpSurrogate.SubProblemObjectiveIndices)
numconst = 0
numvar = len(ANOVAMOPtest1SubpSurrogate.SubProblemVariablesIndices)
problem = ANOVAMOPtest1SubpSurrogate(name, numvar, numobj, numconst, )
lattice_resolution = 4
population_size = 105
pop = Population(problem)
# You can choose the solver (RVEA or NSGAIII)
pop.evolve(NSGAIII)
non_dom_index = pop.non_dominated()
xParetoTemp = pop.individuals[non_dom_index[0]]
xsize = np.shape(xParetoTemp)
fParetoTemp = pop.fitness[non_dom_index[0]] # it seems that the folowing objective calculations are not necessary as they can be returned here
xsize = np.shape(xParetoTemp)
p1m[0] = xParetoTemp # p1m = p1 ... pm (SubProblemsDecisionSpacePareto)
f1m[0] = fParetoTemp # f1m = f1 ... fm (SubProblemsObjectiveSpacePareto)
#refpoint = 2
#volume = 2 ** numobj
#print(pop.hypervolume(refpoint) / volume)
def __init__(
self,
name=None,
num_of_variables=None,
num_of_objectives=None,
num_of_constraints=0,
upper_limits=1,
lower_limits=0,
):
"""Pydocstring is ruthless.
Args:
name:
num_of_variables:
num_of_objectives:
num_of_constraints:
upper_limits:
lower_limits:
"""
super(testProblem, self).__init__(
name,
num_of_variables,
num_of_objectives,
num_of_constraints,
upper_limits,
lower_limits,
)
if name == "ZDT1":
self.obj_func = zdt.ZDT1()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "ZDT2":
self.obj_func = zdt.ZDT2()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "ZDT3":
self.obj_func = zdt.ZDT3()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "ZDT4":
self.obj_func = zdt.ZDT4()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "ZDT5":
self.obj_func = zdt.ZDT5()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "ZDT6":
self.obj_func = zdt.ZDT6()
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ1":
self.obj_func = dtlz.DTLZ1(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ2":
self.obj_func = dtlz.DTLZ2(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ3":
self.obj_func = dtlz.DTLZ3(num_of_objectives, num_of_variables)
self.lower_limits = 0
self.upper_limits = 1
elif name == "DTLZ4":
self.obj_func = dtlz.DTLZ4(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ5":
self.obj_func = dtlz.DTLZ5(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ6":
self.obj_func = dtlz.DTLZ6(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "DTLZ7":
self.obj_func = dtlz.DTLZ7(num_of_objectives, num_of_variables)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "BS1":
self.obj_func = P_BS1(Operation,Input)
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
elif name == "Surrogate":
self.obj_func = P_Surrogate((Operation,M,Input, SubProblemObjectiveIndices,SubProblemVariablesIndices,NumVar,Bounds,lb,ub,FixedIndices, FixedValues, model):
self.lower_limits = self.obj_func.min_bounds
self.upper_limits = self.obj_func.max_bounds
def objectives(self, decision_variables) -> list:
"""Use this method to calculate objective functions.
Args:
decision_variables:
"""
return self.obj_func(decision_variables)
def constraints(self, decision_variables, objective_variables):
"""Calculate constraint violation.
Args:
decision_variables:
objective_variables:
"""
print("Error: Constraints not supported yet.")
#--------------------------------------------------------------------------------
from pyrvea.Population.Population import Population
from pyrvea.Problem.baseProblem import baseProblem
from pyrvea.EAs.RVEA import RVEA
from pyrvea.EAs.NSGAIII import NSGAIII
from optproblems import dtlz
class newProblem(baseProblem):
"""New problem description."""
#def __init__(
#self,
#name = None, # problem name
#num_of_objectives[float] = 0, # M
#data = [], # Sample, Input
#SubProblemObjectiveIndices = [],
#SubProblemVariablesIndices = [],
#num_of_variables = None, # NumVar
#Sub_problem_lb = [],
#Sub_problem_ub = [],
#lower_limits = [],
#upper_limits = [],
#FixedIndices = [],
#FixedValues = [],
#model = [], # SurrogateDataInfo
#num_of_constraints = 0, # Not supported yet
#):
"""
Args:
name :
num_of_objectives :
Input :
SubProblemObjectiveIndices :
SubProblemVariablesIndices :
num_of_variables :
Sub_problem_lb :
Sub_problem_ub :
lower_limits :
upper_limits :
FixedIndices :
FixedValues :
model :
num_of_constraints : Not supported yet
"""
#super().__init__() # call the baseProblem __init__ function
#Sub_problem_lb = Bounds[0,:]
#Sub_problem_ub = Bounds[1,:]
#if name == "P_ALBERTO":
#self.obj_func = P_ALBERTO(Input)
#self.lower_limits = self.obj_func.min_bounds
#self.upper_limits = self.obj_func.max_bounds
#elif name == "P_Surrogate":
#self.obj_func = P_Surrogate(M,Input, SubProblemObjectiveIndices,SubProblemVariablesIndices,NumVar,Bounds,lb,ub,FixedIndices, FixedValues, model):
# self.lower_limits = self.obj_func.min_bounds
# self.upper_limits = self.obj_func.max_bounds
#else:
# raise Exception(Problem,'Not Exist'.format(x))
def objectives(self, decision_variables) -> list:
"""Use this method to calculate objective functions.
Args:
decision_variables: a sample
"""
x = decision_variables
numSample, self.num_of_variables = np.shape(np.matrix(x))
epsilon = 0.1
P1 = np.array([1, 1, 1])
P2 = np.array([1, -1, -1])
P3 = np.array([1, 1, -1])
P4 = np.array([1, -1])
P5 = np.array([-1, 1])
Phi1 = ((x[0:3] - np.ones((numSample,1)) * P1) ** 2).sum(axis=1)
Phi2 = ((x[0:3] - np.ones((numSample,1)) * P2) ** 2).sum(axis=1)
Phi3 = ((x[0:3] - np.ones((numSample,1)) * P3) ** 2).sum(axis=1)
Phi4 = ((x[3:5] - np.ones((numSample,1)) * P4) ** 2).sum(axis=1)
Phi5 = ((x[3:5] - np.ones((numSample,1)) * P5) ** 2).sum(axis=1)
Output = np.empty((numSample,5))
Output[:,0] = Phi1 + epsilon * Phi4
Output[:,1] = Phi2 + epsilon * Phi5
Output[:,2] = Phi3 + epsilon * (Phi4 + Phi5)
Output[:,3] = Phi4 + epsilon * Phi1
Output[:,4] = Phi5 + epsilon * (Phi1 + Phi2)
return Output #objective_values
name = "P_ALBERTO"
#k = 10
numobj = 5
numconst = 0
numvar = 5
problem = newProblem(name, numvar, numobj, numconst, )
lattice_resolution = 4
population_size = 105
pop = Population(problem)
pop.evolve(RVEA)
pop.non_dominated()
refpoint = 2
volume = 2 ** numobj
#print(pop.hypervolume(refpoint) / volume)
#------------------------------
def P_ALBERTO(self, decision_variables):
"""
An example in section 4.1 of the ANOVA-MOP paper (p 3280)
"""
x = decision_variables
numSample, self.num_of_variables = np.shape(decision_variables)
epsilon = 0.007 #0.1
P1 = np.array([1, 1, 1])
P2 = np.array([1, -1, -1])
P3 = np.array([1, 1, -1])
P4 = np.array([1, -1])
P5 = np.array([-1, 1])
Phi1 = ((x[:,0:3] - np.ones((numSample,1)) * P1) ** 2).sum(axis=1)
Phi2 = ((x[:,0:3] - np.ones((numSample,1)) * P2) ** 2).sum(axis=1)
Phi3 = ((x[:,0:3] - np.ones((numSample,1)) * P3) ** 2).sum(axis=1)
Phi4 = ((x[:,3:5] - np.ones((numSample,1)) * P4) ** 2).sum(axis=1)
Phi5 = ((x[:,3:5] - np.ones((numSample,1)) * P5) ** 2).sum(axis=1)
Output = np.empty((numSample,5))
Output[:,0] = Phi1 + epsilon * Phi4
Output[:,4] = Phi2 + epsilon * Phi5
Output[:,2] = Phi3 + epsilon * (Phi4 + Phi5)
Output[:,3] = Phi4 + epsilon * Phi1
Output[:,1] = Phi5 + epsilon * (Phi1 + Phi2)
return Output
def P_Surrogate(self, M,Input, SubProblemObjectiveIndices,SubProblemVariablesIndices,NumVar,Bounds,lb,ub,FixedIndices, FixedValues, model):
"""
Surrogate problem, use in solving the subproblems via RVEA
"""
# MaxValue = Bounds[0,:]
# MinValue = Bounds[1,:]
NumPop = Input.shape[0]
InputTemp = np.zeros((NumPop,len(SubProblemVariablesIndices) + len(FixedIndices)))
InputTemp[:,FixedIndices] = np.matlib.repmat(FixedValues,NumPop,1)
InputTemp[:,SubProblemVariablesIndices] = Input
Input = MapSamples(InputTemp, np.vstack((-np.ones((1,len(lb))), np.ones((1,len(lb))))), np.vstack((lb,ub)))
Output = np.zeros((NumPop,M))
for objective in range(M):
Output[:,objective] = SurrogatePrediction(Input,SurrogateDataInfo[SubProblemObjectiveIndices[objective]])
return Output
def SurrogatePrediction(x0, model): # model stored as Data includes e.g. md, check3, P, MaxIntOrder
"""
Evaluate the objective functions for the given solution x0
"""
md = model.md
check3 = model.check3
P = model.P
MaxIntOrder = model.MaxIntOrder
x0 = MultivariateLegendre2(x0,P,MaxIntOrder)
x0 = x0[:,check3]
Pred = np.matmul(x0,md) # check if the right product has been used here
return Pred
#------------------------
def objectives(self, decision_variables):
"""Objectives function to use in optimization.
Parameters
----------
decision_variables : ndarray
The decision variables
Returns
-------
objectives : ndarray
The objective values
"""
objectives = []
for obj in self.y:
objectives.append(
self.models[obj][0].predict(decision_variables.reshape(1, -1))[0]
)
return objectives
"""
------------------------------------------------------------------------
draft from SubProblem
"""
def SubProblem_objFun(x0,SubProblemObjectiveIndices,SubProblemVariablesIndices,FixedIndices,FixedValues,VariableBounds,model):
"""
#x=x0
"""
numPop = x0.shape[0]
frange = VariableBounds[1,:] - VariableBounds[0,:] # ub - lb
np.delete(frange, FixedIndices,0) # remove a column FixedIndices
x0 = ((x0+1) * np.matlib.repmat(frange,numPop,1)) / 2 + np.matlib.repmat(VariableBounds[0,SubProblemVariablesIndices],numPop,1)
numVar = len(SubProblemVariablesIndices) #+len(FixedIndices)
x = np.zeros((numPop,numVar))
x[:,SubProblemVariablesIndices] = x0
x[:,FixedIndices] = FixedValues
numObj = len(SubProblemObjectiveIndices)
y = np.zeros(1,numObj)
cons = [] # check if it should not be a list
for objective in range(numObj):
y[objective] = SurrogatePrediction(x,model(SubProblemObjectiveIndices[objective]))
return (y, cons)
"""
----------------------------------------------------------------------------
from CheckDecomposability
"""
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sat May 18 11:43:00 2019
@author: alberto
Hi Babooshka,
if I am correct, you need the determination of the rows and columns of a matrix that can form a diagonal block after reordering the rows and columns.
This seems that can be done with an algorithm finding CONNECTED COMPONENTS or CLUSTERS in a graph.
The graph is a bipartite graph where the two sets of nodes are inputs and outputs.
There is a package called PYTHON-IGRAPH that should do both things. It should suffice to rewrite the incidence matrix for the scope and then extract the components.
If you like I can try if it works for our purposes.
-----------
I implemented the solution with a simpler package (scipy) without extra installations.
Here it is.
Should be self explanatory. Refer to incidence matrices as in figure 1 of anovamop paper.
Let me know if it is understandable and working as you expected.
"""
from scipy.sparse import csr_matrix
from scipy.sparse.csgraph import connected_components
bgraph = [
[ 0, 1 , 0 ],
[ 1, 0 , 1 ]
]
print("\n\nincidence matrix: \n")
for row in bgraph :
print(row)
print("\n(rows: output variables, columns: input variables)\n")
nout = len(bgraph) # Number of columns
nin = len(bgraph[0]) # Number of rows
print("in: ",nin," out:",nout,"\n")
# transformation in a suitable graph with ninput + noutput nodes
graph = []
for rigax in range(nin) :
row = []
for x in range(nin) : row.append(0)
for y in range(nout) : row.append(bgraph[y][rigax])
graph.append(row)
for rigay in range(nout) :
row = []
for x in range(nin) : row.append(bgraph[rigay][x])
for y in range(nout) : row.append(0)
graph.append(row)
graph = csr_matrix(graph)
#print(graph)
n_components, labels = connected_components(csgraph=graph, directed=False, return_labels=True)
print("number of components found: ",n_components)
print("components legend : ",labels,"\n")
components = []
for com in range(n_components) :
invarscomp = []
outvarscomp =[]
for inv in range(nin) :
if labels[inv] == com : invarscomp.append(inv)
for outv in range(nout) :
if labels[nin+outv] == com : outvarscomp.append(outv)
components.append([invarscomp,outvarscomp])
print(com+1,"° component, input vars: ",invarscomp," output vars: ",outvarscomp,"\n" )
print("components summary: ",components)
"""
class newProblem(BaseProblem):
"""New problem description."""
def objectives(self, decision_variables):
return dtlz.DTLZ3(self.num_of_objectives,
self.num_of_variables)(decision_variables)
name = "DTLZ3"
k = 10
numobj = 3
numconst = 0
numvar = numobj + k - 1
problem = newProblem(name, numvar, numobj, numconst)
lattice_resolution = 4
population_size = 105
pop = Population(problem,
crossover_type="simulated_binary_crossover",
mutation_type="bounded_polynomial_mutation",
plotting=True)
pop.evolve(RVEA)
pop.non_dominated()
refpoint = 2
volume = 2**numobj
print(pop.hypervolume(refpoint) / volume)
name = "ANOVAMOPtest1" #k = 10 numobj = 5 numconst = 0 numvar = 5 problem = ANOVAMOPtest1(name, numvar, numobj, numconst) lattice_resolution = 4 population_size = 105 pop = Population(problem) #pop = Population( # problem, # crossover_type="simulated_binary_crossover", # mutation_type="bounded_polynomial_mutation", #) pop.evolve(RVEA) non_dom_index = pop.non_dominated() xParetoTemp = pop.individuals[non_dom_index[0]] xsize = np.shape(xParetoTemp) fParetoTemp = pop.fitness[non_dom_index[0]] xsize = np.shape(xParetoTemp) #refpoint = 2 #volume = 2 ** numobj #print(pop.hypervolume(refpoint) / volume)