def max_variance(self,XB, number = 1, exclude = []):
assert XB.shape[0] == self.ndim, 'wrong dimension'
X_min = [0.0]*len(XB)
X_min[0] = XB[0][0]-1.0
while not is_in(X_min,XB):
obj = self.objective
prob = pyOpt.Optimization('Variance Maximization',obj)
for ix in range(XB.shape[0]):
prob.addVar('X%i'%ix,'c',lower=XB[ix,0],upper=XB[ix,1],value=0.)
prob.addObj('Estimated Variance')
opt_ALPSO = pyOpt.ALPSO(pll_type=None)
#opt_ALPSO = pyOpt.ALPSO(pll_type='MP',args=[1.0])
opt_ALPSO.setOption('fileout',0)
opt_ALPSO.setOption('maxOuterIter',10)
opt_ALPSO.setOption('stopCriteria',1)
# opt_ALPSO.setOption('SwarmSize',self.ndim*100)
opt_ALPSO.setOption('SwarmSize',self.ndim*20)
opt_SLSQP = pyOpt.SLSQP()
opt_SLSQP.setOption('IPRINT',-1)
opt_SLSQP.setOption('ACC',1e-5)
vec = []
for index in range(number*10):
print index+1, ' so far: ', len(vec)
[YI_min,X_min,Info] = opt_ALPSO(prob)
[YI_min,X_min,Info] = opt_SLSQP(prob.solution(index),sens_type='FD')
if not is_already_in(X_min,vec+exclude) and is_in(X_min,XB): vec.append(X_min.tolist())
if len(vec) >= number: break
if len(vec) == 1: return vec[0]
return vec
'c',
lower=lower_bound[6],
upper=upper_bound[6],
value=p1[6])
opt_prob.addVar('nua',
'c',
lower=lower_bound[7],
upper=upper_bound[7],
value=p1[7])
opt_prob.addObj('f')
if myrank == 0:
print opt_prob
#optimize
psqp = pyOpt.ALPSO(pll_type='DPM')
psqp.setOption('printOuterIters', 1)
#psqp.setOption('maxOuterIter',1)
#psqp.setOption('stopCriteria',0)
psqp.setOption('SwarmSize', swarmSize)
psqp(opt_prob)
print opt_prob.solution(0)
popt = numpy.zeros(len(p1))
for i in opt_prob._solutions[0]._variables:
popt[i] = opt_prob._solutions[0]._variables[i].__dict__['value']
model = func_ex(popt, ns, pts_l)
ll_opt = dadi.Inference.ll_multinom(model, data)
mu = 3.5e-9
if myrank == 0:
def runOptimizer(self, opt_prob):
# type: (pyOpt.Optimization) -> np._ArrayLike[float]
''' call global followed by local optimizer, return solution '''
import pyOpt
initial = [v.value for v in list(opt_prob.getVarSet().values())]
if self.config['useGlobalOptimization']:
### optimize using pyOpt (global)
sr = random.SystemRandom()
if self.config['globalSolver'] == 'NSGA2':
if parallel:
opt = pyOpt.NSGA2(pll_type='POA') # genetic algorithm
else:
opt = pyOpt.NSGA2()
if self.config['globalOptSize'] % 4:
raise IOError("globalOptSize needs to be a multiple of 4 for NSGA2")
opt.setOption('PopSize', self.config['globalOptSize']) # Population Size (a Multiple of 4)
opt.setOption('maxGen', self.config['globalOptIterations']) # Maximum Number of Generations
opt.setOption('PrintOut', 0) # Flag to Turn On Output to files (0-None, 1-Subset, 2-All)
opt.setOption('xinit', 1) # Use Initial Solution Flag (0 - random population, 1 - use given solution)
opt.setOption('seed', sr.random()) # Random Number Seed 0..1 (0 - Auto based on time clock)
#pCross_real 0.6 Probability of Crossover of Real Variable (0.6-1.0)
opt.setOption('pMut_real', 0.5) # Probablity of Mutation of Real Variables (1/nreal)
#eta_c 10.0 # Distribution Index for Crossover (5-20) must be > 0
#eta_m 20.0 # Distribution Index for Mutation (5-50) must be > 0
#pCross_bin 0.0 # Probability of Crossover of Binary Variable (0.6-1.0)
#pMut_real 0.0 # Probability of Mutation of Binary Variables (1/nbits)
self.iter_max = self.config['globalOptSize']*self.config['globalOptIterations']
elif self.config['globalSolver'] == 'ALPSO':
if parallel:
opt = pyOpt.ALPSO(pll_type='SPM') #augmented lagrange particle swarm optimization
else:
opt = pyOpt.ALPSO() #augmented lagrange particle swarm optimization
opt.setOption('stopCriteria', 0) # stop at max iters
opt.setOption('dynInnerIter', 1) # dynamic inner iter number
opt.setOption('maxInnerIter', 5)
opt.setOption('maxOuterIter', self.config['globalOptIterations'])
opt.setOption('printInnerIters', 1)
opt.setOption('printOuterIters', 1)
opt.setOption('SwarmSize', self.config['globalOptSize'])
opt.setOption('xinit', 1)
opt.setOption('seed', sr.random()*self.mpi_size) #(self.mpi_rank+1)/self.mpi_size)
#opt.setOption('vcrazy', 1e-2)
#TODO: how to properly limit max number of function calls?
# no. func calls = (SwarmSize * inner) * outer + SwarmSize
self.iter_max = opt.getOption('SwarmSize') * opt.getOption('maxInnerIter') * \
opt.getOption('maxOuterIter') + opt.getOption('SwarmSize')
self.iter_max = self.iter_max // self.mpi_size
else:
print("Solver {} not defined".format(self.config['globalSolver']))
sys.exit(1)
# run global optimization
#try:
#reuse history
# opt(opt_prob, store_hst=False, hot_start=True) #, xstart=initial)
#except NameError:
if self.config['verbose']:
print('Running global optimization with {}'.format(self.config['globalSolver']))
self.is_global = True
opt(opt_prob, store_hst=False) #, xstart=initial)
if self.mpi_rank == 0:
print(opt_prob.solution(0))
self.gather_solutions()
### pyOpt local
if self.config['useLocalOptimization']:
print("Runnning local gradient based solver")
# TODO: run local optimization for e.g. the three last best results (global solutions
# could be more or less optimal within their local minima)
# after using global optimization, refine solution with gradient based method init
# optimizer (more or less local)
if self.config['localSolver'] == 'SLSQP':
opt2 = pyOpt.SLSQP() #sequential least squares
opt2.setOption('MAXIT', self.config['localOptIterations'])
if self.config['verbose']:
opt2.setOption('IPRINT', 0)
elif self.config['localSolver'] == 'IPOPT':
opt2 = pyOpt.IPOPT()
opt2.setOption('linear_solver', 'ma57') #mumps or hsl: ma27, ma57, ma77, ma86, ma97 or mkl: pardiso
opt2.setOption('max_iter', self.config['localOptIterations'])
if self.config['verbose']:
opt2.setOption('print_level', 4) #0 none ... 5 max
else:
opt2.setOption('print_level', 0) #0 none ... 5 max
elif self.config['localSolver'] == 'PSQP':
opt2 = pyOpt.PSQP()
opt2.setOption('MIT', self.config['localOptIterations']) # max iterations
#opt2.setOption('MFV', ??) # max function evaluations
elif self.config['localSolver'] == 'COBYLA':
if parallel:
opt2 = pyOpt.COBYLA(pll_type='POA')
else:
opt2 = pyOpt.COBYLA()
opt2.setOption('MAXFUN', self.config['localOptIterations']) # max iterations
opt2.setOption('RHOBEG', 0.1) # initial step size
if self.config['verbose']:
opt2.setOption('IPRINT', 2)
self.iter_max = self.local_iter_max
# use best constrained solution from last run (might be better than what solver thinks)
if len(self.last_best_sol) > 0:
for i in range(len(opt_prob.getVarSet())):
opt_prob.getVar(i).value = self.last_best_sol[i]
if self.config['verbose']:
print('Runing local optimization with {}'.format(self.config['localSolver']))
self.is_global = False
if self.config['localSolver'] in ['COBYLA', 'CONMIN']:
opt2(opt_prob, store_hst=False)
else:
if parallel:
opt2(opt_prob, sens_step=0.1, sens_mode='pgc', store_hst=False)
else:
opt2(opt_prob, sens_step=0.1, store_hst=False)
self.gather_solutions()
if self.mpi_rank == 0:
sol = opt_prob.solution(0)
print(sol)
#sol_vec = np.array([sol.getVar(x).value for x in range(0,len(sol.getVarSet()))])
if len(self.last_best_sol) > 0:
print("using last best constrained solution instead of given solver solution.")
print("testing final solution")
#self.iter_cnt = 0
self.objectiveFunc(self.last_best_sol, test=True)
print("\n")
return self.last_best_sol
else:
print("No feasible solution found!")
sys.exit(-1)
else:
# parallel sub-processes, close
sys.exit(0)
def identifyFeasibleStdFromFeasibleBase(self, xBase):
self.xBase_feas = xBase
# formulate problem as objective function
if self.idf.opt['optInFeasibleParamSpace']:
opt = pyOpt.Optimization('Constrained OLS',
self.minimizeSolToCADFeasible)
else:
opt = pyOpt.Optimization('Constrained OLS',
self.minimizeSolToCADStd)
opt.addObj('u')
'''
x_cons = self.mapStdToConsistent(self.idf.model.xStd[self.start_param:self.idf.model.num_model_params])
test = self.mapConsistentToStd(x_cons)
print(test - self.idf.model.xStd[self.start_param:self.idf.model.num_model_params])
'''
self.addVarsAndConstraints(opt)
# set previous sol as starting point (as primal should be already within constraints for
# most solvers to perform well)
if self.idf.opt['optInFeasibleParamSpace']:
x_cons = self.mapStdToConsistent(
self.idf.model.xStd[self.start_param:self.idf.model.
num_model_params])
for i in range(len(opt.getVarSet())):
#atm, we have 16*no_link + n_dof vars
if i < len(x_cons):
opt.getVar(i).value = x_cons[i]
else:
j = i - len(x_cons)
opt.getVar(i).value = self.model.xStd[
self.idf.model.num_model_params + j]
else:
for i in range(len(opt.getVarSet())):
opt.getVar(i).value = self.model.xStd[i + self.start_link *
self.per_link]
if self.idf.opt['verbose']:
print(opt)
if self.idf.opt['nlOptSolver'] == 'IPOPT':
# not necessarily deterministic
if self.idf.opt['verbose']:
print('Using IPOPT')
solver = pyOpt.IPOPT()
#solver.setOption('linear_solver', 'ma97') #mumps or hsl: ma27, ma57, ma77, ma86, ma97 or mkl: pardiso
#for details, see http://www.gams.com/latest/docs/solvers/ipopt/index.html#IPOPTlinear_solver
solver.setOption('max_iter', self.idf.opt['nlOptMaxIterations'])
solver.setOption('print_level', 3) #0 none ... 5 max
#don't start too far away from inital values (boundaries push even if starting inside feasible set)
solver.setOption('bound_push', 0.0000001)
solver.setOption('bound_frac', 0.0000001)
#don't relax bounds
solver.setOption('bound_relax_factor', 0.0) #1e-16)
elif self.idf.opt['nlOptSolver'] == 'SLSQP':
# solve optimization problem
if self.idf.opt['verbose']:
print('Using SLSQP')
solver = pyOpt.SLSQP(disp_opts=True)
solver.setOption('MAXIT', self.idf.opt['nlOptMaxIterations'])
if self.idf.opt['verbose']:
solver.setOption('IPRINT', 0)
elif self.idf.opt['nlOptSolver'] == 'PSQP':
# solve optimization problem
if self.idf.opt['verbose']:
print('Using PSQP')
solver = pyOpt.PSQP(disp_opts=True)
solver.setOption('MIT', self.idf.opt['nlOptMaxIterations'])
if self.idf.opt['verbose']:
solver.setOption('IPRINT', 0)
elif self.idf.opt['nlOptSolver'] == 'ALPSO':
if self.idf.opt['verbose']:
print('Using ALPSO')
solver = pyOpt.ALPSO(disp_opts=True)
solver.setOption('stopCriteria', 0)
solver.setOption('dynInnerIter', 1) # dynamic inner iter number
solver.setOption('maxInnerIter', 5)
solver.setOption('maxOuterIter',
self.idf.opt['nlOptMaxIterations'])
solver.setOption('printInnerIters', 0)
solver.setOption('printOuterIters', 0)
solver.setOption('SwarmSize', 100)
solver.setOption('xinit', 1)
elif self.idf.opt['nlOptSolver'] == 'NSGA2':
if self.idf.opt['verbose']:
print('Using NSGA2')
solver = pyOpt.NSGA2(disp_opts=True)
solver.setOption('PopSize',
100) # Population Size (a Multiple of 4)
solver.setOption('maxGen', self.config['nlOptMaxIterations']
) # Maximum Number of Generations
solver.setOption(
'PrintOut',
0) # Flag to Turn On Output to files (0-None, 1-Subset, 2-All)
solver.setOption(
'xinit', 1
) # Use Initial Solution Flag (0 - random population, 1 - use given solution)
#solver.serion('seed', sr.random()) # Random Number Seed 0..1 (0 - Auto based on time clock)
#pCross_real 0.6 Probability of Crossover of Real Variable (0.6-1.0)
solver.setOption(
'pMut_real',
0.5) # Probablity of Mutation of Real Variables (1/nreal)
#eta_c 10.0 # Distribution Index for Crossover (5-20) must be > 0
#eta_m 20.0 # Distribution Index for Mutation (5-50) must be > 0
#pCross_bin 0.0 # Probability of Crossover of Binary Variable (0.6-1.0)
#pMut_real 0.0 # Probability of Mutation of Binary Variables (1/nbits)
else:
print('Solver unknown')
self.opt_prob = opt
solver(opt) #run optimizer
# set best solution again (is often different than final solver solution)
if self.last_best_x is not None:
for i in range(len(opt.getVarSet())):
opt.getVar(i).value = self.last_best_x[i]
else:
self.last_best_x = self.model.xStd[self.start_param:]
sol = opt.solution(0)
if self.idf.opt['verbose']:
print(sol)
if self.idf.opt['optInFeasibleParamSpace'] and len(
self.last_best_x) > len(self.model.xStd[self.start_param:]):
# we get consistent parameterized params as solution
x_std = self.mapConsistentToStd(self.last_best_x)
self.model.xStd[self.start_param:self.idf.model.
num_model_params] = x_std
else:
# we get std vars as solution
self.model.xStd[self.start_param:] = self.last_best_x
def optimizePyopt(self, solver, sens, options):
# pyOpt requires empty options to be specified as {}, not None
if options is None: options = {}
eqConstr = ('SLSQP' not in solver and 'CONMIN' not in solver
and 'COBYLA' not in solver and 'FILTERSD' not in solver
and 'SDPEN' not in solver)
# Organize constraints
indLinEq = []
indLinIneq = []
indNonlinEq = []
indNonlinIneq = []
for k, bnd in enumerate(self.linBounds):
if bnd[0] == bnd[1] and eqConstr: indLinEq.append(k)
else: indLinIneq.append(k)
indLinEq = np.array(indLinEq)
indLinIneq = np.array(indLinIneq)
for k, bnd in enumerate(self.nonlinBounds):
if bnd[0] == bnd[1] and eqConstr: indNonlinEq.append(k)
else: indNonlinIneq.append(k)
indNonlinEq = np.array(indNonlinEq)
indNonlinIneq = np.array(indNonlinIneq)
# pyOpt objective
def objectivePyopt(xIn, *args, **kwargs):
x = xIn[:self.nvar]
f = self.objective(x)
g = np.zeros(0, dtype=float)
if len(indLinEq) > 0:
g = np.r_[g, np.dot(self.linMat[indLinEq, :], x)]
if len(indNonlinEq) > 0:
g = np.r_[g,
self.evalNonlinConstraints(x, 'constr', indNonlinEq)]
if len(indLinIneq) > 0:
g = np.r_[g, np.dot(self.linMat[indLinIneq, :], x)]
if len(indNonlinIneq) > 0:
g = np.r_[
g,
self.evalNonlinConstraints(x, 'constr', indNonlinIneq)]
fail = 0
if f >= self.inf or np.any(g >= self.inf): fail = 1
return f, g, fail
# pyOpt gradient
def gradientPyopt(xIn, f, g, *args, **kwargs):
x = xIn[:self.nvar]
df = self.gradient(x)
dg = np.zeros((0, self.nvar), dtype=float)
if len(indLinEq) > 0:
dg = np.r_[dg, self.linMat[indLinEq, :]]
if len(indNonlinEq) > 0:
dg = np.r_[dg,
self.evalNonlinConstraints(x, 'jac', indNonlinEq)]
if len(indLinIneq) > 0:
dg = np.r_[dg, self.linMat[indLinIneq, :]]
if len(indNonlinIneq) > 0:
dg = np.r_[dg,
self.evalNonlinConstraints(x, 'jac', indNonlinIneq)]
fail = 0
if f >= self.inf or np.any(g >= self.inf): fail = 1
return df.reshape((1, -1)), dg, fail
# Instantiate optimization problem
optProb = pyOpt.Optimization('pyopt', objectivePyopt)
# Add objective
optProb.addObj('objective')
# Add variables
optProb.addVarGroup('var',
self.nvar,
type='c',
value=self.varInit,
lower=self.varBounds[:, 0],
upper=self.varBounds[:, 1])
# Add constraints
if len(indLinEq) > 0:
optProb.addConGroup('lin-equality',
len(indLinEq),
type='e',
equal=self.linBounds[indLinEq, 0])
if len(indNonlinEq) > 0:
optProb.addConGroup('nonlin-equality',
len(indNonlinEq),
type='e',
equal=self.nonlinBounds[indNonlinEq, 0])
if len(indLinIneq) > 0:
optProb.addConGroup('lin-inequality',
len(indLinIneq),
type='i',
lower=self.linBounds[indLinIneq, 0],
upper=self.linBounds[indLinIneq, 1])
if len(indNonlinIneq) > 0:
optProb.addConGroup('nonlin-inequality',
len(indNonlinIneq),
type='i',
lower=self.nonlinBounds[indNonlinIneq, 0],
upper=self.nonlinBounds[indNonlinIneq, 1])
# Setup solver
if 'SNOPT' in solver:
optimizer = pyOpt.SNOPT(options=options)
if 'SLSQP' in solver:
optimizer = pyOpt.SLSQP(options=options)
if 'CONMIN' in solver:
optimizer = pyOpt.CONMIN(options=options)
if 'ALGENCAN' in solver:
optimizer = pyOpt.ALGENCAN(options=options)
if 'ALPSO' in solver:
optimizer = pyOpt.ALPSO(options=options)
if 'ALHSO' in solver:
optimizer = pyOpt.ALHSO(options=options)
if 'COBYLA' in solver:
optimizer = pyOpt.COBYLA(options=options)
if 'FILTERSD' in solver:
optimizer = pyOpt.FILTERSD(options=options)
if 'KOPT' in solver:
optimizer = pyOpt.KOPT(options=options)
if 'MIDACO' in solver:
optimizer = pyOpt.MIDACO(options=options)
if 'KSQP' in solver:
optimizer = pyOpt.KSQP(options=options)
if 'SDPEN' in solver:
optimizer = pyOpt.SDPEN(options=options)
if 'SOLVOPT' in solver:
optimizer = pyOpt.SOLVOPT(options=options)
# Run optimization
if sens == 'finite-difference':
optimizer(optProb, sens_type='FD')
else:
optimizer(optProb, sens_type=gradientPyopt)
# Extract solution
j = len(optProb._solutions) - 1
xStar = np.zeros(self.nvar)
for k in range(self.nvar):
xStar[k] = optProb._solutions[j].getVar(k).value
return xStar, objectivePyopt(xStar)[0]
lower=0.0,
upper=1.5 * max(Data),
value=1.0 * max(Data))
opt_prob.addVar('LifeTime', 'c', lower=0.000000001, upper=30.0, value=15.0)
opt_prob.addVar('Offset', 'c', lower=0.0, upper=1.0 * Data[0], value=0.0)
opt_prob.addVar('ScatterAmp',
'c',
lower=0.0,
upper=0.5 * max(Data) / max(AvgIRF),
value=0.1 * max(Data) / max(AvgIRF))
opt_prob.addObj('Minimize Chi**2', value=0.0, optimum=0.0)
print opt_prob
#optimizer = pyOpt.SDPEN()
min_func = pyOpt.ALPSO()
min_func.setOption('printOuterIters', 1000)
min_func.setOption('printInnerIters', 1000)
min_func.setOption('SwarmSize', 1000)
[fstr, xstr, inform] = min_func(opt_prob, disp_opts=True)
print "!------------------------------!"
print opt_prob.solution(0)
print "!------------------------------!"
print fstr
print "!------------------------------!"
print xstr
print "!------------------------------!"
print inform
