def solve (p):
global P
P = p
t0 = time.time()
# Create the problem for Ipopt to solve
nlp = pyipopt.create(P.nvar, P.x_L, P.x_U, P.ncon, P.g_L, P.g_U, P.nnjz, P.nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
# Set Ipopt solve options
nlp.int_option("print_level",5)
nlp.num_option("acceptable_tol",1000.0)
nlp.num_option("tol",1000.0)
nlp.num_option("dual_inf_tol",10000.0)
nlp.num_option("compl_inf_tol",10000.0)
nlp.int_option("max_iter", 1000)
#nlp.str_option("derivative_test","first-order")
stitching_constraint(0.0, 0)
print "Calling solve"
x, zl, zu, constrain_mu, obj, status = nlp.solve(P.x) # Solve the NLP problem with Ipopt
nlp.close()
if status == 0:
solve_time = time.time() - t0
print "Solution found in: ",solve_time," (sec)"
P.x = x
P.solved = True
else:
print "Failed to find solution!"
P.x = x
P.solved = False
def solve_init(self, eval_J, eval_dJ):
nvar = int(2*self.num_cables)
ncon = int(self.num_cables + self.num_cables*(self.num_cables-1)/2)
low_var = -numpy.inf*numpy.ones(nvar,dtype=float)
up_var = numpy.inf*numpy.ones(nvar, dtype=float)
up_var[0], low_var[0] = 0, 0
inf_con = numpy.inf*numpy.ones(ncon, dtype=float)
zero_con = numpy.zeros(ncon, dtype=float)
self.nlp = pyipopt.create(nvar, # Number of controls
low_var, # Lower bounds for Control
up_var, # Upper bounds for Control
ncon, # Number of constraints
-inf_con, # Lower bounds for contraints
zero_con, # Upper bounds for contraints
nvar*ncon, # Number of nonzeros in cons. Jac
0, # Number of nonzeros in cons. Hes
lambda pos: eval_J(pos), # Objective eval
lambda pos: eval_dJ(pos), # Obj. grad eval
self.eval_g, # Constraint evaluation
self.eval_jac_g # Constraint Jacobian evaluation
)
# So it does not violate the boundary constraint
self.nlp.num_option('bound_relax_factor', 0)
self.nlp.int_option('max_iter', 16)
self.nlp.int_option("print_level",6)
def run(x0, eval_jac_g, eval_h):
if x0.shape != (N*D,):
raise "x is wrong dims!"
x_L, x_U = set_x_bounds(lowbnd, upbnd)
start = time.time()
print eval_h.nnz
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, eval_jac_g.nnz, eval_h.nnz, eval_f_adolc, eval_grad_f, eval_g_adolc, eval_jac_g, eval_h)
#nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, eval_jac_g.nnz, 365, eval_f_adolc, eval_grad_f, eval_g_adolc, eval_jac_g)
nlp.int_option('max_iter', maxits)
nlp.num_option('constr_viol_tol', epsg)
nlp.num_option('tol', epsf)
nlp.num_option('acceptable_tol', 1e-3)
nlp.str_option('linear_solver', linear_solver)
nlp.str_option('mu_strategy', 'adaptive')
nlp.num_option('bound_relax_factor', 0)
nlp.str_option('adaptive_mu_globalization','never-monotone-mode')
results = nlp.solve(x0)
print "optimized: ", time.time()-start, "s"
print "Exit flag = ", results[5]
print "Action = ", results[4]
return results[4], results[0]
def runOptimization(self):
pyipopt.set_loglevel(0)
self.nlp = pyipopt.create(
self.optDict['nPrimal'],
self.constraintDict['lowerBound'],
self.constraintDict['upperBound'],
self.optDict['nConstraint'],
self.constraintDict['lowerConstraint'],
self.constraintDict['upperConstraint'],
self.optDict['nJacobianNonZero'],
self.optDict['nHessianNonZero'],
self.evalObjectiveFunction,
self.evalObjectiveFunctionGradient,
self.evalConstraints,
self.evalConstraintsJacbobian)
self.nlp.int_option("max_iter", 150)
# self.nlp.str_option('derivative_test', 'first-order')
# self.nlp.num_option('derivative_test_tol', 1e-2)
# self.nlp.num_option('derivative_test_perturbation', 1e-3)
self.nlp.num_option('point_perturbation_radius', 1e-3)
# self.nlp.int_option('max_iter', 100)
# self.nlp.str_option('derivative_test_print_all', 'yes')
# self.nlp.int_option('derivative_test_first_index', 330)
# self.nlp.num_option('acceptable_constr_viol_tol', 1)
x, zl, zu, constraint_multipliers, obj, status = self.nlp.solve(self.initialGuess)
return x[self.sliceDict['varMu']], x[self.sliceDict['varTheta']], self.optX, self.optY
def optimizePC(channel, noiseIfPower, rate, linkBandwidth, pMax, p0, m, verbosity=0):
''' Uses channel values, PHY parameters and power consumption characteristics to find minimal resource allocation. Returns resource allocation, objective value and IPOPT status.
Input:
channel - 3d array. 0d users, 1d n_tx, 2d n_rx
noiseIfPower - total noise power over the linkbandwidth
rate - target rate in bps
pMax - maximum allowed transmission power
p0 - power consumption at zero transmission (not sleep)
m - power consumption load factor
verbosity - IPOPT verbosity level
Output:
obj - solution objective value
solution - resource share per user
status - IPOPT status '''
# the channel dimensions tell some more parameters
users = channel.shape[0]
n_tx = channel.shape[1]
n_rx = channel.shape[2]
# preparing IPOPT parameters
nvar = users # for readability
x_L = zeros((nvar), dtype=float_) * 0.0
x_U = ones((nvar), dtype=float_) * 1.0
ncon = nvar + 1 # transmit power constraints and the unit sum
g_L = zeros(1+nvar) # unit sum and all power constraints
g_L[0] = 1.
g_U = pMax * ones(1+nvar) # unit sum and all power constraints
g_U[0] = 1.
nnzj = nvar * (1+nvar)
nnzh = 0 #used?
x0 = repeat([1./(nvar+1)],nvar) # Starting point
# IPOPT requires single parameter functions
if n_tx is 2 and n_rx is 2:
eval_f = lambda mus: optimMinPow2x2.eval_f(mus, noiseIfPower, channel, rate, linkBandwidth, p0, m)
eval_grad_f = lambda mus: optimMinPow2x2.eval_grad_f(mus, noiseIfPower, channel, rate, linkBandwidth, p0, m)
eval_g = lambda mus: optimMinPow2x2.eval_g(mus, noiseIfPower, channel, rate, linkBandwidth)
eval_jac_g = lambda mus, flag: optimMinPow2x2.eval_jac_g(mus, noiseIfPower, channel, rate, linkBandwidth, flag)
else:
raise NotImplementedError # other combinations may be needed later
# Call solve()
pyipopt.set_loglevel(min([2,verbosity])) # verbose
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
#nlp.int_option("max_iter", 3000)
#nlp.num_option("tol", 1e-8)
#nlp.num_option("acceptable_tol", 1e-2)
#nlp.int_option("acceptable_iter", 0)
nlp.str_option("derivative_test", "first-order")
nlp.str_option("derivative_test_print_all", "no")
#nlp.str_option("print_options_documentation", "yes")
nlp.int_option("print_level", min([verbosity,12])) # maximum is 12
nlp.str_option("print_user_options", "yes")
solution, zl, zu, obj, status = nlp.solve(x0)
nlp.close()
return obj, solution, status
def ipopt_naive(self,
start=None,
iters=3000,
acc=1E-5,
callback=True,
verbose=False):
"""
x - # of variables
xl - lower bounds of variables
xu - upper bounds of variables
m - # of constraints
gl - lower bounds of constraints
gu - upper bounds of constraints
nnzj - number of nonzero values in jacobian
nnzh - number of nonzero values in hessian (set to 0 if eval_h is not used)
eval_f - objective function
eval_grad_f - calculates gradient of objective function
eval_g - calculates constraint values
eval_jac_g - calculates jacobian
eval_h - calculates hessian (optional, if not used set nnzh to 0)
"""
try:
if not (hasattr(IPOPT, "create")):
print "WARNING: IPOPT not correctly installed (not start)"
return start
except NameError:
print "WARNING: IPOPT is not installed!"
return start
if start is None:
start = self.defaultState.extract(excludeKeys=self.excludeKeys)
if callback == True:
callback = self.robot.cleanupCallback
#start = N.array(start, subok=True)
numConst, lCBounds, uCBounds = self.getConstData()
numVar = start.size
lVBounds = []
uVBounds = []
for i in range(start.size / self.robot.nvars):
lVBounds += self.robot.xLBounds
uVBounds += self.robot.xUBounds
lVBounds = N.array(lVBounds, dtype=float)
uVBounds = N.array(uVBounds, dtype=float)
struct = self.eval_jac_g(start, False)
numNonZero = struct.size
ipprob = IPOPT.create(numVar, lVBounds, uVBounds, numConst, lCBounds,
uCBounds, numNonZero, 0, self.eval_f,
self.eval_grad_f, self.eval_g, self.eval_jac_g)
ipprob.num_option('tol', acc)
ipprob.int_option('max_iter', iters)
# call solve with initial state (start)
x, zl, zu, constraint_multipliers, obj, status = ipprob.solve(start)
ipprob.close()
return x
def solve(self):
x0 = self._robot.configuration()
nbVar = len(x0)
xL = [-math.pi] * nbVar
xU = [math.pi] * nbVar
cL = []
cU = []
nCon = 0
nnzH = 0
ssR = []
ssC = []
for c in self._constraint:
cL.append(c.lBounds())
cU.append(c.uBounds())
ss = c.sparseStruct()
ssR += (np.array(ss[0]) + nCon).tolist()
ssC += ss[1]
nCon += c.constraintCount()
ss = (np.array(ssR), np.array(ssC))
nnzJ = len(ssC)
def evalF(x, user=None):
self._robot.configure(x)
return 0.0
def evalFJ(x, user=None):
self._robot.configure(x)
return np.zeros(nbVar)
def evalC(x, user=None):
self._robot.configure(x)
l = []
for c in self._constraint:
l += c.eval(x)
return np.array(l)
def evalCJ(x, flag, user=None):
if flag:
return ss
else:
l = []
for c in self._constraint:
l += c.evalJ(x)
return np.array(l)
opt = pyipopt.create(nbVar, np.array(xL), np.array(xU),
nCon, np.array(cL), np.array(cU),
nnzJ, nnzH, evalF, evalFJ, evalC, evalCJ)
x, zl, zu, obj, status = opt.solve(np.array(x0))
opt.close()
def start_iteration(self):
"""Perform initial setup before iteration loop begins."""
self._config_ipopt()
# get the initial values of the parameters
for i, val in enumerate(self.get_parameters().values()):
self.design_vals[i] = val.evaluate(self.parent)
self.update_constraints()
x_L = array( [ x.low for x in self.get_parameters().values() ] )
x_U = array( [ x.high for x in self.get_parameters().values() ] )
# Ipopt treats equality and inequality constraints together.
# For equality constraints, just set the g_l and g_u to be
# equal. For this driver, the inequality constraints come
# first. g_l is set to 0.0 and g_l is set to the largest float.
# For the equality constraints, both g_l and g_u are set to zero.
g_L = zeros( self.num_constraints, 'd')
g_U = zeros( self.num_constraints, 'd')
for i in range( self.num_ineq_constraints ):
g_U[i] = sys.float_info.max
# number of non zeros in Jacobian
nnzj = self.num_params * self.num_constraints
# of constraints. Assumed to be dense
# number of non zeros in hessian
nnzh = self.num_params * ( self.num_params + 1 ) / 2
try:
self.nlp = pyipopt.create(
self.num_params, x_L, x_U,
self.num_constraints, g_L, g_U,
nnzj, nnzh,
eval_f, eval_grad_f,
eval_g, eval_jac_g,
intermediate_callback
# f2py lets you pass extra args to
# callback functions
# http://cens.ioc.ee/projects/f2py2e/usersguide/
# index.html#call-back-arguments
# We pass the driver itself to the callbacks
### not using them for now
# eval_f_extra_args = (self,),
# eval_grad_f_extra_args = (self,),
# eval_g_extra_args = (self,),
# eval_jac_g_extra_args = (self,),
# intermediate_cb_extra_args = (self,),
)
self.nlp.set_intermediate_callback( intermediate_callback )
except Exception, err:
self._logger.error(str(err))
raise
def solve_constr(theta0, use_hess = False):
pyipopt.set_loglevel(1)
n = theta0.size
x_L = np.array([pyipopt.NLP_LOWER_BOUND_INF]*n, dtype=float)
x_U = np.array([pyipopt.NLP_UPPER_BOUND_INF]*n, dtype=float)
ncon = pstationtrue.size
g_L = g_U = pstationtrue
nnzj = maskj.sum()
nnzh = maskh.sum()
idxrj, idxcj = np.mgrid[:ncon, :n]
idxrh, idxch = np.mgrid[:n, :n]
eval_c = lambda t: constr(t)
eval_j = lambda t, f: (idxrj[maskj], idxcj[maskj]) if f else np.squeeze(jab(t))[maskj]
if use_hess:
eval_h = lambda t, l, o, f: (idxrh[maskh], idxch[maskh]) if f else np.squeeze(hess_constr(t,l,o))[maskh]
nlp = pyipopt.create(theta0.size, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad, eval_c, eval_j, eval_h)
else:
nlp = pyipopt.create(theta0.size, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad, eval_c, eval_j)
results = nlp.solve(theta0)
nlp.close()
return results
def main():
# verbose
pyipopt.set_loglevel(2)
# define the parameters and their box constraints
nvar = 2
x_L = numpy.array([-3, -3], dtype=float)
x_U = numpy.array([3, 3], dtype=float)
# define the inequality constraints
ncon = 0
g_L = numpy.array([], dtype=float)
g_U = numpy.array([], dtype=float)
# define the number of nonzeros in the jacobian and in the hessian
# there are no nonzeros in the constraint jacobian
nnzj = 0
# there are maximum nonzeros (nvar*(nvar+1))/2 in the lagrangian hessian
nnzh = 3
# create the nonlinear programming model
nlp = pyipopt.create(
nvar,
x_L,
x_U,
ncon,
g_L,
g_U,
nnzj,
nnzh,
eval_f,
eval_grad_f,
eval_g,
eval_jac_g,
eval_h,
apply_new,
)
# define the initial guess
x0 = numpy.array([-1.2, 1], dtype=float)
# compute the results using ipopt
results = nlp.solve(x0)
# free the model
nlp.close()
# report the results
print results
def main():
# verbose
pyipopt.set_loglevel(2)
# define the parameters and their box constraints
nvar = 2
x_L = numpy.array([-3, -3], dtype=float)
x_U = numpy.array([3, 3], dtype=float)
# define the inequality constraints
ncon = 0
g_L = numpy.array([], dtype=float)
g_U = numpy.array([], dtype=float)
# define the number of nonzeros in the jacobian and in the hessian
# there are no nonzeros in the constraint jacobian
nnzj = 0
# there are maximum nonzeros (nvar*(nvar+1))/2 in the lagrangian hessian
nnzh = 3
# create the nonlinear programming model
nlp = pyipopt.create(
nvar,
x_L,
x_U,
ncon,
g_L,
g_U,
nnzj,
nnzh,
eval_f,
eval_grad_f,
eval_g,
eval_jac_g,
eval_h,
apply_new,
)
# define the initial guess
x0 = numpy.array([-1.2, 1], dtype=float)
# compute the results using ipopt
results = nlp.solve(x0)
# free the model
nlp.close()
# report the results
print(results)
def __init__(self,ListStreams,ListUnits,CFlag=5,PrintOpt=0,Xtol=1e-6,iter=100):
self.XFlag=[]
self.X=[]
self.Sigma=[]
self.CFlag=CFlag
self.ListStreams=ListStreams
self.ListUints=ListUnits
self.ConstructXFlag()
#self.ConstructX()
self.Xmeas=self.X
self.JacobianSize()
if (self.CFlag==5 or self.CFlag==6 or self.CFlag==7):
self.JacoNorm()
self.NonZeroJacoRowCol()
self.NonZeroHessRowCol()
self.nnzh=len(self.NonZeroHessRow)
self.nnzj=len(self.NonZeroJacoRow)
self.ConstraintBounds()
self.nlp = pyipopt.create(self.Xlen, self.XLB, self.XUB, self.Glen, self.GLB, self.GUB, self.nnzj, self.nnzh, self.Objective, self.obj_grad, self.Constraints, self.ConstructJaco, self.Hessian)
self.nlp.int_option('print_level',PrintOpt)
self.nlp.int_option('max_iter',iter)
self.nlp.num_option('tol',Xtol)
# self.nlp.num_option('dual_inf_tol',500000)
# self.nlp.str_option('accept_every_trial_step','yes')
# '''==============Derivative Check============================'''
# self.nlp.str_option('derivative_test','only-second-order')
# self.nlp.num_option('derivative_test_perturbation',1e-6)
# self.nlp.num_option('derivative_test_tol',1e-5)
# '''================================================================='''
#self.nlp.str_option('hessian_approximation','limited-memory')
#self.nlp.str_option('nlp_scaling_method','none')
#self.nlp.num_option('nlp_scaling_min_value',0.01)
#self.nlp.str_option('alpha_for_y','full')
#self.nlp.num_option('constr_viol_tol',1e-4)
#nlp.str_option('expect_infeasible_problem','yes')
#nlp.str_option('start_with_resto','yes')
self.X0 = asarray(self.Xmeas,dtype=float_)
self.Validation()
self.ExitFlag={}
self.Xopt, self.zl, self.zu, self.constraint_multipliers, self.obj, self.status = self.nlp.solve(self.X0)
self.ExitFlag[-1]=self.status
for i in range(0):
if (self.status !=0):
self.Xopt, self.zl, self.zu, self.constraint_multipliers, self.obj, self.status = self.nlp.solve(self.Xopt)
self.ExitFlag[i]=self.status
if (self.status==0):
self.OptimIndex=i
print 'sucess ',i
break
self.nlp.close()
def solve (p, warm_start):
global P
P = p
t0 = time.time()
# Create the problem for Ipopt to solve
P.nlp = pyipopt.create(P.nvar, P.x_L, P.x_U, P.ncon, P.g_L, P.g_U, P.nnjz, P.nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
# Set Ipopt solve options
"""
P.nlp.int_option("print_level",1)
P.nlp.num_option("acceptable_tol",10.0)
P.nlp.num_option("tol",10.0)
P.nlp.num_option("dual_inf_tol",1000.0)
P.nlp.num_option("compl_inf_tol",1000.0)
P.nlp.int_option("max_iter", 10000)
P.nlp.str_option("linear_solver", "mumps")
"""
P.nlp.str_option("linear_solver", "mumps")
P.nlp.str_option("jacobian_approximation", "exact")
P.nlp.str_option("hessian_approximation", "limited-memory")
P.nlp.num_option("max_cpu_time", 40.0)
P.nlp.num_option("tol", 10.0)
P.nlp.str_option("print_timing_statistics", "no")
P.nlp.str_option("print_user_options", "no")
P.nlp.int_option("print_level",4)
if warm_start:
P.nlp.str_option('mu_strategy', 'adaptive')
P.nlp.str_option("warm_start_init_point", "yes")
#P.nlp.str_option("derivative_test","first-order")
print "Calling solve"
x, zl, zu, constrain_mu, obj, status = P.nlp.solve(P.x) # Solve the NLP problem with Ipopt
#nlp.close()
if status == 0:
solve_time = time.time() - t0
print "Solution found in: ",solve_time," (sec)"
P.x = x
P.solved = True
else:
print "Failed to find solution!"
P.x = x
P.solved = False
def __build_pyipopt_problem(self):
"""Build the pyipopt problem from the OptimizationProblem instance."""
import pyipopt
from functools import partial
self.rfn = ReducedFunctionalNumPy(self.problem.reduced_functional)
ncontrols = len(self.rfn.get_controls())
(lb, ub) = self.__get_bounds()
(nconstraints, fun_g, jac_g, clb, cub) = self.__get_constraints()
constraints_nnz = nconstraints * ncontrols
# A callback that evaluates the functional and derivative.
J = self.rfn.__call__
dJ = partial(self.rfn.derivative, forget=False)
nlp = pyipopt.create(
len(ub), # length of control vector
lb, # lower bounds on control vector
ub, # upper bounds on control vector
nconstraints, # number of constraints
clb, # lower bounds on constraints,
cub, # upper bounds on constraints,
constraints_nnz, # number of nonzeros in the constraint Jacobian
0, # number of nonzeros in the Hessian
J, # to evaluate the functional
dJ, # to evaluate the gradient
fun_g, # to evaluate the constraints
jac_g) # to evaluate the constraint Jacobian
pyipopt.set_loglevel(1) # turn off annoying pyipopt logging
if rank(self.problem.reduced_functional.mpi_comm()) > 0:
nlp.int_option('print_level',
0) # disable redundant IPOPT output in parallel
else:
nlp.int_option('print_level', 6) # very useful IPOPT output
if isinstance(self.problem, MaximizationProblem):
# multiply objective function by -1 internally in
# ipopt to maximise instead of minimise
nlp.num_option('obj_scaling_factor', -1.0)
self.pyipopt_problem = nlp
def __build_pyipopt_problem(self):
"""Build the pyipopt problem from the OptimizationProblem instance."""
import pyipopt
from functools import partial
self.rfn = ReducedFunctionalNumPy(self.problem.reduced_functional)
ncontrols = len(self.rfn.get_controls())
(lb, ub) = self.__get_bounds()
(nconstraints, fun_g, jac_g, clb, cub) = self.__get_constraints()
constraints_nnz = nconstraints * ncontrols
# A callback that evaluates the functional and derivative.
J = self.rfn.__call__
dJ = partial(self.rfn.derivative, forget=False)
nlp = pyipopt.create(
len(ub), # length of control vector
lb, # lower bounds on control vector
ub, # upper bounds on control vector
nconstraints, # number of constraints
clb, # lower bounds on constraints,
cub, # upper bounds on constraints,
constraints_nnz, # number of nonzeros in the constraint Jacobian
0, # number of nonzeros in the Hessian
J, # to evaluate the functional
dJ, # to evaluate the gradient
fun_g, # to evaluate the constraints
jac_g,
) # to evaluate the constraint Jacobian
pyipopt.set_loglevel(1) # turn off annoying pyipopt logging
if rank(self.problem.reduced_functional.mpi_comm()) > 0:
nlp.int_option("print_level", 0) # disable redundant IPOPT output in parallel
else:
nlp.int_option("print_level", 6) # very useful IPOPT output
if isinstance(self.problem, MaximizationProblem):
# multiply objective function by -1 internally in
# ipopt to maximise instead of minimise
nlp.num_option("obj_scaling_factor", -1.0)
self.pyipopt_problem = nlp
def solve_init(self, eval_J, eval_dJ):
nvar = self.num_angles
low_var = -numpy.infty * numpy.ones(nvar, dtype=float)
up_var = numpy.infty * numpy.ones(nvar, dtype=float)
self.nlp = pyipopt.create(
nvar, # Number of controls
low_var, # Lower bounds for Control
up_var, # Upper bounds for Control
0, # Number of constraints
numpy.array([], dtype=float),
numpy.array([], dtype=float),
0, # Number of nonzeros in cons. Jac
0, # Number of nonzeros in cons. Hes
lambda angle: eval_J(angle), # Objective eval
lambda angle: eval_dJ(angle), # Obj. grad eval
self.func_g,
self.jac_g)
def estimate_alpha_beta(counts, X, ini=None, verbose=0,
use_empty_entries=True):
"""
Estimate the parameters of g
Parameters
----------
counts: ndarray
use_empty_entries: boolean, optional, default: True
whether to use zeroes entries as information or not
"""
m, n = X.shape
nvar = 1
ncon = 0
nnzj = 0
nnzh = 0
x_L = np.array([- 10000.])
x_U = np.array([10000000.])
nlp = pyipopt.create(nvar, x_L, x_U, ncon, x_L,
x_U, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g)
nlp.int_option('max_iter', 100)
if ini is None:
if verbose:
print "Initial values not provided"
ini = np.random.randint(1, 100, size=(1, )) + \
np.random.random(size=(1, ))
results = nlp.solve(ini, (m, n, counts, X, use_empty_entries))
try:
x, _, _, _, _ = results
except ValueError:
x, _, _, _, _, _ = results
# Evaluate counts with new estimated model.
d = euclidean_distances(X)
mask = (np.tri(m, dtype=np.bool) == False) & (counts != 0) & (d != 0)
beta = counts[mask].sum() / (d[mask] ** x[0]).sum()
return x[0], beta
def optimizegraph(opt, max_iter=300, max_cpu_time=100):
x_L, x_U = opt.bound_variables()
g_L, g_U = opt.bound_constraints()
nlp = pyipopt.create(opt.n_var, x_L, x_U, opt.n_cons, g_L, g_U, opt.nnzj, opt.nnzh, \
opt.eval_objective, opt.eval_grad_objective, opt.eval_constraints, opt.eval_jac_constraint)
nlp.num_option('max_cpu_time', max_cpu_time)
nlp.int_option('print_level', 0)
nlp.num_option('tol', 1e-12)
nlp.num_option('acceptable_tol', 1e-12)
x, zl, zu, constraint_multipliers, obj, status = nlp.solve(
initialize_flow(opt))
nlp.close()
# reinforce flow balance
x = reinforce_flowbalance(opt, x)
# scale x
x = np.maximum(x, 0)
s = np.sum(opt.weights) / np.dot(opt.efflen, x)
opt.results = x * s
return opt, status
def _solve(self, x0, A, l, u, xmin, xmax):
""" Solves using the Interior Point OPTimizer.
"""
# Indexes of constrained lines.
il = [i for i,ln in enumerate(self._ln) if 0.0 < ln.rate_a < 1e10]
nl2 = len(il)
neqnln = 2 * self._nb # no. of non-linear equality constraints
niqnln = 2 * len(il) # no. of lines with constraints
user_data = {"A": A, "neqnln": neqnln, "niqnln": niqnln}
self._f(x0)
Jdata = self._dg(x0, False, user_data)
# Hdata = self._h(x0, ones(neqnln + niqnln), None, False, user_data)
lmbda = {"eqnonlin": ones(neqnln),
"ineqnonlin": ones(niqnln)}
H = tril(self._hessfcn(x0, lmbda), format="coo")
self._Hrow, self._Hcol = H.row, H.col
n = len(x0) # the number of variables
xl = xmin
xu = xmax
gl = r_[zeros(2 * self._nb), -Inf * ones(2 * nl2), l]
gu = r_[zeros(2 * self._nb), zeros(2 * nl2), u]
m = len(gl) # the number of constraints
nnzj = len(Jdata) # the number of nonzeros in Jacobian matrix
nnzh = 0#len(H.data) # the number of non-zeros in Hessian matrix
f_fcn, df_fcn, g_fcn, dg_fcn, h_fcn = \
self._f, self._df, self._g, self._dg, self._h
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh,
f_fcn, df_fcn, g_fcn, dg_fcn)#, h_fcn)
# print dir(nlp)
# nlp.str_option("print_options_documentation", "yes")
# nlp.int_option("max_iter", 10)
# x, zl, zu, obj = nlp.solve(x0)
success = nlp.solve(x0, user_data)
nlp.close()
def _solve(self, x0, A, l, u, xmin, xmax):
""" Solves using the Interior Point OPTimizer.
"""
# Indexes of constrained lines.
il = [i for i, ln in enumerate(self._ln) if 0.0 < ln.rate_a < 1e10]
nl2 = len(il)
neqnln = 2 * self._nb # no. of non-linear equality constraints
niqnln = 2 * len(il) # no. of lines with constraints
user_data = {"A": A, "neqnln": neqnln, "niqnln": niqnln}
self._f(x0)
Jdata = self._dg(x0, False, user_data)
# Hdata = self._h(x0, ones(neqnln + niqnln), None, False, user_data)
lmbda = {"eqnonlin": ones(neqnln), "ineqnonlin": ones(niqnln)}
H = tril(self._hessfcn(x0, lmbda), format="coo")
self._Hrow, self._Hcol = H.row, H.col
n = len(x0) # the number of variables
xl = xmin
xu = xmax
gl = r_[zeros(2 * self._nb), -Inf * ones(2 * nl2), l]
gu = r_[zeros(2 * self._nb), zeros(2 * nl2), u]
m = len(gl) # the number of constraints
nnzj = len(Jdata) # the number of nonzeros in Jacobian matrix
nnzh = 0 #len(H.data) # the number of non-zeros in Hessian matrix
f_fcn, df_fcn, g_fcn, dg_fcn, h_fcn = \
self._f, self._df, self._g, self._dg, self._h
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh, f_fcn, df_fcn,
g_fcn, dg_fcn) #, h_fcn)
# print dir(nlp)
# nlp.str_option("print_options_documentation", "yes")
# nlp.int_option("max_iter", 10)
# x, zl, zu, obj = nlp.solve(x0)
success = nlp.solve(x0, user_data)
nlp.close()
def improve_ipopt(x0, prob, *args, **kwargs):
try:
import pyipopt
except ImportError:
raise Exception("PyIpopt package is not installed.")
lb = pyipopt.NLP_LOWER_BOUND_INF
ub = pyipopt.NLP_UPPER_BOUND_INF
g_L = np.zeros(prob.m)
for i in range(prob.m):
if prob.fs[i].relop == '<=':
g_L[i] = lb
g_U = np.zeros(prob.m)
def eval_grad_f(x, user_data = None):
return 2*prob.f0.P.dot(x) + prob.f0.qarray
def eval_g(x, user_data = None):
return np.array([f.eval(x) for f in prob.fs])
jac_grid = np.indices((prob.m, prob.n))
jac_r = jac_grid[0].ravel()
jac_c = jac_grid[1].ravel()
def eval_jac_g(x, flag, user_data = None):
if flag:
return (jac_r, jac_c)
else:
return np.vstack([2*f.P.dot(x)+f.qarray for f in prob.fs])
nlp = pyipopt.create(
prob.n, lb*np.ones(prob.n), ub*np.ones(prob.n),
prob.m, g_L, g_U, prob.m*prob.n, 0,
prob.f0.eval, eval_grad_f,
eval_g, eval_jac_g
)
try:
x, zl, zu, constraint_multipliers, obj, status = nlp.solve(x0)
except:
pass
return x
def OptSolver(bounds, OptObj, max_iter=10):
"""
Function that returns a pyipopt object with OptObj atributes
"""
# Number of Design Variables
nvar = OptObj.nvars
# Upper and lower bounds
x_L = numpy.ones(nvar) * bounds[0]
x_U = numpy.ones(nvar) * bounds[1]
# Number of non-zeros gradients
constraints_nnz = nvar*OptObj.cst_num
acst_L = numpy.array(OptObj.cst_L)
acst_U = numpy.array(OptObj.cst_U)
PyIpOptObj = pyipopt.create(nvar, # number of the design variables
x_L, # lower bounds of the design variables
x_U, # upper bounds of the design variables
OptObj.cst_num, # number of constraints
acst_L, # lower bounds on constraints,
acst_U, # upper bounds on constraints,
constraints_nnz, # number of nonzeros in the constraint Jacobian
0, # number of nonzeros in the Hessian
OptObj.obj_fun, # objective function
OptObj.obj_dfun, # gradient of the objective function
OptObj.cst_fval, # constraint function
OptObj.jacobian ) # gradient of the constraint function
#Parameters
PyIpOptObj.num_option('acceptable_tol', 1.0e-10)
PyIpOptObj.num_option('eta_phi', 1e-12) # eta_phi: Relaxation factor in the Armijo condition.
PyIpOptObj.num_option('theta_max_fact', 30000) # Determines upper bound for constraint violation in the filter.
PyIpOptObj.int_option('max_soc', 20)
PyIpOptObj.int_option('max_iter', max_iter)
PyIpOptObj.int_option('watchdog_shortened_iter_trigger', 20)
PyIpOptObj.int_option('accept_after_max_steps', 5)
pyipopt.set_loglevel(1) # turn off annoying pyipopt logging
PyIpOptObj.int_option('print_level', 6) # very useful IPOPT output
return PyIpOptObj
def main():
cable_positions_numpy = numpy.array(cable_positions)
print(cable_positions)
def eval_robust_J(cable_positions):
if max(eval_g(cable_positions)) > 0:
print("Warning: Inequality constraint is not satisfied")
return numpy.inf
# print eval_g(cable_positions).T
try:
j = eval_J(cable_positions)
# c1 = numpy.array([cable_positions[0],cable_positions[1]])
# c2 = numpy.array([cable_positions[2],cable_positions[3]])
# c3 = numpy.array([cable_positions[4],cable_positions[5]])
# print("Angles between the three cables: %.2f, %.2f, %.2f"
# % (numpy.arccos(numpy.dot(c1-c2,c1-c3)/
# (numpy.sqrt(numpy.dot(c1-c2,c1-c2)
# *numpy.dot(c1-c3,c1-c3))))/(2*numpy.pi)*360,
# numpy.arccos(numpy.dot(c1-c2,c3-c2)/
# (numpy.sqrt(numpy.dot(c1-c2,c1-c2)
# *numpy.dot(c3-c2,c3-c2))))/(2*numpy.pi)*360,
# numpy.arccos(numpy.dot(c3-c1,c3-c2)/
# (numpy.sqrt(numpy.dot(c3-c1,c3-c1)
# *numpy.dot(c3-c2,c3-c2))))/(2*numpy.pi)*360))
except:
print("ERROR: Forward model failed, returning infinity")
j = numpy.inf
exit(1)
#print( "J = ", j)
return j
def eval_robust_dJ(cable_positions):
if max(eval_g(cable_positions)) > 0:
print("Warning: Inequality constraint is not satisfied")
return numpy.inf
dj = eval_dJ(cable_positions)
print("|dj| = ", numpy.dot(dj, dj)**0.5)
return dj
def test_J(cable_positions):
""" Test of optimization without state eq"""
j = 0.5 * numpy.dot(cable_positions, cable_positions)
if max(eval_g(cable_positions)) > 0:
print("Warning: Inequality constraint is not satisfied: ")
print(eval_g(cable_positions))
print(cable_positions)
print("*" * 25)
print("J = ", j)
return j
def test_dJ(cable_positions):
""" Test of optimization without state eq"""
dj = numpy.array(cable_positions)
print("|dj| = ", numpy.dot(dj, dj)**0.5)
return dj
# Check that input data match input variables
nvar = int(2 * num_cables) # Number of controls
ncon = int(num_cables + num_cables *
(num_cables - 1) / 2) # Number of inequality constraints
# Create the NLP model
#pyipopt.set_loglevel(2) # Set verbosity
nlp = pyipopt.create(
nvar, # Number of controls
-numpy.inf * numpy.ones(nvar, dtype=float), # Lower bounds of controls
numpy.inf * numpy.ones(nvar, dtype=float), # Upper bounds of controls
ncon, # Number of inequality constraints
-numpy.inf * numpy.ones(
ncon, dtype=float), # Lower bounds of inequality constraints
numpy.zeros(ncon,
dtype=float), # Upper bounds of inequality constraints
nvar * ncon, # Number of nonzeros in the constraint Jacobian
0, # Number of nonzeros in the Hessian
lambda pos: eval_robust_J(pos), # Objective evaluation
lambda pos: eval_robust_dJ(pos), # Objective gradient evaluation
# lambda pos: test_J(pos), # Objective evaluation
# lambda pos: test_dJ(pos), # Objective gradient evaluation
eval_g, # Constraint evaluation
eval_jac_g, # Constraint Jacobian evaluation
)
nlp.num_option('obj_scaling_factor',
1e-1) # 1e-2 for isoceles example as gradient is bigger
# 1e-1 for 3 cables as gradient is smaller
nlp.int_option('max_iter', 100)
nlp.num_option('acceptable_tol', 1e-2)
nlp.num_option('tol', 1e-2)
nlp.num_option('bound_relax_factor',
0) # So it does not violate the boundary constraint
# nlp.str_option('mu_strategy',"adaptive")
#nlp.str_option('nlp_scaling_method',"gradient-based")
# nlp.num_option("mu_max", 0.1)
# nlp.num_option('mu_init', 0.1)#1e25)
#nlp.num_option('bound_relax_factor', 0) # So it does not violate the boundary constraint
# nlp.num_option('acceptable_tol', 1e-5)
# Solve the optimisation problem
opt_cable_pos = nlp.solve(cable_positions_numpy)[0]
nlp.close()
# Report the results
d = opt_cable_pos.reshape(-1, 2)
for i in range(num_cables):
max_radius = outer_cable_mesh_radius - inner_cable_mesh_radius - distance_from_outer_cable_mesh
print("Cable %d distance from center: %.3e" %
(i, numpy.sqrt(d[i, 0]**2 + d[i, 1]**2)))
for j in range(i + 1, num_cables):
print("Distance between Cable %d and %d: %.3e" %
(i, j, numpy.sqrt(
(d[i, 0] - d[j, 0])**2) + (d[i, 1] - d[j, 1])**2))
print("Cable_positions:")
for i in range(num_cables):
print("%10.3e, %10.3e" % (d[i, 0], d[i, 1]))
update_mesh(opt_cable_pos)
# dolfin.plot(update_mesh(opt_cable_pos))
TempFiles = [
dolfin.File("output/T_ipopt_part%d.pvd" % i)
for i in range(num_cables + 1)
]
eval_J(cable_positions)
for i in range(num_cables + 1):
TempFiles[i] << T.part(i)
eval_J(opt_cable_pos)
for i in range(num_cables + 1):
TempFiles[i] << T.part(i)
c1 = d[0, :]
c2 = d[1, :]
c3 = d[2, :]
print(
numpy.arccos(
numpy.dot(c1 - c2, c1 - c3) / (numpy.sqrt(
numpy.dot(c1 - c2, c1 - c2) * numpy.dot(c1 - c3, c1 - c3)))) /
(2 * numpy.pi) * 360)
print(
numpy.arccos(
numpy.dot(c1 - c2, c3 - c2) / (numpy.sqrt(
numpy.dot(c1 - c2, c1 - c2) * numpy.dot(c3 - c2, c3 - c2)))) /
(2 * numpy.pi) * 360)
print(
numpy.arccos(
numpy.dot(c3 - c1, c3 - c2) / (numpy.sqrt(
numpy.dot(c3 - c1, c3 - c1) * numpy.dot(c3 - c2, c3 - c2)))) /
(2 * numpy.pi) * 360)
def __solve__(
self,
opt_problem={},
sens_type="FD",
store_sol=True,
disp_opts=False,
store_hst=False,
hot_start=False,
sens_mode="",
sens_step={},
*args,
**kwargs
):
"""
Run Optimizer (Optimize Routine)
**Keyword arguments:**
- opt_problem -> INST: Optimization instance
- sens_type -> STR/FUNC: Gradient type, *Default* = 'FD'
- store_sol -> BOOL: Store solution in Optimization class flag,
*Default* = True
- disp_opts -> BOOL: Flag to display options in solution text, *Default*
= False
- store_hst -> BOOL/STR: Flag/filename to store optimization history,
*Default* = False
- hot_start -> BOOL/STR: Flag/filename to read optimization history,
*Default* = False
- sens_mode -> STR: Flag for parallel gradient calculation, *Default* =
''
- sens_step -> FLOAT: Sensitivity setp size, *Default* = {} [corresponds
to 1e-6 (FD), 1e-20(CS)]
Documentation last updated: Feb. 2, 2011 - Peter W. Jansen
"""
self.pll = False
self.myrank = 0
myrank = self.myrank
tmp_file = False
def_fname = self.options["output_file"][1].split(".")[0]
if isinstance(store_hst, str):
if isinstance(hot_start, str):
if myrank == 0:
if store_hst == hot_start:
hos_file = History(hot_start, "r", self)
log_file = History(store_hst + "_tmp", "w", self, opt_problem.name)
tmp_file = True
else:
hos_file = History(hot_start, "r", self)
log_file = History(store_hst, "w", self, opt_problem.name)
# end
# end
self.sto_hst = True
self.h_start = True
elif hot_start:
if myrank == 0:
hos_file = History(store_hst, "r", self)
log_file = History(store_hst + "_tmp", "w", self, opt_problem.name)
tmp_file = True
# end
self.sto_hst = True
self.h_start = True
else:
if myrank == 0:
log_file = History(store_hst, "w", self, opt_problem.name)
# end
self.sto_hst = True
self.h_start = False
# end
elif store_hst:
if isinstance(hot_start, str):
if hot_start == def_fname:
if myrank == 0:
hos_file = History(hot_start, "r", self)
log_file = History(def_fname + "_tmp", "w", self, opt_problem.name)
tmp_file = True
# end
else:
if myrank == 0:
hos_file = History(hot_start, "r", self)
log_file = History(def_fname, "w", self, opt_problem.name)
# end
# end
self.sto_hst = True
self.h_start = True
elif hot_start:
if myrank == 0:
hos_file = History(def_fname, "r", self)
log_file = History(def_fname + "_tmp", "w", self, opt_problem.name)
tmp_file = True
# end
self.sto_hst = True
self.h_start = True
else:
if myrank == 0:
log_file = History(def_fname, "w", self, opt_problem.name)
# end
self.sto_hst = True
self.h_start = False
# end
else:
self.sto_hst = False
self.h_start = False
# end
gradient = Gradient(opt_problem, sens_type, sens_mode, sens_step, *args, **kwargs)
def eval_f(x, user_data=None):
"""IPOPT - Objective Value Function."""
# Variables Groups Handling
if opt_problem.use_groups:
xg = {}
for group in group_ids.keys():
if group_ids[group][1] - group_ids[group][0] == 1:
xg[group] = x[group_ids[group][0]]
else:
xg[group] = x[group_ids[group][0] : group_ids[group][1]]
# end
# end
xn = xg
else:
xn = x
# end
# Flush Output Files
self.flushFiles()
# Evaluate User Function
fail = 0
# if (myrank == 0):
# if self.h_start:
# [vals,hist_end] = hos_file.read(ident=['obj', 'con', 'fail'])
# if hist_end:
# self.h_start = False
# hos_file.close()
# else:
# [ff,gg,fail] = [vals['obj'][0][0],vals['con'][0],int(vals['fail'][0][0])]
# #end
# #end
# end
# if self.pll:
# self.h_start = Bcast(self.h_start,root=0)
# end
# if self.h_start and self.pll:
# [ff,gg,fail] = Bcast([ff,gg,fail],root=0)
# else:
[ff, gg, fail] = opt_problem.obj_fun(xn, *args, **kwargs)
# end
# Store History
if myrank == 0:
if self.sto_hst:
log_file.write(x, "x")
log_file.write(ff, "obj")
log_file.write(gg, "con")
log_file.write(fail, "fail")
# end
# end
# Objective Assigment
if isinstance(ff, complex):
f = ff.astype(float)
else:
f = ff
# end
# Constraints Assigment
g = numpy.zeros(len(opt_problem._constraints.keys()))
for i in xrange(len(opt_problem._constraints.keys())):
if isinstance(gg[i], complex):
g[i] = gg[i].astype(float)
else:
g[i] = gg[i]
# end
# end
return f
def eval_g(x, user_data=None):
# Variables Groups Handling
if opt_problem.use_groups:
xg = {}
for group in group_ids.keys():
if group_ids[group][1] - group_ids[group][0] == 1:
xg[group] = x[group_ids[group][0]]
else:
xg[group] = x[group_ids[group][0] : group_ids[group][1]]
# end
# end
xn = xg
else:
xn = x
# end
# Flush Output Files
self.flushFiles()
# Evaluate User Function
fail = 0
# if (myrank == 0):
# if self.h_start:
# [vals,hist_end] = hos_file.read(ident=['obj', 'con', 'fail'])
# if hist_end:
# self.h_start = False
# hos_file.close()
# else:
# [ff,gg,fail] = [vals['obj'][0][0],vals['con'][0],int(vals['fail'][0][0])]
# end
# end
# end
# if self.pll:
# self.h_start = Bcast(self.h_start,root=0)
# end
# if self.h_start and self.pll:
# [ff,gg,fail] = Bcast([ff,gg,fail],root=0)
# else:
[ff, gg, fail] = opt_problem.obj_fun(xn, *args, **kwargs)
# end
# Store History
if myrank == 0:
if self.sto_hst:
log_file.write(x, "x")
log_file.write(ff, "obj")
log_file.write(gg, "con")
log_file.write(fail, "fail")
# end
# end
# Objective Assigment
if isinstance(ff, complex):
f = ff.astype(float)
else:
f = ff
# end
# Constraints Assigment
g = numpy.zeros(len(opt_problem._constraints.keys()))
for i in xrange(len(opt_problem._constraints.keys())):
if isinstance(gg[i], complex):
g[i] = gg[i].astype(float)
else:
g[i] = gg[i]
# end
# end
return g
def eval_grad_f(x, user_data=None):
"""IPOPT - Objective/Constraint Gradients Function."""
# if self.h_start:
# if (myrank == 0):
# [vals,hist_end] = hos_file.read(ident=['grad_obj','grad_con'])
# if hist_end:
# self.h_start = False
# hos_file.close()
# else:
# dff = vals['grad_obj'][0].reshape((len(opt_problem._objectives.keys()),len(opt_problem._variables.keys())))
# dgg = vals['grad_con'][0].reshape((len(opt_problem._constraints.keys()),len(opt_problem._variables.keys())))
# #end
# #end
# if self.pll:
# self.h_start = Bcast(self.h_start,root=0)
# #end
# if self.h_start and self.pll:
# [dff,dgg] = Bcast([dff,dgg],root=0)
# #end
# end
# if not self.h_start:
[f, g, fail] = opt_problem.obj_fun(x, *args, **kwargs)
dff, dgg = gradient.getGrad(x, group_ids, [f], g, *args, **kwargs)
# Store History
if self.sto_hst and (myrank == 0):
log_file.write(dff, "grad_obj")
log_file.write(dgg, "grad_con")
# end
# Gradient Assignment
df = numpy.zeros(len(opt_problem._variables.keys()))
for i in xrange(len(opt_problem._variables.keys())):
df[i] = dff[0, i]
# end
return df
def eval_grad_g(x, flag, user_data=None):
# if self.h_start:
# if (myrank == 0):
# [vals,hist_end] = hos_file.read(ident=['grad_obj','grad_con'])
# if hist_end:
# self.h_start = False
# hos_file.close()
# else:
# dff = vals['grad_obj'][0].reshape((len(opt_problem._objectives.keys()),len(opt_problem._variables.keys())))
# dgg = vals['grad_con'][0].reshape((len(opt_problem._constraints.keys()),len(opt_problem._variables.keys())))
# #end
# #end
# if self.pll:
# self.h_start = Bcast(self.h_start,root=0)
# #end
# if self.h_start and self.pll:
# [dff,dgg] = Bcast([dff,dgg],root=0)
# #end
# end
# if not self.h_start:
if flag:
a = numpy.zeros(len(opt_problem._variables.keys()) * len(opt_problem._constraints.keys()), int)
b = numpy.zeros(len(opt_problem._variables.keys()) * len(opt_problem._constraints.keys()), int)
for i in xrange(len(opt_problem._constraints.keys())):
for j in xrange(len(opt_problem._variables.keys())):
a[i * len(opt_problem._variables.keys()) + j] = i
b[i * len(opt_problem._variables.keys()) + j] = j
return (a, b)
else:
[f, g, fail] = opt_problem.obj_fun(x, *args, **kwargs)
dff, dgg = gradient.getGrad(x, group_ids, [f], g, *args, **kwargs)
# Store History
if self.sto_hst and (myrank == 0):
log_file.write(dff, "grad_obj")
log_file.write(dgg, "grad_con")
# end
# Gradient Assignment
a = numpy.zeros([len(opt_problem._variables.keys()) * len(opt_problem._constraints.keys())])
for i in xrange(len(opt_problem._constraints.keys())):
for j in xrange(len(opt_problem._variables.keys())):
a[i * len(opt_problem._variables.keys()) + j] = dgg[i, j]
# end
# end
return a
# Variables Handling
nvar = len(opt_problem._variables.keys())
xl = []
xu = []
xx = []
for key in opt_problem._variables.keys():
if opt_problem._variables[key].type == "c":
xl.append(opt_problem._variables[key].lower)
xu.append(opt_problem._variables[key].upper)
xx.append(opt_problem._variables[key].value)
elif opt_problem._variables[key].type == "i":
raise IOError("IPOPT cannot handle integer design variables")
elif opt_problem._variables[key].type == "d":
raise IOError("IPOPT cannot handle discrete design variables")
# end
# end
xl = numpy.array(xl)
xu = numpy.array(xu)
xx = numpy.array(xx)
# Variables Groups Handling
group_ids = {}
if opt_problem.use_groups:
k = 0
for key in opt_problem._vargroups.keys():
group_len = len(opt_problem._vargroups[key]["ids"])
group_ids[opt_problem._vargroups[key]["name"]] = [k, k + group_len]
k += group_len
# end
# end
# Constraints Handling
ncon = len(opt_problem._constraints.keys())
blc = []
buc = []
if ncon > 0:
for key in opt_problem._constraints.keys():
if opt_problem._constraints[key].type == "e":
blc.append(opt_problem._constraints[key].equal)
buc.append(opt_problem._constraints[key].equal)
elif opt_problem._constraints[key].type == "i":
blc.append(opt_problem._constraints[key].lower)
buc.append(opt_problem._constraints[key].upper)
# end
# end
else:
if (store_sol) and (myrank == 0):
print "Optimization Problem Does Not Have Constraints\n"
print "Unconstrained Optimization Initiated\n"
# end
ncon = 1
blc.append(-inf)
buc.append(inf)
# end
blc = numpy.array(blc)
buc = numpy.array(buc)
# Objective Handling
objfunc = opt_problem.obj_fun
nobj = len(opt_problem._objectives.keys())
ff = []
for key in opt_problem._objectives.keys():
ff.append(opt_problem._objectives[key].value)
# end
ff = numpy.array(ff)
# Create an IPOPT instance problem
nnzj = nvar * ncon
nnzh = nvar * nvar
ipopt = pyipopt.create(nvar, xl, xu, ncon, blc, buc, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_grad_g)
# Setup Options
optionss = self.options.copy()
del optionss["defaults"]
for i in optionss:
if not self.options["defaults"][i][1] == optionss[i][1]:
if self.options[i][0].__name__ == "int":
ipopt.int_option(i, self.options[i][1])
if self.options[i][0].__name__ == "float":
ipopt.num_option(i, self.options[i][1])
if self.options[i][0].__name__ == "str":
ipopt.str_option(i, self.options[i][1])
# Run IPOPT
t0 = time.time()
r = ipopt.solve(xx)
sol_time = time.time() - t0
if myrank == 0:
if self.sto_hst:
log_file.close()
if tmp_file:
hos_file.close()
name = hos_file.filename
os.remove(name + ".cue")
os.remove(name + ".bin")
os.rename(name + "_tmp.cue", name + ".cue")
os.rename(name + "_tmp.bin", name + ".bin")
# end
# end
# end
ipopt.close()
# Store Results
sol_inform = {}
print r
sol_inform["value"] = r[-1] # ifail[0]
sol_inform["text"] = self.getInform(r[-1]) # self.getInform(ifail[0])
if store_sol:
sol_name = "IPOPT Solution to " + opt_problem.name
sol_options = copy.copy(self.options)
if "default" in sol_options:
del sol_options["defaults"]
# end
sol_evals = 0
sol_vars = copy.deepcopy(opt_problem._variables)
i = 0
x = r[0]
for key in sol_vars.keys():
sol_vars[key].value = x[i]
i += 1
# end
sol_objs = copy.deepcopy(opt_problem._objectives)
sol_objs[0].value = r[4]
sol_cons = {}
if ncon > 0:
sol_lambda = r[3]
else:
sol_lambda = {}
# end
opt_problem.addSol(
self.__class__.__name__,
sol_name,
objfunc,
sol_time,
sol_evals,
sol_inform,
sol_vars,
sol_objs,
sol_cons,
sol_options,
display_opts=disp_opts,
Lambda=sol_lambda,
Sensitivities=sens_type,
myrank=myrank,
arguments=args,
**kwargs
)
# end
return ff, xx, sol_inform # ifail[0]
def ipoptopf_solver(om, ppopt):
"""Solves AC optimal power flow using IPOPT.
Inputs are an OPF model object and a PYPOWER options vector.
Outputs are a C{results} dict, C{success} flag and C{raw} output dict.
C{results} is a PYPOWER case dict (ppc) with the usual C{baseMVA}, C{bus}
C{branch}, C{gen}, C{gencost} fields, along with the following additional
fields:
- C{order} see 'help ext2int' for details of this field
- C{x} final value of optimization variables (internal order)
- C{f} final objective function value
- C{mu} shadow prices on ...
- C{var}
- C{l} lower bounds on variables
- C{u} upper bounds on variables
- C{nln}
- C{l} lower bounds on nonlinear constraints
- C{u} upper bounds on nonlinear constraints
- C{lin}
- C{l} lower bounds on linear constraints
- C{u} upper bounds on linear constraints
C{success} is C{True} if solver converged successfully, C{False} otherwise
C{raw} is a raw output dict in form returned by MINOS
- C{xr} final value of optimization variables
- C{pimul} constraint multipliers
- C{info} solver specific termination code
- C{output} solver specific output information
@see: L{opf}, L{pips}
@author: Ray Zimmerman (PSERC Cornell)
@author: Carlos E. Murillo-Sanchez (PSERC Cornell & Universidad
Autonoma de Manizales)
@author: Richard Lincoln
"""
import pyipopt
## unpack data
ppc = om.get_ppc()
baseMVA, bus, gen, branch, gencost = \
ppc['baseMVA'], ppc['bus'], ppc['gen'], ppc['branch'], ppc['gencost']
vv, _, nn, _ = om.get_idx()
## problem dimensions
nb = shape(bus)[0] ## number of buses
ng = shape(gen)[0] ## number of gens
nl = shape(branch)[0] ## number of branches
ny = om.getN('var', 'y') ## number of piece-wise linear costs
## linear constraints
A, l, u = om.linear_constraints()
## bounds on optimization vars
_, xmin, xmax = om.getv()
## build admittance matrices
Ybus, Yf, Yt = makeYbus(baseMVA, bus, branch)
## try to select an interior initial point
ll = xmin.copy()
uu = xmax.copy()
ll[xmin == -Inf] = -2e19 ## replace Inf with numerical proxies
uu[xmax == Inf] = 2e19
x0 = (ll + uu) / 2
Varefs = bus[bus[:, BUS_TYPE] == REF, VA] * (pi / 180)
x0[vv['i1']['Va']:vv['iN']['Va']] = Varefs[
0] ## angles set to first reference angle
if ny > 0:
ipwl = find(gencost[:, MODEL] == PW_LINEAR)
# PQ = r_[gen[:, PMAX], gen[:, QMAX]]
# c = totcost(gencost[ipwl, :], PQ[ipwl])
## largest y-value in CCV data
c = gencost.flatten('F')[sub2ind(shape(gencost), ipwl,
NCOST + 2 * gencost[ipwl, NCOST])]
x0[vv['i1']['y']:vv['iN']['y']] = max(c) + 0.1 * abs(max(c))
# x0[vv['i1']['y']:vv['iN']['y']) = c + 0.1 * abs(c)
## find branches with flow limits
il = find((branch[:, RATE_A] != 0) & (branch[:, RATE_A] < 1e10))
nl2 = len(il) ## number of constrained lines
##----- run opf -----
## build Jacobian and Hessian structure
if A is not None and issparse(A):
nA = A.shape[0] ## number of original linear constraints
else:
nA = 0
nx = len(x0)
f = branch[:, F_BUS] ## list of "from" buses
t = branch[:, T_BUS] ## list of "to" buses
Cf = sparse((ones(nl), (arange(nl), f)),
(nl, nb)) ## connection matrix for line & from buses
Ct = sparse((ones(nl), (arange(nl), t)),
(nl, nb)) ## connection matrix for line & to buses
Cl = Cf + Ct
Cb = Cl.T * Cl + speye(nb, nb)
Cl2 = Cl[il, :]
Cg = sparse((ones(ng), (gen[:, GEN_BUS], arange(ng))), (nb, ng))
nz = nx - 2 * (nb + ng)
nxtra = nx - 2 * nb
if nz > 0:
Js = vstack([
hstack([Cb, Cb, Cg, sparse(
(nb, ng)), sparse((nb, nz))]),
hstack([Cb, Cb, sparse(
(nb, ng)), Cg, sparse((nb, nz))]),
hstack([Cl2, Cl2,
sparse((nl2, 2 * ng)),
sparse((nl2, nz))]),
hstack([Cl2, Cl2,
sparse((nl2, 2 * ng)),
sparse((nl2, nz))])
], 'coo')
else:
Js = vstack([
hstack([Cb, Cb, Cg, sparse((nb, ng))]),
hstack([
Cb,
Cb,
sparse((nb, ng)),
Cg,
]),
hstack([
Cl2,
Cl2,
sparse((nl2, 2 * ng)),
]),
hstack([
Cl2,
Cl2,
sparse((nl2, 2 * ng)),
])
], 'coo')
if A is not None and issparse(A):
Js = vstack([Js, A], 'coo')
f, _, d2f = opf_costfcn(x0, om, True)
Hs = tril(d2f + vstack([
hstack([Cb, Cb, sparse((nb, nxtra))]),
hstack([Cb, Cb, sparse((nb, nxtra))]),
sparse((nxtra, nx))
]),
format='coo')
## set options struct for IPOPT
# options = {}
# options['ipopt'] = ipopt_options([], ppopt)
## extra data to pass to functions
userdata = {
'om': om,
'Ybus': Ybus,
'Yf': Yf[il, :],
'Yt': Yt[il, :],
'ppopt': ppopt,
'il': il,
'A': A,
'nA': nA,
'neqnln': 2 * nb,
'niqnln': 2 * nl2,
'Js': Js,
'Hs': Hs
}
## check Jacobian and Hessian structure
#xr = rand(x0.shape)
#lmbda = rand( 2 * nb + 2 * nl2)
#Js1 = eval_jac_g(x, flag, userdata) #(xr, options.auxdata)
#Hs1 = eval_h(xr, 1, lmbda, userdata)
#i1, j1, s = find(Js)
#i2, j2, s = find(Js1)
#if (len(i1) != len(i2)) | (norm(i1 - i2) != 0) | (norm(j1 - j2) != 0):
# raise ValueError, 'something''s wrong with the Jacobian structure'
#
#i1, j1, s = find(Hs)
#i2, j2, s = find(Hs1)
#if (len(i1) != len(i2)) | (norm(i1 - i2) != 0) | (norm(j1 - j2) != 0):
# raise ValueError, 'something''s wrong with the Hessian structure'
## define variable and constraint bounds
# n is the number of variables
n = x0.shape[0]
# xl is the lower bound of x as bounded constraints
xl = xmin
# xu is the upper bound of x as bounded constraints
xu = xmax
neqnln = 2 * nb
niqnln = 2 * nl2
# number of constraints
m = neqnln + niqnln + nA
# lower bound of constraint
gl = r_[zeros(neqnln), -Inf * ones(niqnln), l]
# upper bound of constraints
gu = r_[zeros(neqnln), zeros(niqnln), u]
# number of nonzeros in Jacobi matrix
nnzj = Js.nnz
# number of non-zeros in Hessian matrix, you can set it to 0
nnzh = Hs.nnz
eval_hessian = True
if eval_hessian:
hessian = lambda x, lagrange, obj_factor, flag, user_data=None: \
eval_h(x, lagrange, obj_factor, flag, userdata)
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g, hessian)
else:
nnzh = 0
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g)
nlp.int_option('print_level', 5)
nlp.num_option('tol', 1.0000e-12)
nlp.int_option('max_iter', 250)
nlp.num_option('dual_inf_tol', 0.10000)
nlp.num_option('constr_viol_tol', 1.0000e-06)
nlp.num_option('compl_inf_tol', 1.0000e-05)
nlp.num_option('acceptable_tol', 1.0000e-08)
nlp.num_option('acceptable_constr_viol_tol', 1.0000e-04)
nlp.num_option('acceptable_compl_inf_tol', 0.0010000)
nlp.str_option('mu_strategy', 'adaptive')
iter = 0
def intermediate_callback(algmod,
iter_count,
obj_value,
inf_pr,
inf_du,
mu,
d_norm,
regularization_size,
alpha_du,
alpha_pr,
ls_trials,
user_data=None):
iter = iter_count
return True
nlp.set_intermediate_callback(intermediate_callback)
## run the optimization
# returns final solution x, upper and lower bound for multiplier, final
# objective function obj and the return status of ipopt
x, zl, zu, obj, status, zg = nlp.solve(x0, m, userdata)
info = {
'x': x,
'zl': zl,
'zu': zu,
'obj': obj,
'status': status,
'lmbda': zg
}
nlp.close()
success = (status == 0) | (status == 1)
output = {'iterations': iter}
f, _ = opf_costfcn(x, om)
## update solution data
Va = x[vv['i1']['Va']:vv['iN']['Va']]
Vm = x[vv['i1']['Vm']:vv['iN']['Vm']]
Pg = x[vv['i1']['Pg']:vv['iN']['Pg']]
Qg = x[vv['i1']['Qg']:vv['iN']['Qg']]
V = Vm * exp(1j * Va)
##----- calculate return values -----
## update voltages & generator outputs
bus[:, VA] = Va * 180 / pi
bus[:, VM] = Vm
gen[:, PG] = Pg * baseMVA
gen[:, QG] = Qg * baseMVA
gen[:, VG] = Vm[gen[:, GEN_BUS].astype(int)]
## compute branch flows
f_br = branch[:, F_BUS].astype(int)
t_br = branch[:, T_BUS].astype(int)
Sf = V[f_br] * conj(Yf * V) ## cplx pwr at "from" bus, p.u.
St = V[t_br] * conj(Yt * V) ## cplx pwr at "to" bus, p.u.
branch[:, PF] = Sf.real * baseMVA
branch[:, QF] = Sf.imag * baseMVA
branch[:, PT] = St.real * baseMVA
branch[:, QT] = St.imag * baseMVA
## line constraint is actually on square of limit
## so we must fix multipliers
muSf = zeros(nl)
muSt = zeros(nl)
if len(il) > 0:
muSf[il] = 2 * info['lmbda'][2 * nb +
arange(nl2)] * branch[il,
RATE_A] / baseMVA
muSt[il] = 2 * info['lmbda'][2 * nb + nl2 +
arange(nl2)] * branch[il,
RATE_A] / baseMVA
## update Lagrange multipliers
bus[:, MU_VMAX] = info['zu'][vv['i1']['Vm']:vv['iN']['Vm']]
bus[:, MU_VMIN] = info['zl'][vv['i1']['Vm']:vv['iN']['Vm']]
gen[:, MU_PMAX] = info['zu'][vv['i1']['Pg']:vv['iN']['Pg']] / baseMVA
gen[:, MU_PMIN] = info['zl'][vv['i1']['Pg']:vv['iN']['Pg']] / baseMVA
gen[:, MU_QMAX] = info['zu'][vv['i1']['Qg']:vv['iN']['Qg']] / baseMVA
gen[:, MU_QMIN] = info['zl'][vv['i1']['Qg']:vv['iN']['Qg']] / baseMVA
bus[:, LAM_P] = info['lmbda'][nn['i1']['Pmis']:nn['iN']['Pmis']] / baseMVA
bus[:, LAM_Q] = info['lmbda'][nn['i1']['Qmis']:nn['iN']['Qmis']] / baseMVA
branch[:, MU_SF] = muSf / baseMVA
branch[:, MU_ST] = muSt / baseMVA
## package up results
nlnN = om.getN('nln')
## extract multipliers for nonlinear constraints
kl = find(info['lmbda'][:2 * nb] < 0)
ku = find(info['lmbda'][:2 * nb] > 0)
nl_mu_l = zeros(nlnN)
nl_mu_u = r_[zeros(2 * nb), muSf, muSt]
nl_mu_l[kl] = -info['lmbda'][kl]
nl_mu_u[ku] = info['lmbda'][ku]
## extract multipliers for linear constraints
lam_lin = info['lmbda'][2 * nb + 2 * nl2 +
arange(nA)] ## lmbda for linear constraints
kl = find(lam_lin < 0) ## lower bound binding
ku = find(lam_lin > 0) ## upper bound binding
mu_l = zeros(nA)
mu_l[kl] = -lam_lin[kl]
mu_u = zeros(nA)
mu_u[ku] = lam_lin[ku]
mu = {
'var': {'l': info['zl'], 'u': info['zu']},
'nln': {'l': nl_mu_l, 'u': nl_mu_u}, \
'lin': {'l': mu_l, 'u': mu_u}
}
results = ppc
results['bus'], results['branch'], results['gen'], \
results['om'], results['x'], results['mu'], results['f'] = \
bus, branch, gen, om, x, mu, f
pimul = r_[results['mu']['nln']['l'] - results['mu']['nln']['u'],
results['mu']['lin']['l'] - results['mu']['lin']['u'],
-ones(ny > 0),
results['mu']['var']['l'] - results['mu']['var']['u']]
raw = {'xr': x, 'pimul': pimul, 'info': info['status'], 'output': output}
return results, success, raw
def __solver__(self, p):
if not pyipoptInstalled:
p.err('you should have pyipopt installed')
# try:
# os.close(1); os.close(2) # may not work for non-Unix OS
# except:
# pass
nvar = p.n
x_L = p.lb
x_U = p.ub
ncon = p.nc + p.nh + p.b.size + p.beq.size
g_L, g_U = zeros(ncon), zeros(ncon)
g_L[:p.nc] = -inf
g_L[p.nc+p.nh:p.nc+p.nh+p.b.size] = -inf
# IPOPT non-linear constraints, both eq and ineq
if p.isFDmodel:
r = []
if p.nc != 0: r.append(p._getPattern(p.user.c))
if p.nh != 0: r.append(p._getPattern(p.user.h))
if p.nb != 0: r.append(p.A)
if p.nbeq != 0: r.append(p.Aeq)
if len(r)>0:
if all([isinstance(elem, ndarray) for elem in r]):
r = vstack(r)
else:
r = Vstack(r)
if isspmatrix(r):
from scipy import __version__
if __version__.startswith('0.7.3') or __version__.startswith('0.7.2') or __version__.startswith('0.7.1') or __version__.startswith('0.7.0'):
p.pWarn('updating scipy to version >= 0.7.4 is very recommended for the problem with the solver IPOPT')
else:
r = array([])
if isspmatrix(r):
I, J, _ = Find(r)
# DON'T remove it!
I, J = array(I, int64), array(J, int64)
elif isinstance(r, ndarray):
if r.size == 0:
I, J= array([], dtype=int64),array([], dtype=int64)
else:
I, J = where(r)
else:
p.disp('unimplemented type:%s' % str(type(r))) # dense matrix?
nnzj = len(I)
else:
I, J = where(ones((ncon, p.n)))
#I, J = None, None
nnzj = ncon * p.n #TODO: reduce it
def eval_g(x):
r = array(())
if p.userProvided.c: r = p.c(x)
if p.userProvided.h: r = hstack((r, p.h(x)))
r = hstack((r, p._get_AX_Less_B_residuals(x), p._get_AeqX_eq_Beq_residuals(x)))
return r
def eval_jac_g(x, flag, userdata = (I, J)):
(I, J) = userdata
if flag and p.isFDmodel:
return (I, J)
r = []
if p.userProvided.c: r.append(p.dc(x))
if p.userProvided.h: r.append(p.dh(x))
if p.nb != 0: r.append(p.A)
if p.nbeq != 0: r.append(p.Aeq)
# TODO: fix it!
if any([isspmatrix(elem) for elem in r]):
r = Vstack([(atleast_2d(elem) if elem.ndim < 2 else elem) for elem in r])
elif len(r)!=0:
r = vstack(r)
if p.isFDmodel:
# TODO: make it more properly
R = (r.tocsr() if isspmatrix(r) else r)[I, J]
if isspmatrix(R):
return R.A
elif isinstance(R, ndarray):
return R
else: p.err('bug in OpenOpt-ipopt connection, inform OpenOpt developers, type(R) = %s' % type(R))
if flag:
return (I, J)
else:
if isspmatrix(r): r = r.A
return r.flatten()
""" This function might be buggy, """ # // comment by Eric
nnzh = 0
def eval_h(lagrange, obj_factor, flag):
return None
# def apply_new(x):
# return True
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, p.f, p.df, eval_g, eval_jac_g)
if self.optFile == 'auto':
lines = ['# generated automatically by OpenOpt\n','print_level 0\n']
lines.append('tol ' + str(p.ftol)+ '\n')
lines.append('constr_viol_tol ' + str(p.contol)+ '\n')
lines.append('max_iter ' + str(min(15000, p.maxIter))+ '\n')
if self.options != '' :
for s in re.split(',|;', self.options):
lines.append(s.strip().replace('=', ' ', 1) + '\n')
if p.nc == 0:
lines.append('jac_d_constant yes\n')
if p.nh == 0:
lines.append('jac_c_constant yes\n')
if p.castFrom.lower() in ('lp', 'qp', 'llsp'):
lines.append('hessian_constant yes\n')
ipopt_opt_file = open('ipopt.opt', 'w')
ipopt_opt_file.writelines(lines)
ipopt_opt_file.close()
try:
x, zl, zu, obj = nlp.solve(p.x0)[:4]
if p.point(p.xk).betterThan(p.point(x)):
obj = p.fk
p.xk = p.xk.copy() # for more safety
else:
p.xk, p.fk = x.copy(), obj
if p.istop == 0: p.istop = 1000
finally:
nlp.close()
def test_example1(self):
nvar = 4
x_L = np.ones((nvar), dtype=np.float_) * 1.0
x_U = np.ones((nvar), dtype=np.float_) * 5.0
ncon = 2
g_L = np.array([25.0, 40.0])
g_U = np.array([2.0*pow(10.0, 19), 40.0])
def eval_f(x, user_data = None):
assert len(x) == 4
return x[0] * x[3] * (x[0] + x[1] + x[2]) + x[2]
def eval_grad_f(x, user_data = None):
assert len(x) == 4
grad_f = np.array([
x[0] * x[3] + x[3] * (x[0] + x[1] + x[2]) ,
x[0] * x[3],
x[0] * x[3] + 1.0,
x[0] * (x[0] + x[1] + x[2])
], np.float_)
return grad_f;
def eval_g(x, user_data= None):
assert len(x) == 4
return np.array([
x[0] * x[1] * x[2] * x[3],
x[0]*x[0] + x[1]*x[1] + x[2]*x[2] + x[3]*x[3]
], np.float_)
nnzj = 8
def eval_jac_g(x, flag, user_data = None):
if flag:
return (np.array([0, 0, 0, 0, 1, 1, 1, 1]),
np.array([0, 1, 2, 3, 0, 1, 2, 3]))
else:
assert len(x) == 4
return np.array([ x[1]*x[2]*x[3],
x[0]*x[2]*x[3],
x[0]*x[1]*x[3],
x[0]*x[1]*x[2],
2.0*x[0],
2.0*x[1],
2.0*x[2],
2.0*x[3] ])
nnzh = 10
def eval_h(x, lagrange, obj_factor, flag, user_data = None):
if flag:
hrow = [0, 1, 1, 2, 2, 2, 3, 3, 3, 3]
hcol = [0, 0, 1, 0, 1, 2, 0, 1, 2, 3]
return (np.array(hcol), np.array(hrow))
else:
values = zeros((10), np.float_)
values[0] = obj_factor * (2*x[3])
values[1] = obj_factor * (x[3])
values[2] = 0
values[3] = obj_factor * (x[3])
values[4] = 0
values[5] = 0
values[6] = obj_factor * (2*x[0] + x[1] + x[2])
values[7] = obj_factor * (x[0])
values[8] = obj_factor * (x[0])
values[9] = 0
values[1] += lagrange[0] * (x[2] * x[3])
values[3] += lagrange[0] * (x[1] * x[3])
values[4] += lagrange[0] * (x[0] * x[3])
values[6] += lagrange[0] * (x[1] * x[2])
values[7] += lagrange[0] * (x[0] * x[2])
values[8] += lagrange[0] * (x[0] * x[1])
values[0] += lagrange[1] * 2
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
x0 = np.array([1.0, 5.0, 5.0, 1.0])
pi0 = np.array([1.0, 1.0])
"""
print x0
print nvar, ncon, nnzj
print x_L, x_U
print g_L, g_U
print eval_f(x0)
print eval_grad_f(x0)
print eval_g(x0)
a = eval_jac_g(x0, True)
print "a = ", a[1], a[0]
print eval_jac_g(x0, False)
print eval_h(x0, pi0, 1.0, False)
print eval_h(x0, pi0, 1.0, True)
"""
r = nlp.solve(x0)
nlp.close()
# print "Solution of the primal variables, x"
# print r["x"]
# print "Solution of the bound multipliers, z_L and z_U"
# print r["mult_xL"], r["mult_xU"]
# print "Solution of the constraint multiplier, lambda"
# print r["mult_g"]
# print "Objective value"
# print "f(x*) =", r["f"]
# print "Constraint value"
# print "g(x*) =", r["g"]
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h)
r = nlp.solve(x0)
values[6] += lagrange[0] * (x[1] * x[2])
values[7] += lagrange[0] * (x[0] * x[2])
values[8] += lagrange[0] * (x[0] * x[1])
values[0] += lagrange[1] * 2
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h)
x0 = array([1.0, 5.0, 5.0, 1.0])
pi0 = array([1.0, 1.0])
"""
print x0
print nvar, ncon, nnzj
print x_L, x_U
print g_L, g_U
print eval_f(x0)
print eval_grad_f(x0)
print eval_g(x0)
a = eval_jac_g(x0, True)
print "a = ", a[1], a[0]
print eval_jac_g(x0, False)
def pyipopt_qp(H, c, A, gl, gu, x0=None, xl=None, xu=None,
hessian=False, **kw_args):
"""IPOPT based quadratic program solver.
Solves the quadratic programming problem:
min 0.5*x'Hx + c'x
x
subject to:
gl <= Ax <= gu
xl <= x <= xu
@param H: quadratic cost coefficient matrix
@param c: linear cost coefficients
@param A: linear constraint matrix
@param gl: lower bound of constraints
@param gu: upper bound of constraints
@param x0: initial starting point
@param xl: lower bound of x as bounded constraints
@param xu: upper bound of x as bounded constraints
@param finite_differences: evaluate Hessian by finite differences
@param kw_args: IPOPT options
"""
userdata = {'H': H, 'c': c, 'A': A}
if x0 is None:
x0 = zeros(A.shape[1])
# n is the number of variables
n = x0.shape[0]
# number of linear constraints (zero nln)
m = A.shape[0]
# number of nonzeros in Jacobi matrix
nnzj = A.nnz
if gl is None:
gl = -Inf * ones(m)
if gu is None:
gu = Inf * ones(m)
if (xl is None) or (len(xl) == 0):
xl = -Inf * ones(n)
if (xu is None) or (len(xu) == 0):
xu = Inf * ones(n)
if not hessian:
nnzh = 0
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh,
qp_f, qp_grad_f, qp_g, qp_jac_g)
else:
Hl = tril(H, format='coo')
# number of non-zeros in Hessian matrix
nnzh = Hl.nnz
eval_h = lambda x, lagrange, obj_factor, flag, \
userdata=None: qp_h(x, lagrange, obj_factor, flag, Hl)
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh,
qp_f, qp_grad_f, qp_g, qp_jac_g, eval_h)
for k, v in kw_args.iteritems():
if isinstance(v, int):
nlp.int_option(k, v)
elif isinstance(v, basestring):
nlp.str_option(k, v)
else:
nlp.num_option(k, v)
# returns x, upper and lower bound for multiplier, final
# objective function obj and the return status of IPOPT
result = nlp.solve(x0, m, userdata)
## final values for the primal variables
x = result[0]
## final values for the lower bound multipliers
zl = result[1]
## final values for the upper bound multipliers
zu = result[2]
## final value of the objective
obj = result[3]
## status of the algorithm
status = result[4]
## final values for the constraint multipliers
zg = result[5]
nlp.close()
return x, zl, zu, obj, status, zg
values[6] += lagrange[0] * (x[1] * x[2])
values[7] += lagrange[0] * (x[0] * x[2])
values[8] += lagrange[0] * (x[0] * x[1])
values[0] += lagrange[1] * 2
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g)
x0 = array([1.0, 5.0, 5.0, 1.0])
pi0 = array([1.0, 1.0])
"""
print x0
print nvar, ncon, nnzj
print x_L, x_U
print g_L, g_U
print eval_f(x0)
print eval_grad_f(x0)
print eval_g(x0)
a = eval_jac_g(x0, True)
print "a = ", a[1], a[0]
print eval_jac_g(x0, False)
print eval_h(x0, pi0, 1.0, False)
def main():
prob = {}
prob["n"] = 20
prob["qdim"] = 2
prob["udim"] = 2
prob["dt"] = 1.0 / (prob["n"] - 1)
qdim = prob['qdim']
start = np.array([0] * qdim + [0] * qdim)
end = np.array([10] * qdim + [0] * qdim)
n = prob["n"]
q_v_arr_lst = [
np.linspace(start[i], end[i], n) for i in range(2 * prob['qdim'])
]
u_arr = np.ones((prob["udim"], n)) * 2
X_2d = np.vstack([q_v_arr_lst, u_arr])
X_init = X_2d.T.flatten()
x_L = np.ones(X_init.size) * -100
x_U = np.ones(X_init.size) * 100
X_sample = np.random.uniform(x_L, x_U)
#set the cost and the gradient of the cost
ctrl_cost = cost.Control_Cost(prob)
eval_f_adolc = aa.Func_Adolc(ctrl_cost, X_sample, scaler=True)
eval_grad_f_adolc = aa.Eval_Grad_F_Adolc(eval_f_adolc.id)
# eval_grad_f_adolc = aa.Func_Adolc(ctrl_cost.eval_grad_f, X_sample)
#set the constriant function for points at specific time
points = [(0, start), (n - 1, end)]
# p_index, p_g_func = [constraint.get_point_constriant(prob, t, g)
# for (t, g) in points]
p_index_g_piar = [
constraint.get_point_constriant(prob, t, g) for (t, g) in points
]
p_index_iter, p_g_func_iter = zip(*p_index_g_piar)
p_index_lst = list(p_index_iter)
p_g_lst = list(p_g_func_iter)
# p_g_adolc_lst = [aa.Func_Adolc(g, X_sample[i])
# for (i, g) in p_index_g_piar]
#set the dynamic constriants
block = dy.Block(prob["qdim"], prob["udim"])
dynamics = block.dynamics
d_index, d_g_func = constraint.get_dynamic_constriants(
prob, dynamics, range(0, n - 1))
# d_g_adolc = aa.Func_Adolc(d_g_func, X_sample[d_index[0]])
#all dynamcis shares the same approximation function
# d_g_adolc_lst = [d_g_adolc for i in d_index]
d_g_lst = [d_g_func for i in d_index]
index_lst = p_index_lst + d_index
# eval_g_adolc_lst = p_g_adolc_lst + d_g_adolc_lst
eval_g_lst = p_g_lst + d_g_lst
# X_sample_lst = [X_sample[i] for i in index_lst]
#
# g_adolc_x_pair = zip(eval_g_adolc_lst, X_sample_lst)
#
# eval_jac_adolc_lst = [aa.Eval_Jac_G_Adolc(g.id, x)
# for (g, x) in g_adolc_x_pair]
eval_g = constraint.Stacked_Constriants(eval_g_lst, index_lst)
eval_g_adolc = aa.Func_Adolc(eval_g, X_sample)
eval_jac_g_adolc = aa.Eval_Jac_G_Adolc(eval_g_adolc.id, X_sample)
# eval_g_adolc = aa.Func_Adolc(eval_g, X_sample)
# eval_jac_g = constraint.Stacked_Constriants_Jacobian(eval_g_lst ,
# eval_jac_lst,
# index_lst)
nvar = X_init.size
ncon = eval_g(X_init).size
eval_lagrangian = constraint.Eval_Lagrangian(ctrl_cost, eval_g)
#x, lagrangian, obj_factor
x_lag_lst = [X_sample, np.ones(ncon), 1]
x_lag_arr = np.hstack(x_lag_lst)
eval_lagrangian_adolc = aa.Func_Adolc(eval_lagrangian, x_lag_lst)
eval_h_adolc = aa.Eval_h_adolc(eval_lagrangian_adolc.id, x_lag_arr)
maks = eval_h_adolc(X_init, np.ones(ncon), 1, True)
H = eval_h_adolc(X_init, np.ones(ncon), 1, False)
g_L = np.zeros(ncon)
g_U = np.zeros(ncon)
nnzj = eval_jac_g_adolc.nnzj
nnzh = eval_h_adolc.nnzh
import ipdb
ipdb.set_trace() # XXX BREAKPOINT
# nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, 0, eval_f_adolc ,
# eval_grad_f_adolc, eval_g_adolc, eval_jac_g_adolc)
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh,
eval_f_adolc, eval_grad_f_adolc, eval_g_adolc,
eval_jac_g_adolc, eval_h_adolc)
output, zl, zu, constraint_multipliers, obj, status = nlp.solve(X_init)
output_2d = output.reshape(n, -1)
return output_2d, prob
print
print "Calling solver with set of vars"
print "nvar = ", nvar
print "x_L = ", x_L
print "x_U = ", x_U
print "g_L = ", g_L
print "g_U = ", g_U
print "ncon = ", ncon
print "nnzj, nnzh = ", nnzj, ", ", nnzh
print "x0 = ", x0
# Call solve()
#pyipopt.set_loglevel(2) # verbose
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, anon_eval_f, anon_eval_grad_f, anon_eval_g, anon_eval_jac_g)
#nlp.int_option("max_iter", 3000)
#nlp.num_option("tol", 1e-8)
#nlp.num_option("acceptable_tol", 1e-2)
#nlp.int_option("acceptable_iter", 0)
nlp.str_option("derivative_test", "first-order")
nlp.str_option("derivative_test_print_all", "no")
#nlp.str_option("print_options_documentation", "yes")
nlp.str_option("print_user_options", "yes")
#nlp.int_option("print_level", 12)
x, zl, zu, obj, status = nlp.solve(x0)
nlp.close()
# Print results
print
print "Solution of the primal variables, x"
ineq_constr_jac(x),
axis=0)
i_s = []
j_s = []
if flag:
for i in range(len(x0)):
for j in range(len(ret)):
i_s.append(i)
j_s.append(j)
return (np.array(i_s), np.array(j_s))
else:
return ret.flatten()
nnzj = len(eval_jac_g(x0, False))
g_L = np.zeros((len(eval_g(x0))))
g_U = np.zeros_like(g_L)
g_U[neq:] = 1e10
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, avg_cost,
avg_cost_grad, eval_g, eval_jac_g)
nlp.num_option('bound_relax_factor', 0.1)
nlp.str_option("mu_strategy", "adaptive")
nlp.str_option("derivative_test", "first-order")
nlp.str_option('warm_start_init_point', 'yes')
nlp.str_option('linear_solver', 'mumps')
print(datetime.datetime.now(), ": Going to call solve")
x, zl, zu, constraint_multipliers, obj, status = nlp.solve(x0)
nlp.close()
print(x)
def get_ipopt_miu(self):
"""Optimized miu
Calculate optimal miu. Called when opt=True is passed to loop().
...
Args:
None
Returns
nd.array: Array of optimal miu, n = params.tmax
"""
try:
import pyipopt
except ImportError:
pyipopt = None
print('OPTIMIZATION ERROR: It appears that you do not have '
'pyipopt installed. Please install it before running '
'optimization.')
x0 = np.concatenate(
(np.linspace(0, 1, 40) ** (1 - np.linspace(0, 1, 40)), np.ones(20))
)
M = 0
nnzj = 0
nnzh = 0
xl = np.zeros(self.params.tmax)
xu = np.ones(self.params.tmax)
xl[0] = .005
xu[0] = .005
xl[-20:] = 1
gl = np.zeros(M)
gu = np.ones(M) * 4.0
def eval_f(_x0):
if (_x0 == self.opt_x).all() and self.opt_obj is not None:
return self.opt_obj
else:
self.opt_x = _x0.copy()
return self.obj_loop(_x0)
def eval_grad_f(_x0):
if (_x0 == self.opt_x).all() and self.opt_grad_f is not None:
return self.opt_grad_f
else:
self.opt_x = _x0.copy()
return self.grad_loop(_x0)
def eval_g(x):
return np.zeros(M)
def eval_jac_g(x, flag):
if flag:
return [], []
else:
return np.empty(M)
pyipopt.set_loglevel(1)
nlp = pyipopt.create(
self.opt_vars, xl, xu, M, gl, gu, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g,
)
nlp.num_option('constr_viol_tol', 8e-7)
nlp.int_option('max_iter', 30)
nlp.num_option('max_cpu_time', 60)
nlp.num_option('tol', self.opt_tol)
# nlp.num_option('acceptable_tol', 1e-4)
# nlp.int_option('acceptable_iter', 4)
nlp.num_option('obj_scaling_factor', -1e+0)
nlp.int_option('print_level', 0)
nlp.str_option('linear_solver', 'ma57')
# nlp.str_option('derivative_test', 'first-order')
x = nlp.solve(x0)[0]
nlp.close()
return x
def test_ipopt_optimization():
"""
This test checks
1. the sparse functionality of pyadolc
2. the execution speed compared to the direct sparse computation
3. run the optimization with the derivatives provided by pyadolc
IPOPT is an interior point algorithm to solve
min f(x)
x in R^n
s.t. g_L <= g(x) <= g_U
x_L <= x <= x_U
this test fails probably because of a bug in pyipopt
"""
try:
import pyipopt
except:
#print '"pyipopt is not installed, skipping test'
#return
raise NotImplementedError("pyipopt is not installed, skipping test")
try:
import scipy.sparse as sparse
except:
#print '"pyipopt is not installed, skipping test'
#return
raise NotImplementedError("scipy is not installed, skipping test")
import time
nvar = 4
x_L = numpy.ones((nvar), dtype=numpy.float_) * 1.0
x_U = numpy.ones((nvar), dtype=numpy.float_) * 5.0
ncon = 2
g_L = numpy.array([25.0, 40.0])
g_U = numpy.array([2.0 * pow(10.0, 19), 40.0])
def eval_f(x, user_data=None):
assert len(x) == 4
return x[0] * x[3] * (x[0] + x[1] + x[2]) + x[2]
def eval_grad_f(x, user_data=None):
assert len(x) == 4
grad_f = numpy.array([
x[0] * x[3] + x[3] * (x[0] + x[1] + x[2]), x[0] * x[3],
x[0] * x[3] + 1.0, x[0] * (x[0] + x[1] + x[2])
])
return grad_f
def eval_g(x, user_data=None):
assert len(x) == 4
return numpy.array([
x[0] * x[1] * x[2] * x[3],
x[0] * x[0] + x[1] * x[1] + x[2] * x[2] + x[3] * x[3]
])
nnzj = 8
def eval_jac_g(x, flag, user_data=None):
if flag:
return (numpy.array([0, 0, 0, 0, 1, 1, 1,
1]), numpy.array([0, 1, 2, 3, 0, 1, 2, 3]))
else:
assert len(x) == 4
return numpy.array([
x[1] * x[2] * x[3], x[0] * x[2] * x[3], x[0] * x[1] * x[3],
x[0] * x[1] * x[2], 2.0 * x[0], 2.0 * x[1], 2.0 * x[2],
2.0 * x[3]
])
nnzh = 10
def eval_h(x, lagrange, obj_factor, flag, user_data=None):
if flag:
hrow = [0, 1, 1, 2, 2, 2, 3, 3, 3, 3]
hcol = [0, 0, 1, 0, 1, 2, 0, 1, 2, 3]
return (numpy.array(hcol, dtype=int), numpy.array(hrow, dtype=int))
else:
values = numpy.zeros((10), numpy.float_)
values[0] = obj_factor * (2 * x[3])
values[1] = obj_factor * (x[3])
values[2] = 0
values[3] = obj_factor * (x[3])
values[4] = 0
values[5] = 0
values[6] = obj_factor * (2 * x[0] + x[1] + x[2])
values[7] = obj_factor * (x[0])
values[8] = obj_factor * (x[0])
values[9] = 0
values[1] += lagrange[0] * (x[2] * x[3])
values[3] += lagrange[0] * (x[1] * x[3])
values[4] += lagrange[0] * (x[0] * x[3])
values[6] += lagrange[0] * (x[1] * x[2])
values[7] += lagrange[0] * (x[0] * x[2])
values[8] += lagrange[0] * (x[0] * x[1])
values[0] += lagrange[1] * 2
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
x0 = numpy.array([1.0, 5.0, 5.0, 1.0])
pi0 = numpy.array([1.0, 1.0])
# check that adolc gives the same answers as derivatives calculated by hand
trace_on(1)
ax = adouble(x0)
independent(ax)
ay = eval_f(ax)
dependent(ay)
trace_off()
trace_on(2)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay)
trace_off()
trace_on(3)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay[0])
trace_off()
trace_on(4)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay[1])
trace_off()
def eval_f_adolc(x, user_data=None):
return function(1, x)[0]
def eval_grad_f_adolc(x, user_data=None):
return gradient(1, x)
def eval_g_adolc(x, user_data=None):
return function(2, x)
def eval_jac_g_adolc(x, flag, user_data=None):
options = numpy.array([1, 1, 0, 0], dtype=int)
result = colpack.sparse_jac_no_repeat(2, x, options)
if flag:
return (numpy.asarray(result[1], dtype=int),
numpy.asarray(result[2], dtype=int))
else:
return result[3]
def eval_h_adolc(x, lagrange, obj_factor, flag, user_data=None):
if flag:
# return sparsity pattern of the hessian
if eval_h_adolc.firstrun == True:
# this can be done more elegantly based directly on
# the sparsity pattern returned by adolc.sparse.hess_pat
options = numpy.array([0, 0], dtype=int)
result_f = colpack.sparse_hess_no_repeat(1, x, options)
result_g0 = colpack.sparse_hess_no_repeat(3, x, options)
result_g1 = colpack.sparse_hess_no_repeat(4, x, options)
Hf = sparse.coo_matrix(
(result_f[3], (result_f[1], result_f[2])), shape=(4, 4))
Hg0 = sparse.coo_matrix(
(result_g0[3], (result_g0[1], result_g0[2])), shape=(4, 4))
Hg1 = sparse.coo_matrix(
(result_g1[3], (result_g1[1], result_g1[2])), shape=(4, 4))
H = Hf + Hg0 + Hg1
H = H.tocoo()
hrow = H.row
hcol = H.col
hpat = (numpy.array(hcol,
dtype=int), numpy.array(hrow, dtype=int))
eval_h_adolc.hpat = hpat
eval_h_adolc.firstrun = False
return eval_h_adolc.hpat
else:
# compute sparse hessian
assert numpy.ndim(x) == 1
assert numpy.size(x) == 4
options = numpy.array([0, 0], dtype=int)
result_f = colpack.sparse_hess_no_repeat(1, x, options)
result_g0 = colpack.sparse_hess_no_repeat(3, x, options)
result_g1 = colpack.sparse_hess_no_repeat(4, x, options)
Hf = sparse.coo_matrix((result_f[3], (result_f[1], result_f[2])),
shape=(4, 4))
Hg0 = sparse.coo_matrix(
(result_g0[3], (result_g0[1], result_g0[2])), shape=(4, 4))
Hg1 = sparse.coo_matrix(
(result_g1[3], (result_g1[1], result_g1[2])), shape=(4, 4))
H = Hf + Hg0 + Hg1
H = H.tocoo()
values = numpy.zeros((10), float)
values[:] = H.data
return values
eval_h_adolc.firstrun = True
# function of f
assert_almost_equal(eval_f(x0), eval_f_adolc(x0))
# gradient of f
assert_array_almost_equal(eval_grad_f(x0), eval_grad_f_adolc(x0))
# function of g
assert_array_almost_equal(eval_g(x0), function(2, x0))
# sparse jacobian of g
assert_array_equal(eval_jac_g_adolc(x0, True)[0], eval_jac_g(x0, True)[0])
assert_array_equal(eval_jac_g_adolc(x0, True)[1], eval_jac_g(x0, True)[1])
assert_array_equal(eval_jac_g_adolc(x0, False), eval_jac_g(x0, False))
# sparse hessian of the lagrangian
lagrange = numpy.ones(2, dtype=float)
obj_factor = 1.
x0 = numpy.random.rand(4)
result = (eval_h(x0, lagrange, obj_factor,
False), eval_h(x0, lagrange, obj_factor, True))
result_adolc = (eval_h_adolc(x0, lagrange, obj_factor, False),
eval_h_adolc(x0, lagrange, obj_factor, True))
H = sparse.coo_matrix(result, shape=(4, 4))
H_adolc = sparse.coo_matrix(result_adolc, shape=(4, 4))
H = H.todense()
H_adolc = H_adolc.todense()
assert_array_almost_equal(H, H_adolc.T)
# test optimization with PYIPOPT
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g, eval_h)
start_time = time.time()
result = nlp.solve(x0)
end_time = time.time()
nlp.close()
pure_python_optimization_time = end_time - start_time
nlp_adolc = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh,
eval_f_adolc, eval_grad_f_adolc, eval_g_adolc,
eval_jac_g_adolc, eval_h_adolc)
start_time = time.time()
result_adolc = nlp_adolc.solve(x0)
end_time = time.time()
nlp_adolc.close()
adolc_optimization_time = end_time - start_time
print('optimization time with derivatives computed by adolc = ',
adolc_optimization_time)
print('optimization time with derivatives computed by hand = ',
pure_python_optimization_time)
assert adolc_optimization_time / pure_python_optimization_time < 10
# this works with the pyipopt version from code.google.com
assert_array_almost_equal(result[0], result_adolc[0])
assert_array_almost_equal(result[1], result_adolc[1])
assert_array_almost_equal(result[2], result_adolc[2])
assert_array_almost_equal(result[3], result_adolc[3])
def iterate(k_init, n_agents, gp_old):
# IPOPT PARAMETERS below "
N = 3 * n_agents # number of vars
M = 3 * n_agents + 1 # number of constraints
NELE_JAC = N * M
NELE_HESS = (N**2 - N) / 2 + N # number of non-zero entries of Hess matrix
# Vector of variables -> solution of non-linear equation system
X = np.empty(N)
LAM = np.empty(M) # multipliers
G = np.empty(M) # (in-)equality constraints
# Vector of lower and upper bounds
G_L = np.empty(M)
G_U = np.empty(M)
X_L = np.empty(N)
X_U = np.empty(N)
Z_L = np.empty(N)
Z_U = np.empty(N)
# get coords of an individual points
grid_pt_box = k_init
X_L[:n_agents] = c_bar
X_U[:n_agents] = c_up
X_L[n_agents:2 * n_agents] = l_bar
X_U[n_agents:2 * n_agents] = l_up
X_L[2 * n_agents:3 * n_agents] = inv_bar
X_U[2 * n_agents:3 * n_agents] = inv_up
# Set bounds for the constraints
G_L[:n_agents] = c_bar
G_U[:n_agents] = c_up
G_L[n_agents:2 * n_agents] = l_bar
G_U[n_agents:2 * n_agents] = l_up
G_L[2 * n_agents:3 * n_agents] = inv_bar
G_U[2 * n_agents:3 * n_agents] = inv_up
G_L[3 * n_agents] = 0.0 # both values set to 0 for equality contraints
G_U[3 * n_agents] = 0.0
# initial guesses for first iteration
cons_init = 0.5 * (X_U[:n_agents] - X_L[:n_agents]) + X_L[:n_agents]
lab_init = 0.5 * (X_U[n_agents:2 * n_agents] -
X_L[n_agents:2 * n_agents]) + X_L[n_agents:2 * n_agents]
inv_init = 0.5 * (X_U[2 * n_agents:3 * n_agents] -
X_L[2 * n_agents:3 * n_agents]) + X_L[2 * n_agents:3 *
n_agents]
X[:n_agents] = cons_init
X[n_agents:2 * n_agents] = lab_init
X[2 * n_agents:3 * n_agents] = inv_init
# Create ev_f, eval_f, eval_grad_f, eval_g, eval_jac_g for given k_init and n_agent
def eval_f(X):
return EV_F_ITER(X, k_init, n_agents, gp_old)
def eval_grad_f(X):
return EV_GRAD_F_ITER(X, k_init, n_agents, gp_old)
def eval_g(X):
return EV_G_ITER(X, k_init, n_agents)
def eval_jac_g(X, flag):
return EV_JAC_G_ITER(X, flag, k_init, n_agents)
# First create a handle for the Ipopt problem
nlp = pyipopt.create(N, X_L, X_U, M, G_L, G_U, NELE_JAC, NELE_HESS, eval_f,
eval_grad_f, eval_g, eval_jac_g)
nlp.num_option("obj_scaling_factor", -1.00)
nlp.num_option("tol", 1e-6)
nlp.num_option("acceptable_tol", 1e-5)
nlp.str_option("derivative_test", "first-order")
nlp.str_option("hessian_approximation", "limited-memory")
nlp.int_option("print_level", 0)
x, z_l, z_u, constraint_multipliers, obj, status = nlp.solve(X)
nlp.close()
# x: Solution of the primal variables
# z_l, z_u: Solution of the bound multipliers
# constraint_multipliers: Solution of the constraint multipliers
# obj: Objective value
# status: Exit Status
# Unpack Consumption, Labor, and Investment
c = x[:n_agents]
l = x[n_agents:2 * n_agents]
inv = x[2 * n_agents:3 * n_agents]
to_print = np.hstack((obj, x))
# === debug
#f=open("results.txt", 'a')
#np.savetxt(f, np.transpose(to_print) #, fmt=len(x)*'%10.10f ')
#for num in to_print:
# f.write(str(num)+"\t")
#f.write("\n")
#f.close()
return obj, c, l, inv
def ipoptopf_solver(om, ppopt):
"""Solves AC optimal power flow using IPOPT.
Inputs are an OPF model object and a PYPOWER options vector.
Outputs are a C{results} dict, C{success} flag and C{raw} output dict.
C{results} is a PYPOWER case dict (ppc) with the usual C{baseMVA}, C{bus}
C{branch}, C{gen}, C{gencost} fields, along with the following additional
fields:
- C{order} see 'help ext2int' for details of this field
- C{x} final value of optimization variables (internal order)
- C{f} final objective function value
- C{mu} shadow prices on ...
- C{var}
- C{l} lower bounds on variables
- C{u} upper bounds on variables
- C{nln}
- C{l} lower bounds on nonlinear constraints
- C{u} upper bounds on nonlinear constraints
- C{lin}
- C{l} lower bounds on linear constraints
- C{u} upper bounds on linear constraints
C{success} is C{True} if solver converged successfully, C{False} otherwise
C{raw} is a raw output dict in form returned by MINOS
- C{xr} final value of optimization variables
- C{pimul} constraint multipliers
- C{info} solver specific termination code
- C{output} solver specific output information
@see: L{opf}, L{pips}
@author: Ray Zimmerman (PSERC Cornell)
@author: Carlos E. Murillo-Sanchez (PSERC Cornell & Universidad
Autonoma de Manizales)
@author: Richard Lincoln
"""
import pyipopt
## unpack data
ppc = om.get_ppc()
baseMVA, bus, gen, branch, gencost = \
ppc['baseMVA'], ppc['bus'], ppc['gen'], ppc['branch'], ppc['gencost']
vv, _, nn, _ = om.get_idx()
## problem dimensions
nb = shape(bus)[0] ## number of buses
ng = shape(gen)[0] ## number of gens
nl = shape(branch)[0] ## number of branches
ny = om.getN('var', 'y') ## number of piece-wise linear costs
## linear constraints
A, l, u = om.linear_constraints()
## bounds on optimization vars
_, xmin, xmax = om.getv()
## build admittance matrices
Ybus, Yf, Yt = makeYbus(baseMVA, bus, branch)
## try to select an interior initial point
ll = xmin.copy(); uu = xmax.copy()
ll[xmin == -Inf] = -2e19 ## replace Inf with numerical proxies
uu[xmax == Inf] = 2e19
x0 = (ll + uu) / 2
Varefs = bus[bus[:, BUS_TYPE] == REF, VA] * (pi / 180)
x0[vv['i1']['Va']:vv['iN']['Va']] = Varefs[0] ## angles set to first reference angle
if ny > 0:
ipwl = find(gencost[:, MODEL] == PW_LINEAR)
# PQ = r_[gen[:, PMAX], gen[:, QMAX]]
# c = totcost(gencost[ipwl, :], PQ[ipwl])
## largest y-value in CCV data
c = gencost.flatten('F')[sub2ind(shape(gencost), ipwl, NCOST + 2 * gencost[ipwl, NCOST])]
x0[vv['i1']['y']:vv['iN']['y']] = max(c) + 0.1 * abs(max(c))
# x0[vv['i1']['y']:vv['iN']['y']) = c + 0.1 * abs(c)
## find branches with flow limits
il = find((branch[:, RATE_A] != 0) & (branch[:, RATE_A] < 1e10))
nl2 = len(il) ## number of constrained lines
##----- run opf -----
## build Jacobian and Hessian structure
if A is not None and issparse(A):
nA = A.shape[0] ## number of original linear constraints
else:
nA = 0
nx = len(x0)
f = branch[:, F_BUS] ## list of "from" buses
t = branch[:, T_BUS] ## list of "to" buses
Cf = sparse((ones(nl), (arange(nl), f)), (nl, nb)) ## connection matrix for line & from buses
Ct = sparse((ones(nl), (arange(nl), t)), (nl, nb)) ## connection matrix for line & to buses
Cl = Cf + Ct
Cb = Cl.T * Cl + speye(nb, nb)
Cl2 = Cl[il, :]
Cg = sparse((ones(ng), (gen[:, GEN_BUS], arange(ng))), (nb, ng))
nz = nx - 2 * (nb + ng)
nxtra = nx - 2 * nb
if nz > 0:
Js = vstack([
hstack([Cb, Cb, Cg, sparse((nb, ng)), sparse((nb, nz))]),
hstack([Cb, Cb, sparse((nb, ng)), Cg, sparse((nb, nz))]),
hstack([Cl2, Cl2, sparse((nl2, 2 * ng)), sparse((nl2, nz))]),
hstack([Cl2, Cl2, sparse((nl2, 2 * ng)), sparse((nl2, nz))])
], 'coo')
else:
Js = vstack([
hstack([Cb, Cb, Cg, sparse((nb, ng))]),
hstack([Cb, Cb, sparse((nb, ng)), Cg, ]),
hstack([Cl2, Cl2, sparse((nl2, 2 * ng)), ]),
hstack([Cl2, Cl2, sparse((nl2, 2 * ng)), ])
], 'coo')
if A is not None and issparse(A):
Js = vstack([Js, A], 'coo')
f, _, d2f = opf_costfcn(x0, om, True)
Hs = tril(d2f + vstack([
hstack([Cb, Cb, sparse((nb, nxtra))]),
hstack([Cb, Cb, sparse((nb, nxtra))]),
sparse((nxtra, nx))
]), format='coo')
## set options struct for IPOPT
# options = {}
# options['ipopt'] = ipopt_options([], ppopt)
## extra data to pass to functions
userdata = {
'om': om,
'Ybus': Ybus,
'Yf': Yf[il, :],
'Yt': Yt[il, :],
'ppopt': ppopt,
'il': il,
'A': A,
'nA': nA,
'neqnln': 2 * nb,
'niqnln': 2 * nl2,
'Js': Js,
'Hs': Hs
}
## check Jacobian and Hessian structure
#xr = rand(x0.shape)
#lmbda = rand( 2 * nb + 2 * nl2)
#Js1 = eval_jac_g(x, flag, userdata) #(xr, options.auxdata)
#Hs1 = eval_h(xr, 1, lmbda, userdata)
#i1, j1, s = find(Js)
#i2, j2, s = find(Js1)
#if (len(i1) != len(i2)) | (norm(i1 - i2) != 0) | (norm(j1 - j2) != 0):
# raise ValueError, 'something''s wrong with the Jacobian structure'
#
#i1, j1, s = find(Hs)
#i2, j2, s = find(Hs1)
#if (len(i1) != len(i2)) | (norm(i1 - i2) != 0) | (norm(j1 - j2) != 0):
# raise ValueError, 'something''s wrong with the Hessian structure'
## define variable and constraint bounds
# n is the number of variables
n = x0.shape[0]
# xl is the lower bound of x as bounded constraints
xl = xmin
# xu is the upper bound of x as bounded constraints
xu = xmax
neqnln = 2 * nb
niqnln = 2 * nl2
# number of constraints
m = neqnln + niqnln + nA
# lower bound of constraint
gl = r_[zeros(neqnln), -Inf * ones(niqnln), l]
# upper bound of constraints
gu = r_[zeros(neqnln), zeros(niqnln), u]
# number of nonzeros in Jacobi matrix
nnzj = Js.nnz
# number of non-zeros in Hessian matrix, you can set it to 0
nnzh = Hs.nnz
eval_hessian = True
if eval_hessian:
hessian = lambda x, lagrange, obj_factor, flag, user_data=None: \
eval_h(x, lagrange, obj_factor, flag, userdata)
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh,
eval_f, eval_grad_f, eval_g, eval_jac_g, hessian)
else:
nnzh = 0
nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh,
eval_f, eval_grad_f, eval_g, eval_jac_g)
nlp.int_option('print_level', 5)
nlp.num_option('tol', 1.0000e-12)
nlp.int_option('max_iter', 250)
nlp.num_option('dual_inf_tol', 0.10000)
nlp.num_option('constr_viol_tol', 1.0000e-06)
nlp.num_option('compl_inf_tol', 1.0000e-05)
nlp.num_option('acceptable_tol', 1.0000e-08)
nlp.num_option('acceptable_constr_viol_tol', 1.0000e-04)
nlp.num_option('acceptable_compl_inf_tol', 0.0010000)
nlp.str_option('mu_strategy', 'adaptive')
iter = 0
def intermediate_callback(algmod, iter_count, obj_value, inf_pr, inf_du,
mu, d_norm, regularization_size, alpha_du, alpha_pr, ls_trials,
user_data=None):
iter = iter_count
return True
nlp.set_intermediate_callback(intermediate_callback)
## run the optimization
# returns final solution x, upper and lower bound for multiplier, final
# objective function obj and the return status of ipopt
x, zl, zu, obj, status, zg = nlp.solve(x0, m, userdata)
info = {'x': x, 'zl': zl, 'zu': zu, 'obj': obj, 'status': status, 'lmbda': zg}
nlp.close()
success = (status == 0) | (status == 1)
output = {'iterations': iter}
f, _ = opf_costfcn(x, om)
## update solution data
Va = x[vv['i1']['Va']:vv['iN']['Va']]
Vm = x[vv['i1']['Vm']:vv['iN']['Vm']]
Pg = x[vv['i1']['Pg']:vv['iN']['Pg']]
Qg = x[vv['i1']['Qg']:vv['iN']['Qg']]
V = Vm * exp(1j * Va)
##----- calculate return values -----
## update voltages & generator outputs
bus[:, VA] = Va * 180 / pi
bus[:, VM] = Vm
gen[:, PG] = Pg * baseMVA
gen[:, QG] = Qg * baseMVA
gen[:, VG] = Vm[gen[:, GEN_BUS].astype(int)]
## compute branch flows
f_br = branch[:, F_BUS].astype(int)
t_br = branch[:, T_BUS].astype(int)
Sf = V[f_br] * conj(Yf * V) ## cplx pwr at "from" bus, p.u.
St = V[t_br] * conj(Yt * V) ## cplx pwr at "to" bus, p.u.
branch[:, PF] = Sf.real * baseMVA
branch[:, QF] = Sf.imag * baseMVA
branch[:, PT] = St.real * baseMVA
branch[:, QT] = St.imag * baseMVA
## line constraint is actually on square of limit
## so we must fix multipliers
muSf = zeros(nl)
muSt = zeros(nl)
if len(il) > 0:
muSf[il] = 2 * info['lmbda'][2 * nb + arange(nl2)] * branch[il, RATE_A] / baseMVA
muSt[il] = 2 * info['lmbda'][2 * nb + nl2 + arange(nl2)] * branch[il, RATE_A] / baseMVA
## update Lagrange multipliers
bus[:, MU_VMAX] = info['zu'][vv['i1']['Vm']:vv['iN']['Vm']]
bus[:, MU_VMIN] = info['zl'][vv['i1']['Vm']:vv['iN']['Vm']]
gen[:, MU_PMAX] = info['zu'][vv['i1']['Pg']:vv['iN']['Pg']] / baseMVA
gen[:, MU_PMIN] = info['zl'][vv['i1']['Pg']:vv['iN']['Pg']] / baseMVA
gen[:, MU_QMAX] = info['zu'][vv['i1']['Qg']:vv['iN']['Qg']] / baseMVA
gen[:, MU_QMIN] = info['zl'][vv['i1']['Qg']:vv['iN']['Qg']] / baseMVA
bus[:, LAM_P] = info['lmbda'][nn['i1']['Pmis']:nn['iN']['Pmis']] / baseMVA
bus[:, LAM_Q] = info['lmbda'][nn['i1']['Qmis']:nn['iN']['Qmis']] / baseMVA
branch[:, MU_SF] = muSf / baseMVA
branch[:, MU_ST] = muSt / baseMVA
## package up results
nlnN = om.getN('nln')
## extract multipliers for nonlinear constraints
kl = find(info['lmbda'][:2 * nb] < 0)
ku = find(info['lmbda'][:2 * nb] > 0)
nl_mu_l = zeros(nlnN)
nl_mu_u = r_[zeros(2 * nb), muSf, muSt]
nl_mu_l[kl] = -info['lmbda'][kl]
nl_mu_u[ku] = info['lmbda'][ku]
## extract multipliers for linear constraints
lam_lin = info['lmbda'][2 * nb + 2 * nl2 + arange(nA)] ## lmbda for linear constraints
kl = find(lam_lin < 0) ## lower bound binding
ku = find(lam_lin > 0) ## upper bound binding
mu_l = zeros(nA)
mu_l[kl] = -lam_lin[kl]
mu_u = zeros(nA)
mu_u[ku] = lam_lin[ku]
mu = {
'var': {'l': info['zl'], 'u': info['zu']},
'nln': {'l': nl_mu_l, 'u': nl_mu_u}, \
'lin': {'l': mu_l, 'u': mu_u}
}
results = ppc
results['bus'], results['branch'], results['gen'], \
results['om'], results['x'], results['mu'], results['f'] = \
bus, branch, gen, om, x, mu, f
pimul = r_[
results['mu']['nln']['l'] - results['mu']['nln']['u'],
results['mu']['lin']['l'] - results['mu']['lin']['u'],
-ones(ny > 0),
results['mu']['var']['l'] - results['mu']['var']['u']
]
raw = {'xr': x, 'pimul': pimul, 'info': info['status'], 'output': output}
return results, success, raw
def main(traj_filename):
prob = {}
prob["n"] = 10
prob["qdim"] = 2
prob["udim"] = 1
# prob["dt"] = 0.1/( prob["n"]-1)
prob["dt"] = 0.1
p_L = [-2, -1*np.pi, -2, -np.pi, -3]
p_U = [2, 1*np.pi, 2, np.pi, 3]
x_L = np.tile(p_L, prob["n"])
x_U = np.tile(p_U, prob["n"])
qdim = prob['qdim']
start = np.array([0]*qdim+[0]*qdim)
start[1] = -np.pi
start[1] = 0
end = np.array([0]*qdim+[0]*qdim)
n = prob["n"]
q_v_arr_lst = [np.linspace(start[i], end[i], n) for i in range(2*prob['qdim'])]
u_arr = np.ones((prob["udim"], n))*0.001
X_2d = np.vstack([q_v_arr_lst, u_arr])
X_init = X_2d.T.flatten()
# set the control cost
X_sample = np.random.uniform(x_L, x_U)
#set the cost and the gradient of the cost
ctrl_cost = cost.Control_Cost(prob)
eval_f_adolc = aa.Func_Adolc(ctrl_cost, X_sample, scaler=True)
eval_grad_f_adolc = aa.Eval_Grad_F_Adolc(eval_f_adolc.id)
# eval_grad_f_adolc = aa.Func_Adolc(ctrl_cost.eval_grad_f, X_sample)
#set the constriant function for points at specific time
g1 = np.array([-np.pi, 0])
g2 = np.array([0, 0])
points = [(0, g1), (n-1, g2)]
# points = [(n-1, end)]
# points = [(0, start)]
# p_index, p_g_func = [constraint.get_point_constriant(prob, t, g)
# for (t, g) in points]
# q and v of the pole
dims = np.array([1,1+prob["qdim"]])
p_index_g_piar = [constraint.get_point_constriant(prob, t, g, dims)
for (t, g) in points]
p_index_iter, p_g_func_iter = zip(*p_index_g_piar)
p_index_lst = list(p_index_iter)
p_g_lst = list(p_g_func_iter)
# p_g_adolc_lst = [aa.Func_Adolc(g, X_sample[i])
# for (i, g) in p_index_g_piar]
D_factory= constraint.Dynamics_constriant
model_path = "/home/tao/src/gym/gym/envs/mujoco/assets/inverted_pendulum.xml"
model = dy.make_model(model_path)
sim = mujoco_py.MjSim(model)
qdim = model.nq
udim = model.nu
cart = dy.Mujoco_Dynamics(model, sim, qdim, udim)
dynamics = cart.dynamics
d_index, d_g_func = constraint.get_dynamic_constriants(prob,
dynamics,
range(0, n-1))
# d_g_adolc = aa.Func_Adolc(d_g_func, X_sample[d_index[0]])
#all dynamcis shares the same approximation function
# d_g_adolc_lst = [d_g_adolc for i in d_index]
d_g_lst = [d_g_func for i in d_index]
index_lst = p_index_lst + d_index
eval_g_lst = p_g_lst + d_g_lst
# X_sample_lst = [X_sample[i] for i in index_lst]
#
# g_adolc_x_pair = zip(eval_g_adolc_lst, X_sample_lst)
#
# eval_jac_adolc_lst = [aa.Eval_Jac_G_Adolc(g.id, x)
# for (g, x) in g_adolc_x_pair]
eval_g = constraint.Stacked_Constriants(eval_g_lst, index_lst)
eval_g_adolc = aa.Func_Adolc(eval_g, X_sample)
eval_jac_g_adolc = aa.Eval_Jac_G_Adolc(eval_g_adolc.id, X_sample)
# eval_g_adolc = aa.Func_Adolc(eval_g, X_sample)
# eval_jac_g = constraint.Stacked_Constriants_Jacobian(eval_g_lst ,
# eval_jac_lst,
# index_lst)
nvar = X_init.size
ncon = eval_g(X_init).size
eval_lagrangian = constraint.Eval_Lagrangian(ctrl_cost, eval_g)
#x, lagrangian, obj_factor
x_lag_lst = [X_sample, np.ones(ncon), 1]
x_lag_arr = np.hstack(x_lag_lst)
eval_lagrangian_adolc = aa.Func_Adolc(eval_lagrangian, x_lag_lst)
eval_h_adolc = aa.Eval_h_adolc(eval_lagrangian_adolc.id, x_lag_arr)
maks = eval_h_adolc(X_init, np.ones(ncon), 1, True)
H = eval_h_adolc(X_init, np.ones(ncon), 1, False)
g_L = np.zeros(ncon)
g_U = np.zeros(ncon)
nnzj = eval_jac_g_adolc.nnzj
nnzh = eval_h_adolc.nnzh
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, 0, eval_f_adolc ,
eval_grad_f_adolc, eval_g_adolc, eval_jac_g_adolc)
# nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f_adolc ,
# eval_grad_f_adolc, eval_g_adolc, eval_jac_g_adolc, eval_h_adolc)
output, zl, zu, constraint_multipliers, obj, status = nlp.solve(X_init)
output_2d = output.reshape(n, -1)
output.dump(traj_filename)
return output_2d, prob
def cons_opt(obj, cons, Vars, x0):
"""nlinprog
Parameters
----------
obj :
cons :
Returns
-------
Notes
------
"""
nvars = len(Vars)
x_L = np.array((pyipopt.NLP_LOWER_BOUND_INF,)*nvars)
x_U = np.array((pyipopt.NLP_UPPER_BOUND_INF,)*nvars)
#x_L = -20.*np.ones(nvars)
#x_U = 20.*np.ones(nvars)
g_L, g_U, g = [], [], []
for gc in group_cons_by_ub_lb(cons):
g_L.append(gc.lb)
g_U.append(gc.ub)
g.append(gc.expr)
ncon = len(g)
g_L, g_U = np.array(g_L), np.array(g_U)
eval_g = ft.partial(list2array_wrap, sym.lambdify(Vars, g))
js = jac(Vars, g)
jrow, jcol, jdata = np.asarray(js.row, dtype=int), np.asarray(js.col, dtype=int), js.data
eval_jac_g = ft.partial(eval_jac_cons, (jrow, jcol, sym.lambdify(Vars, jdata.tolist())))
eval_f = ft.partial(eval_expr, sym.lambdify(Vars, obj))
eval_grad_f = ft.partial(eval_grad_obj, sym.lambdify(Vars, grad(Vars, obj)))
#eval_hessian_f = ft.partial(eval_expr, sym.lambdify(Vars, sym.hessian(obj, Vars)))
nnzj = js.nnz
nnzh = 0
if debug:
for gi, lb, ub in zip(g, g_L, g_U):
print('{} \in [{}, {}]'.format(gi, lb, ub))
nlp = pyipopt.create(nvars, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
# Verbosity level \in [0, 12]
nlp.int_option('print_level', print_level)
nlp.num_option('constr_viol_tol', constr_viol_tol)
res_x, zl, zu, constraint_multipliers, res_obj, status = nlp.solve(x0)
nlp.close()
if debug:
def print_variable(variable_name, value):
for i in xrange(len(value)):
print(variable_name + "["+str(i)+"] =", value[i])
print()
print("Solution of the primal variables, x")
print_variable("x", res_x)
print()
print("Solution of the bound multipliers, z_L and z_U")
print_variable("z_L", zl)
print_variable("z_U", zu)
print()
print("Solution of the constraint multipliers, lambda")
print_variable("lambda", constraint_multipliers)
print()
print("Objective value")
print("f(x*) =", res_obj)
# Return codes in IpReturnCodes_inc.h
print('status:', status)
return spec.OPTRES(res_obj, res_x, 'OK', status in (0, 1))
## IPOPT SOLUTION
start = time.time()
nvar = data.numX
xl = np.zeros(data.numX)
xu = 2e19*np.ones(data.numX)
m = 0
gl = np.zeros(1)
gu = 2e19*np.ones(1)
g_L = np.array([], dtype=float)
g_U = np.array([], dtype=float)
nnzj = 0
nnzh = int(data.numX * (data.numX + 1) / 2)
nlp = pyipopt.create(nvar, xl, xu, m, g_L, g_U, nnzj, nnzh, evaluateFunction, evaluateGradient, eval_g, eval_jac_g, evaluateHessian)
nlp.num_option('tol', 1e-5)
nlp.int_option("print_level", 5)
nlp.str_option('hessian_approximation', 'limited-memory')
nlp.str_option('mu_strategy', 'adaptive')
nlp.str_option('mu_oracle', 'probing')
nlp.str_option('linear_solver', 'ma97')
nlp.num_option('acceptable_tol', 1e-2)
nlp.int_option("acceptable_iter", 5)
nlp.num_option('acceptable_obj_change_tol', 5e-1)
mylines = [float(myline) for myline in [line.rstrip('\n') for line in open('/home/wilmer/Dropbox/IpOptSolver/currentIntensities.txt')]]
data.currentIntensities = np.array(mylines)
x, zl, zu, constraint_multipliers, obj, status = nlp.solve(data.currentIntensities)
print('solved in ' + str(time.time()-start) + ' seconds')
# PYTHON scipy.optimize solution
def min_ipopt(self, XP0, xtrace=None):
"""
Minimize f starting from XP0 using IPOPT.
Returns the minimizing state, the minimum function value, and the
termination information.
"""
if self.taped == False:
self.tape_A(xtrace)
#Is this time consuming - creates a new instance of the class at each call
#Need to move outside this function
#Or figure out how to do this WITH a function instead of a class
eval_h_adolc = self.Eval_h_adolc(XP0)
nnzh = eval_h_adolc.nnz
#Includes NPest
nvar = len(XP0)
#Use bounds - Includes parameters. Does this work for time dependent parameters?
x_L = np.asarray(self.bounds)[:, 0]
x_U = np.asarray(self.bounds)[:, 1]
ncon = 0
g_L = np.array([])
g_U = np.array([])
nnzj = 0
nlp_adolc = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh,
A_taped, gradA_taped, eval_g, eval_jac_g,
eval_h_adolc)
#Setting default settings to match minAone with adjustments
nlp_adolc.num_option('tol', 1e-6)
nlp_adolc.str_option('mu_strategy', 'adaptive')
nlp_adolc.str_option('adaptive_mu_globalization',
'never-monotone-mode')
nlp_adolc.int_option('max_iter', 1000)
nlp_adolc.str_option('linear_solver', 'ma97')
nlp_adolc.num_option('bound_relax_factor', 0)
#IPOPT distinguishes between num, int, and str options
#Only supports 3 options atm, could for loop through values with if statements...
if self.opt_args is not None:
if 'max_iter' in boundz:
nlp_adolc.int_option('max_iter', boundz.get('max_iter'))
if 'tol' in boundz:
nlp_adolc.num_option('tol', boundz.get('tol'))
if 'linear_solver' in boundz:
nlp_adolc.str_option('linear_solver',
boundz.get('linear_solver'))
# start the optimization
print("Beginning optimization...")
tstart = time.time()
XPmin, _, _, _, Amin, status = nlp_adolc.solve(XP0)
#deleting objects
nlp_adolc.close()
del eval_h_adolc
print("Optimization complete!")
print("Time = {0} s".format(time.time() - tstart))
#print("Exit flag = {0}".format(status))
#print("Exit message: {0}".format(res.message))
#print("Iterations = {0}".format(res.nit))
print("Obj. function value = {0}\n".format(Amin))
return XPmin, Amin, status
def get_ipopt_options(rd, lb, ub, tol, max_iter, **kwargs):
"""Get options for IPOPT module (interior point algorithm)
See `<https://projects.coin-or.org/Ipopt>`
rd : :py:class`dolfin_adjoint.ReducedFunctional`
The reduced functional
lb : list
Lower bound on the control
ub : list
Upper bound on the control
tol : float
Tolerance
max_iter : int
Maximum number of iterations
*Returns*
nlp : ipopt instance
A nonlinear ipopt problem
"""
ncontrols = len(ub)
nconstraints = 0
empty = np.array([], dtype=float)
clb = empty
cub = empty
constraints_nnz = nconstraints * ncontrols
# The constraint function, should do nothing
def fun_g(x, user_data=None):
return empty
# The constraint Jacobian
def jac_g(x, flag, user_data=None):
if flag:
rows = np.array([], dtype=int)
cols = np.array([], dtype=int)
return (rows, cols)
else:
return empty
J = rd.__call__
dJ = rd.derivative
nlp = pyipopt.create(
ncontrols, # length of control vector
lb, # lower bounds on control vector
ub, # upper bounds on control vector
0, # number of constraints
clb, # lower bounds on constraints,
cub, # upper bounds on constraints,
0, # number of nonzeros in the constraint Jacobian
0, # number of nonzeros in the Hessian
J, # to evaluate the functional
dJ, # to evaluate the gradient
fun_g, # to evaluate the constraints
jac_g,
) # to evaluate the constraint Jacobian
pyipopt.set_loglevel(1)
return nlp
return j.tolist() dummyh, hrow, hcol = _amplpy.eval_H( nlp.x0, nlp.pi0, 1, 1.0 ) print hrow, hcol dummyh, hrow, hcol = nlp.hess( nlp.x0, nlp.pi0, 1.0 ) def eval_h(x, lagrange, obj_factor, flag): if flag: return (hrow.tolist(), hcol.tolist()) else: temph, dummyr, dummyc = nlp.hess(array(x), array(lagrange), obj_factor) assert len(temph) == nnzh return temph.tolist() print n, m print xl, xu print nnzj print gl, gu print x0 print eval_f(x0) print eval_grad_f(x0) print eval_g(x0) print eval_jac_g(x0, False) print eval_jac_g(x0, True) nlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, 0, eval_f, eval_grad_f, eval_g, eval_jac_g, ) print nlp #nlp.solve(x0)
def MCdo(b, i):
user = os.path.expanduser("~")
tempOut = robj.r("rm(list=ls())")
tempOut = robj.r("i <- %i"%(i+1))
tempOut = robj.r("b <- %i"%(i*(1000)+(b+1) ))
tempOut = robj.r("source('CMLE_unstable_support.r')")
regrMat = np.array((robj.r("do.call(cbind, regr)")))
regr = OrderedDict([('SA',np.ones((regrMat.shape[0],0))),('VA',regrMat[:,0:1]),('CB',np.ones((regrMat.shape[0],0))),('barWA',regrMat[:,1:2]),('barWB',regrMat[:,2:4]),('bara',regrMat[:,4:5]),('VB',regrMat[:,5:6])])
Y = np.array((robj.r("Y")))
x0 = np.array((robj.r("unname(x0)")))
xL = np.array((robj.r("unname(xL)")))
tPL = np.array((robj.r("unname(out.2step$time )")))
def fll(x):
return LL_jo(x, Y, regr)
def heq(x):
return const_cmle(x, Y, regr)
np.random.seed(b)
while np.isinf(fll(adouble(x0)).val):
x0 = np.random.uniform(size=len(x0))
ccd = os.getcwd()
if not os.path.exists(user + '/Documents/adolc%i_%i'%(i, b)):
os.makedirs(user + '/Documents/adolc%i_%i'%(i, b))
os.chdir(user + '/Documents/adolc%i_%i'%(i, b))
adolc.trace_on(1)
ax = adolc.adouble(np.zeros(len(x0)))
adolc.independent(ax)
ay = fll(ax)
adolc.dependent(ay)
adolc.trace_off()
# trace constraint function
adolc.trace_on(2)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = heq(ax)
adolc.dependent(ay)
adolc.trace_off()
npar = len(x0)
def adFun(x):
return adolc.function(1, x)
def grFun(x):
return adolc.gradient(1, x)
def const_adolc(x):
return adolc.function(2,x)
def jac_adolc(x):
return adolc.jacobian(2,x)
def lagrangian(x, lagrange, obj_factor):
return obj_factor*fll(x) + np.dot(lagrange, heq(x))
#Jacobian
#### initalize it
class jac_c_adolc:
def __init__(self, x):
options = None
result = adolc.colpack.sparse_jac_no_repeat(2,x,options)
self.nnz = result[0]
self.rind = np.asarray(result[1],dtype=int)
self.cind = np.asarray(result[2],dtype=int)
self.values = np.asarray(result[3],dtype=float)
def __call__(self, x, flag, user_data=None):
if flag:
return (self.rind, self.cind)
else:
result = adolc.colpack.sparse_jac_repeat(2, x, self.nnz, self.rind,
self.cind, self.values)
return result[3]
##### create the function
Jac_c_adolc = jac_c_adolc(x0)
###Hessian
# trace lagrangian function
adolc.trace_on(3)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = lagrangian(ax, xL, np.array([1.0]))
adolc.dependent(ay)
adolc.trace_off()
M = Y.shape[1]
nreal = npar-M
given = {'rind': np.concatenate((np.kron(np.arange(nreal), np.ones(npar,dtype='int')), np.arange(nreal, npar))),
'cind': np.concatenate((np.tile(np.arange(npar), nreal), np.arange(nreal, npar)))}
mask = np.where(given['rind'] <= given['cind'])
given['rind'] = given['rind'][mask]
given['cind'] = given['cind'][mask]
def hessLag_adolc(x, lagrange, obj_factor, flag, user_data=None):
if flag:
result = (given['rind'], given['cind'])
else:
result = np.ravel(adolc.hessian(3, x)[given['rind'],given['cind']], order="C")
return result
H2 = hessLag_adolc(x0, xL, 1.0, False)
H2a = hessLag_adolc(x0, xL, 1.0, True)
nnzh = len(given['rind'])
##Optimization
#PRELIMS: other things to pass to IPOPT
nvar = len(x0) #number of variables in the problem
x_L = np.array([-np.inf]*nvar, dtype=float) #box contraints on variables (none)
x_U = np.array([np.inf]*nvar, dtype=float)
#PRELIMS:define the (in)equality constraints
ncon = heq(ax).shape[0] #number of constraints
g_L = np.array([0]*ncon, dtype=float) #constraints are to equal 0
g_U = np.array([0]*ncon, dtype=float) #constraints are to equal 0
#PRELIMS: define the number of nonzeros in the jacobian
val = Jac_c_adolc(x0, False)
nnzj = len(val)
# create the nonlinear programming model
nlp2 = pyipopt.create(
nvar,
x_L,
x_U,
ncon,
g_L,
g_U,
nnzj,
nnzh,
adFun,
grFun,
const_adolc,
Jac_c_adolc,
hessLag_adolc
)
nlp2.num_option('expect_infeasible_problem_ctol', 1e-15)
nlp2.int_option('max_iter', 100)
nlp2.num_option('dual_inf_tol', 1e-3)
nlp2.num_option('constr_viol_tol', 1e-3)
nlp2.num_option('tol', 1e-6)
nlp2.int_option('print_level', 0)
t1 = time()
out = nlp2.solve(x0)
t1 = time()-t1
t1 = t1 + tPL
# free the model
nlp2.close()
os.chdir(ccd)
shutil.rmtree(user + '/Documents/adolc%i_%i'%(i, b))
output = np.concatenate((out[0][:6], t1, np.array([out[5]])))
return output
def _minimize(self, initial_val, loss_grad_func, equality_funcs,
equality_grad_funcs, inequality_funcs, inequality_grad_funcs,
packed_bounds, step_callback, optimizer_kwargs):
# initial value should be float64
initial_val = np.array(initial_val, dtype=np.float_)
# objective function
def eval_f(x, user_data=None):
loss, _ = loss_grad_func(x)
return np.array(loss, dtype=np.float_)
# gradient of objective function
def eval_grad_f(x, user_data=None):
_, grad_f = loss_grad_func(x)
return np.array(grad_f, dtype=np.float_)
# gradient function (first inequalities then equalities)
def eval_g(x, user_data=None):
inequalities = [inequality_funcs[i](x) for i in range(nineqcon)]
equalities = [equality_funcs[i](x) for i in range(neqcon)]
return np.array(inequalities + equalities, dtype=np.float_).reshape(ncon, )
# hessian of the lagrangian (first inequalities then equalities)
def eval_h(x, lagrange, obj_factor, flag, user_data=None):
rows, cols = np.tril_indices(nvar)
if flag:
return (np.array(rows, dtype=np.int_), np.array(cols, dtype=np.int_))
else:
loss = [loss_hessian_func(x)]
inequalities = [inequality_hessian_funcs[i](x) for i in range(nineqcon)]
equalities = [equality_hessian_funcs[i](x) for i in range(neqcon)]
values = np.zeros([nvar, nvar])
values += obj_factor * loss[0][0]
for idc in range(nineqcon):
values += lagrange[idc] * inequalities[idc][0]
for idc in range(neqcon):
values += lagrange[idc + nineqcon] * equalities[idc][0]
return np.array(values.reshape(nvar, nvar)[rows, cols], dtype=np.float_)
# jacobian for gradient (first inequalities the equalities)
def eval_jac_g(x, flag, user_data=None):
rows, cols = np.indices((ncon, nvar))
if flag:
return (np.array(rows.reshape(-1, 1), dtype=np.int_), np.array(cols.reshape(-1, 1), dtype=np.int_))
else:
inequalities = [inequality_grad_funcs[i](x) for i in range(nineqcon)]
equalities = [equality_grad_funcs[i](x) for i in range(neqcon)]
values = np.empty([ncon, nvar])
for idc in range(nineqcon):
values[idc, :] = inequalities[idc][0]
for idc in range(neqcon):
values[idc + nineqcon, :] = equalities[idc][0]
return np.array(values.reshape(ncon * nvar, ), dtype=np.float_)
# box constraints on the variables
nvar = int(np.sum([np.prod(self._vars[i].get_shape().as_list()) for i in range(len(self._vars))]))
if self._packed_bounds is None:
x_L = -np.ones((nvar), dtype=np.float_) * np.inf
x_U = np.ones((nvar), dtype=np.float_) * np.inf
else:
x_L, x_U = zip(*self._packed_bounds)
x_L = np.array(x_L, dtype=np.float_)
x_U = np.array(x_U, dtype=np.float_)
# inequality constraints as g(x)>=0 and equality constraints as h(x)=0
nineqcon = len(self._inequalities)
neqcon = len(self._equalities)
ncon = nineqcon + neqcon
g_L_ineq = np.zeros((nineqcon), dtype=np.float_)
g_U_ineq = np.ones((nineqcon), dtype=np.float_) * 2.0 * pow(10.0, 19)
g_L_eq = np.zeros((neqcon), dtype=np.float_)
g_U_eq = np.zeros((neqcon), dtype=np.float_)
g_L = np.concatenate((g_L_ineq, g_L_eq), axis=0)
g_U = np.concatenate((g_U_ineq, g_U_eq), axis=0)
nnzj = nvar * ncon
nnzh = int(nvar * (nvar + 1) / 2)
minimize_args = [nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g]
# create nlp in ipopt
import pyipopt
# log_level decides if logs from pyipopt are desired -- these are logs on
# top of what is returned from ipopt set by "print_level"; see below
if "log_level" in optimizer_kwargs["options"]:
pyipopt.set_loglevel(optimizer_kwargs["options"]["log_level"])
nlp = pyipopt.create(*minimize_args)
# check https://www.coin-or.org/Ipopt/documentation/node40.html
# for more options and default settings
# default print_level=5
# default max_iter=3000
# default tol=1e-8
for optvar in optimizer_kwargs["options"]:
if optvar is "log_level":
print
elif type(optimizer_kwargs["options"][optvar]) is np.str:
nlp.str_option(optvar, optimizer_kwargs["options"][optvar])
elif type(optimizer_kwargs["options"][optvar]) is np.int:
nlp.int_option(optvar, optimizer_kwargs["options"][optvar])
else:
nlp.num_option(optvar, optimizer_kwargs["options"][optvar])
result_x, zl, zu, constraint_multipliers, result_f, status = nlp.solve(initial_val)
nlp.close()
message_lines = [
'Optimization terminated with:',
' Message: %s',
' Objective function value: %f',
]
message_args = [status, result_f]
logging.info('\n'.join(message_lines), *message_args)
print("Optimization terminated with message: {}".format(status))
return result_x
def test_ipopt_optimization():
"""
This test checks
1. the sparse functionality of pyadolc
2. the execution speed compared to the direct sparse computation
3. run the optimization with the derivatives provided by pyadolc
IPOPT is an interior point algorithm to solve
min f(x)
x in R^n
s.t. g_L <= g(x) <= g_U
x_L <= x <= x_U
"""
try:
import pyipopt
except:
#print '"pyipopt is not installed, skipping test'
#return
raise NotImplementedError("pyipopt is not installed, skipping test")
import time
nvar = 4
x_L = numpy.ones((nvar), dtype=numpy.float_) * 1.0
x_U = numpy.ones((nvar), dtype=numpy.float_) * 5.0
ncon = 2
g_L = numpy.array([25.0, 40.0])
g_U = numpy.array([2.0*pow(10.0, 19), 40.0])
def eval_f(x, user_data = None):
assert len(x) == 4
return x[0] * x[3] * (x[0] + x[1] + x[2]) + x[2]
def eval_grad_f(x, user_data = None):
assert len(x) == 4
grad_f = numpy.array([
x[0] * x[3] + x[3] * (x[0] + x[1] + x[2]) ,
x[0] * x[3],
x[0] * x[3] + 1.0,
x[0] * (x[0] + x[1] + x[2])
])
return grad_f;
def eval_g(x, user_data= None):
assert len(x) == 4
return numpy.array([
x[0] * x[1] * x[2] * x[3],
x[0]*x[0] + x[1]*x[1] + x[2]*x[2] + x[3]*x[3]
])
nnzj = 8
def eval_jac_g(x, flag, user_data = None):
if flag:
return (numpy.array([0, 0, 0, 0, 1, 1, 1, 1]),
numpy.array([0, 1, 2, 3, 0, 1, 2, 3]))
else:
assert len(x) == 4
return numpy.array([ x[1]*x[2]*x[3],
x[0]*x[2]*x[3],
x[0]*x[1]*x[3],
x[0]*x[1]*x[2],
2.0*x[0],
2.0*x[1],
2.0*x[2],
2.0*x[3] ])
nnzh = 10
def eval_h(x, lagrange, obj_factor, flag, user_data = None):
if flag:
hrow = [0, 1, 1, 2, 2, 2, 3, 3, 3, 3]
hcol = [0, 0, 1, 0, 1, 2, 0, 1, 2, 3]
return (numpy.array(hcol,dtype=int), numpy.array(hrow,dtype=int))
else:
values = numpy.zeros((10), numpy.float_)
values[0] = obj_factor * (2*x[3])
values[1] = obj_factor * (x[3])
values[2] = 0
values[3] = obj_factor * (x[3])
values[4] = 0
values[5] = 0
values[6] = obj_factor * (2*x[0] + x[1] + x[2])
values[7] = obj_factor * (x[0])
values[8] = obj_factor * (x[0])
values[9] = 0
values[1] += lagrange[0] * (x[2] * x[3])
values[3] += lagrange[0] * (x[1] * x[3])
values[4] += lagrange[0] * (x[0] * x[3])
values[6] += lagrange[0] * (x[1] * x[2])
values[7] += lagrange[0] * (x[0] * x[2])
values[8] += lagrange[0] * (x[0] * x[1])
values[0] += lagrange[1] * 2
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
x0 = numpy.array([1.0, 5.0, 5.0, 1.0])
pi0 = numpy.array([1.0, 1.0])
# check that adolc gives the same answers as derivatives calculated by hand
trace_on(1)
ax = adouble(x0)
independent(ax)
ay = eval_f(ax)
dependent(ay)
trace_off()
trace_on(2)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay)
trace_off()
trace_on(3)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay[0])
trace_off()
trace_on(4)
ax = adouble(x0)
independent(ax)
ay = eval_g(ax)
dependent(ay[1])
trace_off()
def eval_f_adolc(x, user_data = None):
return function(1,x)[0]
def eval_grad_f_adolc(x, user_data = None):
return gradient(1,x)
def eval_g_adolc(x, user_data= None):
return function(2,x)
def eval_jac_g_adolc(x, flag, user_data = None):
options = numpy.array([1,1,0,0],dtype=int)
result = sparse.sparse_jac_no_repeat(2,x,options)
if flag:
return (numpy.asarray(result[1],dtype=int), numpy.asarray(result[2],dtype=int))
else:
return result[3]
def eval_h_adolc(x, lagrange, obj_factor, flag, user_data = None):
options = numpy.array([0,0],dtype=int)
assert numpy.ndim(x) == 1
assert numpy.size(x) == 4
result_f = sparse.sparse_hess_no_repeat(1, x, options)
result_g0 = sparse.sparse_hess_no_repeat(3, x,options)
result_g1 = sparse.sparse_hess_no_repeat(4, x,options)
Hf = scipy.sparse.coo_matrix( (result_f[3], (result_f[1], result_f[2])), shape=(4, 4))
Hg0 = scipy.sparse.coo_matrix( (result_g0[3], (result_g0[1], result_g0[2])), shape=(4, 4))
Hg1 = scipy.sparse.coo_matrix( (result_g1[3], (result_g1[1], result_g1[2])), shape=(4, 4))
H = Hf + Hg0 + Hg1
H = H.tocoo()
if flag:
hrow = H.row
hcol = H.col
return (numpy.array(hcol,dtype=int), numpy.array(hrow,dtype=int))
else:
values = numpy.zeros((10), float)
values[:] = H.data
return values
# function of f
assert_almost_equal(eval_f(x0), eval_f_adolc(x0))
# gradient of f
assert_array_almost_equal(eval_grad_f(x0), eval_grad_f_adolc(x0))
# function of g
assert_array_almost_equal(eval_g(x0), function(2,x0))
# sparse jacobian of g
assert_array_equal(eval_jac_g_adolc(x0,True)[0], eval_jac_g(x0,True)[0])
assert_array_equal(eval_jac_g_adolc(x0,True)[1], eval_jac_g(x0,True)[1])
assert_array_equal(eval_jac_g_adolc(x0,False), eval_jac_g(x0,False))
# sparse hessian of the lagrangian
lagrange = numpy.ones(2,dtype=float)
obj_factor = 1.
x0 = numpy.random.rand(4)
result = (eval_h(x0, lagrange, obj_factor, False), eval_h(x0, lagrange, obj_factor, True))
result_adolc = (eval_h_adolc(x0, lagrange, obj_factor, False), eval_h_adolc(x0, lagrange, obj_factor, True))
H = scipy.sparse.coo_matrix( result, shape=(4, 4))
H_adolc = scipy.sparse.coo_matrix( result_adolc, shape=(4, 4))
H = H.todense()
H_adolc = H_adolc.todense()
assert_array_almost_equal( H, H_adolc.T)
# test optimization with PYIPOPT
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h)
start_time = time.time()
result = nlp.solve(x0)
end_time = time.time()
nlp.close()
pure_python_optimization_time = end_time - start_time
nlp_adolc = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f_adolc, eval_grad_f_adolc, eval_g_adolc, eval_jac_g_adolc, eval_h_adolc)
start_time = time.time()
result_adolc = nlp_adolc.solve(x0)
end_time = time.time()
nlp_adolc.close()
adolc_optimization_time = end_time - start_time
print 'optimization time with derivatives computed by adolc = ', adolc_optimization_time
print 'optimization time with derivatives computed by hand = ',pure_python_optimization_time
assert adolc_optimization_time / pure_python_optimization_time < 10
# this works with the pyipopt version from code.google.com
assert_array_almost_equal(result[0], result_adolc[0])
assert_array_almost_equal(result[1], result_adolc[1])
assert_array_almost_equal(result[2], result_adolc[2])
assert_array_almost_equal(result[3], result_adolc[3])
print eval_f(x0) assert len(eval_grad_f(x0)) == n print eval_grad_f(x0) assert len(eval_g(x0)) == m a = eval_jac_g(x0, True) print "row, col ", a[0], a[1] print eval_jac_g(x0, False) print eval_h(True, True, True, True) print eval_h(nlp.x0, nlp.pi0, 1.0, True) def gf2(x, data = None): return array([0.0 for i in xrange(m)]) def f2(x, data = None): return 0 """ print "Solving problem with approximate hession calculation now" problem = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, 0, eval_f, eval_grad_f, eval_g, eval_jac_g) problem.solve(x0) problem.close() print "Solving problem with exact hession calculation now" mynlp = pyipopt.create(n, xl, xu, m, gl, gu, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h, applynew) mynlp.solve(x0) mynlp.close()
def Ipopt_Solve(problem):
"""Solves a Nexus optimization problem using ipopt
Assumptions:
You can actually install ipopt on your machine
Source:
N/A
Inputs:
problem [nexus()]
Outputs:
result [array]
Properties Used:
None
"""
# Pull out the basic problem
inp = problem.optimization_problem.inputs
obj = problem.optimization_problem.objective
con = problem.optimization_problem.constraints
# Number of input variables and constrains
nvar = len(inp)
ncon = len(con)
# Set inputs
ini = inp[:,1] # Initials
bnd = inp[:,2] # Bounds
scl = inp[:,3] # Scale
# Scaled initials
x0 = ini/scl
x0 = x0.astype(float)
# Nonzero jacobians and hessians, fix this
nnzj = ncon*nvar
nnzh = nvar*nvar
# Bounds for inputs and constraints
flbd = np.zeros_like(ini)
fubd = np.zeros_like(ini)
for ii in xrange(0,nvar):
flbd[ii] = (bnd[ii][0]/scl[ii])
fubd[ii] = (bnd[ii][1]/scl[ii])
g_L = np.zeros_like(con)
g_U = np.zeros_like(con)
# Setup constraints
for ii in xrange(0,len(con)):
name = con[ii][0]
edge = con[ii][2]
if con[ii][1]=='<':
g_L[ii] = -np.inf
g_U[ii] = edge
elif con[ii][1]=='>':
g_L[ii] = edge
g_U[ii] = np.inf
elif con[ii][1]=='=':
g_L[ii] = edge
g_U[ii] = edge
# Instantiate the problem and set objective
import pyipopt #import down here to allow SUAVE to run without the user having Ipopt
flbd = flbd.astype(float)
fubd = fubd.astype(float)
g_L = g_L.astype(float)
g_U = g_U.astype(float)
# Create the problem
nlp = pyipopt.create(nvar, flbd, fubd, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g)
nlp.str_option('derivative_test_print_all','yes')
nlp.str_option('derivative_test','first-order')
# Solve the problem
result = nlp.solve(x0,problem)
nlp.close()
return result
## IPOPT SOLUTION
start = time.time()
nvar = data.numX
xl = np.zeros(data.numX)
xu = 2e19 * np.ones(data.numX)
m = 0
gl = np.zeros(1)
gu = 2e19 * np.ones(1)
g_L = np.array([], dtype=float)
g_U = np.array([], dtype=float)
nnzj = 0
nnzh = int(data.numX * (data.numX + 1) / 2)
nlp = pyipopt.create(nvar, xl, xu, m, g_L, g_U, nnzj, nnzh, evaluateFunction,
evaluateGradient, eval_g, eval_jac_g, evaluateHessian)
nlp.num_option('tol', 1e-5)
nlp.int_option("print_level", 5)
nlp.str_option('hessian_approximation', 'limited-memory')
nlp.str_option('mu_strategy', 'adaptive')
nlp.str_option('mu_oracle', 'probing')
nlp.str_option('linear_solver', 'ma97')
nlp.num_option('acceptable_tol', 1e-2)
nlp.int_option("acceptable_iter", 5)
nlp.num_option('acceptable_obj_change_tol', 5e-1)
mylines = [
float(myline) for myline in [
line.rstrip('\n') for line in open(
'/home/wilmer/Dropbox/IpOptSolver/currentIntensities.txt')
]
]
# sparse jacobian of g
assert_array_equal(eval_jac_g_adolc(x0, True)[0], eval_jac_g(x0, True)[0])
assert_array_equal(eval_jac_g_adolc(x0, True)[1], eval_jac_g(x0, True)[1])
assert_array_equal(eval_jac_g_adolc(x0, False), eval_jac_g(x0, False))
# test optimization with PYIPOPT
nvar = 4
x_L = numpy.ones((nvar), dtype=numpy.float_) * 1.0
x_U = numpy.ones((nvar), dtype=numpy.float_) * 5.0
ncon = 2
g_L = numpy.array([25.0, 40.0])
g_U = numpy.array([2.0 * pow(10.0, 19), 40.0])
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g, eval_h)
start_time = time.time()
result = nlp.solve(x0)
end_time = time.time()
nlp.close()
pure_python_optimization_time = end_time - start_time
nlp_adolc = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh,
eval_f_adolc, eval_grad_f_adolc, eval_g_adolc,
eval_jac_g_adolc, eval_h_adolc)
start_time = time.time()
result_adolc = nlp_adolc.solve(x0)
end_time = time.time()
nlp_adolc.close()
def Ipopt_Solve(problem):
"""Solves a Nexus optimization problem using ipopt
Assumptions:
You can actually install ipopt on your machine
Source:
N/A
Inputs:
problem [nexus()]
Outputs:
result [array]
Properties Used:
None
"""
# Pull out the basic problem
inp = problem.optimization_problem.inputs
obj = problem.optimization_problem.objective
con = problem.optimization_problem.constraints
# Number of input variables and constrains
nvar = len(inp)
ncon = len(con)
# Set inputs
ini = inp[:, 1] # Initials
bndl = inp[:, 2] # Bounds
bndu = inp[:, 3] # Bounds
scl = inp[:, 4] # Scale
# Scaled initials
x0 = ini / scl
x0 = x0.astype(float)
# Nonzero jacobians and hessians, fix this
nnzj = ncon * nvar
nnzh = nvar * nvar
# Bounds for inputs and constraints
flbd = np.zeros_like(ini)
fubd = np.zeros_like(ini)
for ii in range(0, nvar):
flbd[ii] = (bndl[ii] / scl[ii])
fubd[ii] = (bndu[ii] / scl[ii])
g_L = np.zeros_like(con)
g_U = np.zeros_like(con)
# Setup constraints
for ii in range(0, len(con)):
name = con[ii][0]
edge = con[ii][2]
if con[ii][1] == '<':
g_L[ii] = -np.inf
g_U[ii] = edge
elif con[ii][1] == '>':
g_L[ii] = edge
g_U[ii] = np.inf
elif con[ii][1] == '=':
g_L[ii] = edge
g_U[ii] = edge
# Instantiate the problem and set objective
import pyipopt #import down here to allow SUAVE to run without the user having Ipopt
flbd = flbd.astype(float)
fubd = fubd.astype(float)
g_L = g_L.astype(float)
g_U = g_U.astype(float)
# Create the problem
nlp = pyipopt.create(nvar, flbd, fubd, ncon, g_L, g_U, nnzj, nnzh, eval_f,
eval_grad_f, eval_g, eval_jac_g)
nlp.str_option('derivative_test_print_all', 'yes')
nlp.str_option('derivative_test', 'first-order')
# Solve the problem
result = nlp.solve(x0, problem)
nlp.close()
return result
def __solve__(self,
opt_problem={},
sens_type='FD',
store_sol=True,
disp_opts=False,
store_hst=False,
hot_start=False,
sens_mode='',
sens_step={},
*args,
**kwargs):
"""
Run Optimizer (Optimize Routine)
**Keyword arguments:**
- opt_problem -> INST: Optimization instance
- sens_type -> STR/FUNC: Gradient type, *Default* = 'FD'
- store_sol -> BOOL: Store solution in Optimization class flag,
*Default* = True
- disp_opts -> BOOL: Flag to display options in solution text, *Default*
= False
- store_hst -> BOOL/STR: Flag/filename to store optimization history,
*Default* = False
- hot_start -> BOOL/STR: Flag/filename to read optimization history,
*Default* = False
- sens_mode -> STR: Flag for parallel gradient calculation, *Default* =
''
- sens_step -> FLOAT: Sensitivity setp size, *Default* = {} [corresponds
to 1e-6 (FD), 1e-20(CS)]
Documentation last updated: Feb. 2, 2011 - Peter W. Jansen
"""
self.pll = False
self.myrank = 0
myrank = self.myrank
tmp_file = False
def_fname = self.options['output_file'][1].split('.')[0]
if isinstance(store_hst, str):
if isinstance(hot_start, str):
if (myrank == 0):
if (store_hst == hot_start):
hos_file = History(hot_start, 'r', self)
log_file = History(store_hst + '_tmp', 'w', self,
opt_problem.name)
tmp_file = True
else:
hos_file = History(hot_start, 'r', self)
log_file = History(store_hst, 'w', self,
opt_problem.name)
self.sto_hst = True
self.hot_start = True
elif hot_start:
if (myrank == 0):
hos_file = History(store_hst, 'r', self)
log_file = History(store_hst + '_tmp', 'w', self,
opt_problem.name)
tmp_file = True
self.sto_hst = True
self.hot_start = True
else:
if (myrank == 0):
log_file = History(store_hst, 'w', self, opt_problem.name)
self.sto_hst = True
self.hot_start = False
elif store_hst:
if isinstance(hot_start, str):
if (hot_start == def_fname):
if (myrank == 0):
hos_file = History(hot_start, 'r', self)
log_file = History(def_fname + '_tmp', 'w', self,
opt_problem.name)
tmp_file = True
else:
if (myrank == 0):
hos_file = History(hot_start, 'r', self)
log_file = History(def_fname, 'w', self,
opt_problem.name)
self.sto_hst = True
self.hot_start = True
elif hot_start:
if (myrank == 0):
hos_file = History(def_fname, 'r', self)
log_file = History(def_fname + '_tmp', 'w', self,
opt_problem.name)
tmp_file = True
self.sto_hst = True
self.hot_start = True
else:
if (myrank == 0):
log_file = History(def_fname, 'w', self, opt_problem.name)
self.sto_hst = True
self.hot_start = False
else:
self.sto_hst = False
self.hot_start = False
gradient = Gradient(opt_problem, sens_type, sens_mode, sens_step,
*args, **kwargs)
def eval_f(x, user_data=None):
"""IPOPT - Objective Value Function."""
# Variables Groups Handling
if opt_problem.use_groups:
xg = {}
for group in group_ids.keys():
if (group_ids[group][1] - group_ids[group][0] == 1):
xg[group] = x[group_ids[group][0]]
else:
xg[group] = x[group_ids[group][0]:group_ids[group][1]]
xn = xg
else:
xn = x
# Flush Output Files
self.flushFiles()
# Evaluate User Function
fail = 0
# if (myrank == 0):
# if self.hot_start:
# [vals,hist_end] = hos_file.read(ident=['obj', 'con', 'fail'])
# if hist_end:
# self.hot_start = False
# hos_file.close()
# else:
# [ff,gg,fail] = [vals['obj'][0][0],vals['con'][0],int(vals['fail'][0][0])]
#
#
# if self.pll:
# self.hot_start = Bcast(self.hot_start,root=0)
# if self.hot_start and self.pll:
# [ff,gg,fail] = Bcast([ff,gg,fail],root=0)
# else:
[ff, gg, fail] = opt_problem.obj_fun(xn, *args, **kwargs)
# Store History
if (myrank == 0):
if self.sto_hst:
log_file.write(x, 'x')
log_file.write(ff, 'obj')
log_file.write(gg, 'con')
log_file.write(fail, 'fail')
# Objective Assigment
if isinstance(ff, complex):
f = ff.astype(float)
else:
f = ff
# Constraints Assigment
g = numpy.zeros(len(opt_problem._constraints.keys()))
for i in range(len(opt_problem._constraints.keys())):
if isinstance(gg[i], complex):
g[i] = gg[i].astype(float)
else:
g[i] = gg[i]
return f
def eval_g(x, user_data=None):
# Variables Groups Handling
if opt_problem.use_groups:
xg = {}
for group in group_ids.keys():
if (group_ids[group][1] - group_ids[group][0] == 1):
xg[group] = x[group_ids[group][0]]
else:
xg[group] = x[group_ids[group][0]:group_ids[group][1]]
xn = xg
else:
xn = x
# Flush Output Files
self.flushFiles()
# Evaluate User Function
fail = 0
# if (myrank == 0):
# if self.hot_start:
# [vals,hist_end] = hos_file.read(ident=['obj', 'con', 'fail'])
# if hist_end:
# self.hot_start = False
# hos_file.close()
# else:
# [ff,gg,fail] = [vals['obj'][0][0],vals['con'][0],int(vals['fail'][0][0])]
# if self.pll:
# self.hot_start = Bcast(self.hot_start,root=0)
# if self.hot_start and self.pll:
# [ff,gg,fail] = Bcast([ff,gg,fail],root=0)
# else:
[ff, gg, fail] = opt_problem.obj_fun(xn, *args, **kwargs)
# Store History
if (myrank == 0):
if self.sto_hst:
log_file.write(x, 'x')
log_file.write(ff, 'obj')
log_file.write(gg, 'con')
log_file.write(fail, 'fail')
# Objective Assigment
if isinstance(ff, complex):
f = ff.astype(float)
else:
f = ff
# Constraints Assigment
g = numpy.zeros(len(opt_problem._constraints.keys()))
for i in range(len(opt_problem._constraints.keys())):
if isinstance(gg[i], complex):
g[i] = gg[i].astype(float)
else:
g[i] = gg[i]
return g
def eval_grad_f(x, user_data=None):
"""IPOPT - Objective/Constraint Gradients Function."""
# if self.hot_start:
# if (myrank == 0):
# [vals,hist_end] = hos_file.read(ident=['grad_obj','grad_con'])
# if hist_end:
# self.hot_start = False
# hos_file.close()
# else:
# dff = vals['grad_obj'][0].reshape((len(opt_problem._objectives.keys()),len(opt_problem._variables.keys())))
# dgg = vals['grad_con'][0].reshape((len(opt_problem._constraints.keys()),len(opt_problem._variables.keys())))
#
#
# if self.pll:
# self.hot_start = Bcast(self.hot_start,root=0)
#
# if self.hot_start and self.pll:
# [dff,dgg] = Bcast([dff,dgg],root=0)
#
# if not self.hot_start:
[f, g, fail] = opt_problem.obj_fun(x, *args, **kwargs)
dff, dgg = gradient.getGrad(x, group_ids, [f], g, *args, **kwargs)
# Store History
if self.sto_hst and (myrank == 0):
log_file.write(dff, 'grad_obj')
log_file.write(dgg, 'grad_con')
# Gradient Assignment
df = numpy.zeros(len(opt_problem._variables.keys()))
for i in range(len(opt_problem._variables.keys())):
df[i] = dff[0, i]
return df
def eval_grad_g(x, flag, user_data=None):
# if self.hot_start:
# if (myrank == 0):
# [vals,hist_end] = hos_file.read(ident=['grad_obj','grad_con'])
# if hist_end:
# self.hot_start = False
# hos_file.close()
# else:
# dff = vals['grad_obj'][0].reshape((len(opt_problem._objectives.keys()),len(opt_problem._variables.keys())))
# dgg = vals['grad_con'][0].reshape((len(opt_problem._constraints.keys()),len(opt_problem._variables.keys())))
#
#
# if self.pll:
# self.hot_start = Bcast(self.hot_start,root=0)
#
# if self.hot_start and self.pll:
# [dff,dgg] = Bcast([dff,dgg],root=0)
#
# if not self.hot_start:
if flag:
a = numpy.zeros(
len(opt_problem._variables.keys()) *
len(opt_problem._constraints.keys()), int)
b = numpy.zeros(
len(opt_problem._variables.keys()) *
len(opt_problem._constraints.keys()), int)
for i in range(len(opt_problem._constraints.keys())):
for j in range(len(opt_problem._variables.keys())):
a[i * len(opt_problem._variables.keys()) + j] = i
b[i * len(opt_problem._variables.keys()) + j] = j
return (a, b)
else:
[f, g, fail] = opt_problem.obj_fun(x, *args, **kwargs)
dff, dgg = gradient.getGrad(x, group_ids, [f], g, *args,
**kwargs)
# Store History
if self.sto_hst and (myrank == 0):
log_file.write(dff, 'grad_obj')
log_file.write(dgg, 'grad_con')
# Gradient Assignment
a = numpy.zeros([
len(opt_problem._variables.keys()) *
len(opt_problem._constraints.keys())
])
for i in range(len(opt_problem._constraints.keys())):
for j in range(len(opt_problem._variables.keys())):
a[i * len(opt_problem._variables.keys()) +
j] = dgg[i, j]
return a
# Variables Handling
nvar = len(opt_problem._variables.keys())
xl = []
xu = []
xx = []
for key in opt_problem._variables.keys():
if opt_problem._variables[key].type == 'c':
xl.append(opt_problem._variables[key].lower)
xu.append(opt_problem._variables[key].upper)
xx.append(opt_problem._variables[key].value)
elif opt_problem._variables[key].type == 'i':
raise IOError('IPOPT cannot handle integer design variables')
elif opt_problem._variables[key].type == 'd':
raise IOError('IPOPT cannot handle discrete design variables')
xl = numpy.array(xl)
xu = numpy.array(xu)
xx = numpy.array(xx)
# Variables Groups Handling
group_ids = {}
if opt_problem.use_groups:
k = 0
for key in opt_problem._vargroups.keys():
group_len = len(opt_problem._vargroups[key]['ids'])
group_ids[opt_problem._vargroups[key][
'name']] = [k, k + group_len]
k += group_len
# Constraints Handling
ncon = len(opt_problem._constraints.keys())
blc = []
buc = []
if ncon > 0:
for key in opt_problem._constraints.keys():
if opt_problem._constraints[key].type == 'e':
blc.append(opt_problem._constraints[key].equal)
buc.append(opt_problem._constraints[key].equal)
elif opt_problem._constraints[key].type == 'i':
blc.append(opt_problem._constraints[key].lower)
buc.append(opt_problem._constraints[key].upper)
else:
if ((store_sol) and (myrank == 0)):
print("Optimization Problem Does Not Have Constraints\n")
print("Unconstrained Optimization Initiated\n")
ncon = 1
blc.append(-inf)
buc.append(inf)
blc = numpy.array(blc)
buc = numpy.array(buc)
# Objective Handling
objfunc = opt_problem.obj_fun
nobj = len(opt_problem._objectives.keys())
ff = []
for key in opt_problem._objectives.keys():
ff.append(opt_problem._objectives[key].value)
ff = numpy.array(ff)
# Create an IPOPT instance problem
nnzj = nvar * ncon
nnzh = nvar * nvar
ipopt = pyipopt.create(nvar, xl, xu, ncon, blc, buc, nnzj, nnzh,
eval_f, eval_grad_f, eval_g, eval_grad_g)
# Setup Options
optionss = self.options.copy()
del optionss['defaults']
for i in optionss:
if not self.options['defaults'][i][1] == optionss[i][1]:
if self.options[i][0].__name__ == 'int':
ipopt.int_option(i, self.options[i][1])
if self.options[i][0].__name__ == 'float':
ipopt.num_option(i, self.options[i][1])
if self.options[i][0].__name__ == 'str':
ipopt.str_option(i, self.options[i][1])
# Run IPOPT
t0 = time.time()
r = ipopt.solve(xx)
sol_time = time.time() - t0
if (myrank == 0):
if self.sto_hst:
log_file.close()
if tmp_file:
hos_file.close()
name = hos_file.filename
os.remove(name + '.cue')
os.remove(name + '.bin')
os.rename(name + '_tmp.cue', name + '.cue')
os.rename(name + '_tmp.bin', name + '.bin')
ipopt.close()
# Store Results
sol_inform = {}
print(r)
sol_inform['value'] = r[-1] # ifail[0]
sol_inform['text'] = self.getInform(r[-1]) # self.getInform(ifail[0])
if store_sol:
sol_name = 'IPOPT Solution to ' + opt_problem.name
sol_options = copy.copy(self.options)
if 'default' in sol_options:
del sol_options['defaults']
sol_evals = 0
sol_vars = copy.deepcopy(opt_problem._variables)
i = 0
x = r[0]
for key in sol_vars.keys():
sol_vars[key].value = x[i]
i += 1
sol_objs = copy.deepcopy(opt_problem._objectives)
sol_objs[0].value = r[4]
sol_cons = {}
if ncon > 0:
sol_lambda = r[3]
else:
sol_lambda = {}
opt_problem.addSol(
self.__class__.__name__,
sol_name,
objfunc,
sol_time,
sol_evals,
sol_inform,
sol_vars,
sol_objs,
sol_cons,
sol_options,
display_opts=disp_opts,
Lambda=sol_lambda,
Sensitivities=sens_type,
myrank=myrank,
arguments=args,
**kwargs)
return ff, xx, sol_inform # ifail[0]
def solve(self, x0=None, options={}):
"""Solve the nonlinear problem from a specified starting point.
Arguments:
x0 -- array of length self.nvars with initial guesses for values of
self.variables (in that order), or a scalar to be used as the initial
guess for all variables; if not supplied, all variables will be
set initially to 1.0.
options -- dictionary {'option': value} of options to pass to
IPOPT. Options will be taken from (in increasing order of priority)
IPOPT's default values, the module-wide default_ipopt_options, those
specified in self.fixed_variable_treatment or self.ipopt_options,
this argument, or the ipopt.opt file (if any) in the working directory.
Errors will result if invalid names or values for options are given
here.
Returns:
x -- array of variable values returned by IPOPT (ordered as in
self.variables)
Also sets self.x, self.zl, self.zu, self.constraint_multipliers,
self.obj_value, and self.status to x, the lower and upper
bound multipliers, the Lagrange multipliers associated with
the constraint functions, the final objective function value,
and the optimzation status, respectively, and sets self.soln
to a dictionary mapping each variable key to its value in
self.x.
This method does not recompile the objective, constraint and
derivative functions if self.compile() has already been
called. If anything except bounds on variables, bounds on
constraints, and/or parameter values has been changed since
the last time self.compile() has been called, self.compile
must be called again before solving, or this method will
attempt to solve an out-of-date version of the problem.
Each call to solve does create a new pyipopt problem instance
as self.nlp, closing the old one beforehand if self.active_nlp
is true, and resetting self.active_nlp to true afterwards.
"""
if not self._compiled:
self.compile()
if x0 is None:
x0 = np.ones(self.nvar)
elif np.isscalar(x0):
x0 = x0 * np.ones(self.nvar)
else:
# Assume this is a vector. Check its length, as trying to
# proceed with an inappropriately sized starting point may
# lead to crashes.
if len(x0) != self.nvar:
message = (
'Starting point has wrong dimension (needed %d, given %d)'
% (self.nvar, len(x0)))
raise ValueError(message)
self.close_nlp()
x_L, x_U = self.make_variable_bound_vectors()
g_L, g_U = self.make_constraint_bound_vectors()
# The following avoids a segmentation fault that results
# when bound methods are supplied as evaluation functions
# to pyipopt, which I don't really understand:
eval_f = lambda x, user_data=None: self.eval_f(x)
eval_grad_f = lambda x, user_data=None: self.eval_grad_f(x)
eval_g = lambda x, user_data=None: self.eval_g(x)
eval_jac_g = lambda x, flag, user_data=None: self.eval_jac_g(x, flag)
eval_h= lambda x, lagrange, obj_factor, flag, user_data=None: \
self.eval_h(x,lagrange,obj_factor,flag)
self.nlp = pyipopt.create(self.nvar, x_L, x_U, self.ncon, g_L, g_U,
self._nnzj, self._nnzh, eval_f, eval_grad_f,
eval_g, eval_jac_g, eval_h)
self.active_nlp = True
# Handle generic IPOPT options
all_options = {}
all_options.update(default_ipopt_options)
all_options['fixed_variable_treatment'] = self.fixed_variable_treatment
all_options.update(self.ipopt_options)
all_options.update(options)
for option, value in all_options.iteritems():
if isinstance(value, str):
self.nlp.str_option(option, value)
elif isinstance(value, int):
self.nlp.int_option(option, value)
else:
self.nlp.num_option(option, value)
(self.x, self.zl, self.zu, self.constraint_multipliers, self.obj_value,
self.status) = self.nlp.solve(x0)
self.soln = dict(zip(self.variables, self.x))
if self.status not in (0, 1):
raise OptimizationFailure('IPOPT exited with status %d' %
self.status)
return self.x.copy()
# sparse jacobian of g
assert_array_equal(eval_jac_g_adolc(x0, True)[0], eval_jac_g(x0, True)[0])
assert_array_equal(eval_jac_g_adolc(x0, True)[1], eval_jac_g(x0, True)[1])
assert_array_equal(eval_jac_g_adolc(x0, False), eval_jac_g(x0, False))
# test optimization with PYIPOPT
nvar = 4
x_L = numpy.ones((nvar), dtype=numpy.float_) * 1.0
x_U = numpy.ones((nvar), dtype=numpy.float_) * 5.0
ncon = 2
g_L = numpy.array([25.0, 40.0])
g_U = numpy.array([2.0 * pow(10.0, 19), 40.0])
nlp = pyipopt.create(nvar, x_L, x_U, ncon, g_L, g_U, nnzj, nnzh, eval_f, eval_grad_f, eval_g, eval_jac_g, eval_h)
start_time = time.time()
result = nlp.solve(x0)
end_time = time.time()
nlp.close()
pure_python_optimization_time = end_time - start_time
nlp_adolc = pyipopt.create(
nvar,
x_L,
x_U,
ncon,
g_L,
g_U,
nnzj,
#PRELIMS: define the number of nonzeros in the jacobian
val = Jac_c_adolc(x0, False)
nnzj = len(val)
# create the nonlinear programming model
nlp2 = pyipopt.create(
nvar,
x_L,
x_U,
ncon,
g_L,
g_U,
nnzj,
nnzh,
adFun,
grFun,
const_adolc,
Jac_c_adolc,
hessLag_adolc
)
nlp2.num_option('expect_infeasible_problem_ctol', 1e-15)
nlp2.int_option('max_iter', 5000)
nlp2.num_option('dual_inf_tol', 1e-5)
nlp2.num_option('constr_viol_tol', 1e-5)
nlp2.num_option('tol', 1e-6)
def solve(self):
""" Solves AC optimal power flow.
"""
case = self.om.case
base_mva = case.base_mva
# TODO: Explain this value.
self.opt["cost_mult"] = 1e-4
# Unpack the OPF model.
bs, ln, gn, _ = self._unpack_model(self.om)
# Compute problem dimensions.
ipol, _, nb, nl, _, ny, nxyz = self._dimension_data(bs, ln, gn)
# Compute problem dimensions.
ng = len(gn)
# gpol = [g for g in gn if g.pcost_model == POLYNOMIAL]
# Indexes of constrained lines.
il = array([i for i,l in enumerate(ln) if 0.0 < l.rate_a < 1e10])
nl2 = len(il)
# Linear constraints (l <= A*x <= u).
A, l, u = self.om.linear_constraints()
# AA, bb = self._linear_constraints(self.om)
_, xmin, xmax = self._var_bounds()
# Select an interior initial point for interior point solver.
x0 = self._initial_interior_point(bs, gn, xmin, xmax, ny)
# Build admittance matrices.
Ybus, Yf, Yt = case.Y
# Optimisation variables.
Va = self.om.get_var("Va")
Vm = self.om.get_var("Vm")
Pg = self.om.get_var("Pg")
Qg = self.om.get_var("Qg")
# Adds a constraint on the reference bus angles.
# xmin, xmax = self._ref_bus_angle_constraint(bs, Va, xmin, xmax)
def f_fcn(x, user_data=None):
""" Evaluates the objective function.
"""
p_gen = x[Pg.i1:Pg.iN + 1] # Active generation in p.u.
q_gen = x[Qg.i1:Qg.iN + 1] # Reactive generation in p.u.
# Polynomial cost of P and Q.
xx = r_[p_gen, q_gen] * base_mva
if len(ipol) > 0:
f = sum([g.total_cost(xx[i]) for i,g in enumerate(gn)])
else:
f = 0
# Piecewise linear cost of P and Q.
if ny:
y = self.om.get_var("y")
ccost = csr_matrix((ones(ny),
(range(y.i1, y.iN + 1), zeros(ny))), shape=(nxyz, 1)).T
f = f + ccost * x
else:
ccost = zeros((1, nxyz))
# TODO: Generalised cost term.
return f
def df_fcn(x, usr_data=None):
""" Calculates gradient of the objective function.
"""
p_gen = x[Pg.i1:Pg.iN + 1] # Active generation in p.u.
q_gen = x[Qg.i1:Qg.iN + 1] # Reactive generation in p.u.
xx = r_[p_gen, q_gen] * base_mva
if ny > 0:
y = self.om.get_var("y")
iy = range(y.i1, y.iN + 1)
ccost = \
csr_matrix((ones(ny), (iy, zeros(ny))), shape=(nxyz, 1)).T
else:
ccost = zeros((1, nxyz))
# TODO: Generalised cost term.
iPg = range(Pg.i1, Pg.iN + 1)
iQg = range(Qg.i1, Qg.iN + 1)
# Polynomial cost of P and Q.
df_dPgQg = zeros((2 * ng, 1)) # w.r.t p.u. Pg and Qg
# df_dPgQg[ipol] = matrix([g.poly_cost(xx[i], 1) for g in gpol])
# for i, g in enumerate(gn):
# der = polyder(list(g.p_cost))
# df_dPgQg[i] = polyval(der, xx[i]) * base_mva
for i in ipol:
df_dPgQg[i] = \
base_mva * polyval(polyder(list(gn[i].p_cost)), xx[i])
df = zeros((nxyz, 1))
df[iPg] = df_dPgQg[:ng]
df[iQg] = df_dPgQg[ng:ng + ng]
# Piecewise linear cost of P and Q.
df = df + ccost.T
# TODO: Generalised cost term.
return asarray(df).flatten()
def g_fcn(x, usr_data=None):
""" Evaluates the non-linear constraint values.
"""
Pgen = x[Pg.i1:Pg.iN + 1] # Active generation in p.u.
Qgen = x[Qg.i1:Qg.iN + 1] # Reactive generation in p.u.
for i, g in enumerate(gn):
g.p = Pgen[i] * base_mva # active generation in MW
g.q = Qgen[i] * base_mva # reactive generation in MVAr
# Rebuild the net complex bus power injection vector in p.u.
Sbus = case.getSbus(bs)
Vang = x[Va.i1:Va.iN + 1]
Vmag = x[Vm.i1:Vm.iN + 1]
V = Vmag * exp(1j * Vang)
# Evaluate the power flow equations.
mis = V * conj(Ybus * V) - Sbus
# Equality constraints (power flow).
g = r_[mis.real, # active power mismatch for all buses
mis.imag] # reactive power mismatch for all buses
# Inequality constraints (branch flow limits).
# (line constraint is actually on square of limit)
flow_max = array([(l.rate_a / base_mva)**2 for l in ln])
# FIXME: There must be a more elegant method for this.
for i, v in enumerate(flow_max):
if v == 0.0:
flow_max[i] = Inf
if self.flow_lim == IFLOW:
If = Yf * V
It = Yt * V
# Branch current limits.
h = r_[(If * conj(If)) - flow_max,
(If * conj(It)) - flow_max]
else:
i_fbus = [e.from_bus._i for e in ln]
i_tbus = [e.to_bus._i for e in ln]
# Complex power injected at "from" bus (p.u.).
Sf = V[i_fbus] * conj(Yf * V)
# Complex power injected at "to" bus (p.u.).
St = V[i_tbus] * conj(Yt * V)
if self.flow_lim == PFLOW: # active power limit, P (Pan Wei)
# Branch real power limits.
h = r_[Sf.real()**2 - flow_max,
St.real()**2 - flow_max]
elif self.flow_lim == SFLOW: # apparent power limit, |S|
# Branch apparent power limits.
h = r_[(Sf * conj(Sf)) - flow_max,
(St * conj(St)) - flow_max].real
else:
raise ValueError
return r_[g, h]
def dg_fcn(x, flag, usr_data=None):
""" Calculates the Jacobian matrix. It takes two arguments, the
first is the variable x and the second is a Boolean flag. If
the flag is true, the function returns a tuple of arrays
(row, col) to indicate the sparse structure of the Jacobian
matrix. If the flag is false the function returns the values
of the Jacobian matrix with length nnzj.
"""
iVa = range(Va.i1, Va.iN + 1)
iVm = range(Vm.i1, Vm.iN + 1)
iPg = range(Pg.i1, Pg.iN + 1)
iQg = range(Qg.i1, Qg.iN + 1)
iVaVmPgQg = r_[iVa, iVm, iPg, iQg].T
Vang = x[Va.i1:Va.iN + 1]
Vmag = x[Vm.i1:Vm.iN + 1]
V = Vmag * exp(1j * Vang)
# Compute partials of injected bus powers.
dSbus_dVm, dSbus_dVa = case.dSbus_dV(Ybus, V)
i_gbus = [gen.bus._i for gen in gn]
neg_Cg = csr_matrix((-ones(ng), (i_gbus, range(ng))), (nb, ng))
# Transposed Jacobian of the power balance equality constraints.
dg = lil_matrix((nxyz, 2 * nb))
blank = csr_matrix((nb, ng))
dg[iVaVmPgQg, :] = vstack([
hstack([dSbus_dVa.real, dSbus_dVm.real, neg_Cg, blank]),
hstack([dSbus_dVa.imag, dSbus_dVm.imag, blank, neg_Cg])
], "csr").T
# Compute partials of flows w.r.t V.
if self.flow_lim == IFLOW:
dFf_dVa, dFf_dVm, dFt_dVa, dFt_dVm, Ff, Ft = \
case.dIbr_dV(Yf, Yt, V)
else:
dFf_dVa, dFf_dVm, dFt_dVa, dFt_dVm, Ff, Ft = \
case.dSbr_dV(Yf, Yt, V, bs, ln)
if self.flow_lim == PFLOW:
dFf_dVa = dFf_dVa.real
dFf_dVm = dFf_dVm.real
dFt_dVa = dFt_dVa.real
dFt_dVm = dFt_dVm.real
Ff = Ff.real
Ft = Ft.real
# Squared magnitude of flow (complex power, current or real power).
df_dVa, df_dVm, dt_dVa, dt_dVm = \
case.dAbr_dV(dFf_dVa, dFf_dVm, dFt_dVa, dFt_dVm, Ff, Ft)
# Construct Jacobian of inequality constraints (branch limits) and
# transpose it.
dh = lil_matrix((nxyz, 2 * nl))
dh[r_[iVa, iVm].T, :] = vstack([hstack([df_dVa, df_dVm]),
hstack([dt_dVa, dt_dVm])], "csr").T
J = vstack([dg, dh, A]).tocoo()
if flag:
return (J.row, J.col)
else:
return J.data
def h_fcn(x, lagrange, obj_factor, flag, usr_data=None):
""" Evaluates the Hessian of the Lagrangian.
"""
neqnln = 2 * nb
niqnln = 2 * len(il) # no. of lines with constraints
Pgen = x[Pg.i1:Pg.iN + 1] # Active generation in p.u.
Qgen = x[Qg.i1:Qg.iN + 1] # Reactive generation in p.u.
for i, g in enumerate(gn):
g.p = Pgen[i] * base_mva # active generation in MW
g.q = Qgen[i] * base_mva # reactive generation in MVAr
Vang = x[Va.i1:Va.iN + 1]
Vmag = x[Vm.i1:Vm.iN + 1]
V = Vmag * exp(1j * Vang)
nxtra = nxyz - 2 * nb
#------------------------------------------------------------------
# Evaluate d2f.
#------------------------------------------------------------------
d2f_dPg2 = lil_matrix((ng, 1)) # w.r.t p.u. Pg
d2f_dQg2 = lil_matrix((ng, 1)) # w.r.t p.u. Qg]
for i in ipol:
d2f_dPg2[i, 0] = polyval(polyder(list(gn[i].p_cost), 2),
Pg.v0[i] * base_mva) * base_mva**2
# for i in ipol:
# d2f_dQg2[i] = polyval(polyder(list(gn[i].p_cost), 2),
# Qg.v0[i] * base_mva) * base_mva**2
i = r_[range(Pg.i1, Pg.iN + 1), range(Qg.i1, Qg.iN + 1)]
d2f = csr_matrix((vstack([d2f_dPg2, d2f_dQg2]).toarray().flatten(),
(i, i)), shape=(nxyz, nxyz))
# TODO: Generalised cost model.
d2f = d2f * self.opt["cost_mult"]
#------------------------------------------------------------------
# Evaluate Hessian of power balance constraints.
#------------------------------------------------------------------
eqnonlin = lagrange[:neqnln]
# nlam = len(lagrange["eqnonlin"]) / 2
nlam = len(eqnonlin) / 2
lamP = eqnonlin[:nlam]
lamQ = eqnonlin[nlam:nlam + nlam]
Gpaa, Gpav, Gpva, Gpvv = case.d2Sbus_dV2(Ybus, V, lamP)
Gqaa, Gqav, Gqva, Gqvv = case.d2Sbus_dV2(Ybus, V, lamQ)
d2G = vstack([
hstack([
vstack([hstack([Gpaa, Gpav]),
hstack([Gpva, Gpvv])]).real +
vstack([hstack([Gqaa, Gqav]),
hstack([Gqva, Gqvv])]).imag,
csr_matrix((2 * nb, nxtra))]),
hstack([
csr_matrix((nxtra, 2 * nb)),
csr_matrix((nxtra, nxtra))
])
], "csr")
#------------------------------------------------------------------
# Evaluate Hessian of flow constraints.
#------------------------------------------------------------------
ineqnonlin = lagrange[neqnln:neqnln + niqnln]
nmu = len(ineqnonlin) / 2
muF = ineqnonlin[:nmu]
muT = ineqnonlin[nmu:nmu + nmu]
if self.flow_lim == "I":
dIf_dVa, dIf_dVm, dIt_dVa, dIt_dVm, If, It = \
case.dIbr_dV(Yf, Yt, V)
Hfaa, Hfav, Hfva, Hfvv = \
case.d2AIbr_dV2(dIf_dVa, dIf_dVm, If, Yf, V, muF)
Htaa, Htav, Htva, Htvv = \
case.d2AIbr_dV2(dIt_dVa, dIt_dVm, It, Yt, V, muT)
else:
f = [e.from_bus._i for e in ln]
t = [e.to_bus._i for e in ln]
# Line-bus connection matrices.
Cf = csr_matrix((ones(nl), (range(nl), f)), (nl, nb))
Ct = csr_matrix((ones(nl), (range(nl), t)), (nl, nb))
dSf_dVa, dSf_dVm, dSt_dVa, dSt_dVm, Sf, St = \
case.dSbr_dV(Yf, Yt, V)
if self.flow_lim == PFLOW:
Hfaa, Hfav, Hfva, Hfvv = \
case.d2ASbr_dV2(dSf_dVa.real(), dSf_dVm.real(),
Sf.real(), Cf, Yf, V, muF)
Htaa, Htav, Htva, Htvv = \
case.d2ASbr_dV2(dSt_dVa.real(), dSt_dVm.real(),
St.real(), Ct, Yt, V, muT)
elif self.flow_lim == SFLOW:
Hfaa, Hfav, Hfva, Hfvv = \
case.d2ASbr_dV2(dSf_dVa, dSf_dVm, Sf, Cf, Yf, V, muF)
Htaa, Htav, Htva, Htvv = \
case.d2ASbr_dV2(dSt_dVa, dSt_dVm, St, Ct, Yt, V, muT)
else:
raise ValueError
d2H = vstack([
hstack([
vstack([hstack([Hfaa, Hfav]),
hstack([Hfva, Hfvv])]) +
vstack([hstack([Htaa, Htav]),
hstack([Htva, Htvv])]),
csr_matrix((2 * nb, nxtra))
]),
hstack([
csr_matrix((nxtra, 2 * nb)),
csr_matrix((nxtra, nxtra))
])
], "csr")
H = d2f + d2G + d2H
if flag:
return (H.row, H.col)
else:
return H.data
n = len(x0) # the number of variables
gl = r_[zeros(2 * nb), -Inf * ones(2 * nl2), l]
gu = r_[zeros(2 * nb), zeros(2 * nl2), u]
m = len(gl) # the number of constraints
nnzj = 0 # the number of nonzeros in Jacobian matrix
nnzh = 0 # the number of non-zeros in Hessian matrix
nlp = pyipopt.create(n, xmin, xmax, m, gl, gu, nnzj, nnzh,
f_fcn, df_fcn, g_fcn, dg_fcn, h_fcn)
# x, zl, zu, obj = nlp.solve(x0)
success = nlp.solve(x0)
nlp.close()
print "Success:", success
print "Solution of the primal variables, x"
# print x
print "Solution of the bound multipliers, z_L and z_U"
# print zl, zu
print "Objective value"