def add_affine_cuts(nlp_result, solve_data, config):
with time_code(solve_data.timing, "affine cut generation"):
m = solve_data.linear_GDP
if config.calc_disjunctive_bounds:
with time_code(solve_data.timing, "disjunctive variable bounding"):
TransformationFactory(
'contrib.compute_disj_var_bounds').apply_to(
m,
solver=config.mip_solver
if config.obbt_disjunctive_bounds else None)
config.logger.info("Adding affine cuts.")
GDPopt = m.GDPopt_utils
counter = 0
for var, val in zip(GDPopt.variable_list, nlp_result.var_values):
if val is not None and not var.fixed:
var.value = val
for constr in constraints_in_True_disjuncts(m, config):
# Note: this includes constraints that are deactivated in the current model (linear_GDP)
disjunctive_var_bounds = disjunctive_bounds(constr.parent_block())
if constr.body.polynomial_degree() in (1, 0):
continue
vars_in_constr = list(identify_variables(constr.body))
if any(var.value is None for var in vars_in_constr):
continue # a variable has no values
# mcpp stuff
mc_eqn = mc(constr.body, disjunctive_var_bounds)
# mc_eqn = mc(constr.body)
ccSlope = mc_eqn.subcc()
cvSlope = mc_eqn.subcv()
ccStart = mc_eqn.concave()
cvStart = mc_eqn.convex()
ub_int = min(
constr.upper,
mc_eqn.upper()) if constr.has_ub() else mc_eqn.upper()
lb_int = max(
constr.lower,
mc_eqn.lower()) if constr.has_lb() else mc_eqn.lower()
parent_block = constr.parent_block()
# Create a block on which to put outer approximation cuts.
aff_utils = parent_block.component('GDPopt_aff')
if aff_utils is None:
aff_utils = parent_block.GDPopt_aff = Block(
doc="Block holding affine constraints")
aff_utils.GDPopt_aff_cons = ConstraintList()
aff_cuts = aff_utils.GDPopt_aff_cons
concave_cut = sum(ccSlope[var] * (var - var.value)
for var in vars_in_constr) + ccStart >= lb_int
convex_cut = sum(cvSlope[var] * (var - var.value)
for var in vars_in_constr) + cvStart <= ub_int
aff_cuts.add(expr=concave_cut)
aff_cuts.add(expr=convex_cut)
counter += 2
config.logger.info("Added %s affine cuts" % counter)
def add_affine_cuts(nlp_result, solve_data, config):
with time_code(solve_data.timing, "affine cut generation"):
m = solve_data.linear_GDP
if config.calc_disjunctive_bounds:
with time_code(solve_data.timing, "disjunctive variable bounding"):
TransformationFactory('contrib.compute_disj_var_bounds').apply_to(
m,
solver=config.mip_solver if config.obbt_disjunctive_bounds else None
)
config.logger.info("Adding affine cuts.")
GDPopt = m.GDPopt_utils
counter = 0
for var, val in zip(GDPopt.variable_list, nlp_result.var_values):
if val is not None and not var.fixed:
var.value = val
for constr in constraints_in_True_disjuncts(m, config):
# Note: this includes constraints that are deactivated in the current model (linear_GDP)
disjunctive_var_bounds = disjunctive_bounds(constr.parent_block())
if constr.body.polynomial_degree() in (1, 0):
continue
vars_in_constr = list(
identify_variables(constr.body))
if any(var.value is None for var in vars_in_constr):
continue # a variable has no values
# mcpp stuff
mc_eqn = mc(constr.body, disjunctive_var_bounds)
# mc_eqn = mc(constr.body)
ccSlope = mc_eqn.subcc()
cvSlope = mc_eqn.subcv()
ccStart = mc_eqn.concave()
cvStart = mc_eqn.convex()
ub_int = min(constr.upper, mc_eqn.upper()) if constr.has_ub() else mc_eqn.upper()
lb_int = max(constr.lower, mc_eqn.lower()) if constr.has_lb() else mc_eqn.lower()
parent_block = constr.parent_block()
# Create a block on which to put outer approximation cuts.
aff_utils = parent_block.component('GDPopt_aff')
if aff_utils is None:
aff_utils = parent_block.GDPopt_aff = Block(
doc="Block holding affine constraints")
aff_utils.GDPopt_aff_cons = ConstraintList()
aff_cuts = aff_utils.GDPopt_aff_cons
concave_cut = sum(ccSlope[var] * (var - var.value)
for var in vars_in_constr
) + ccStart >= lb_int
convex_cut = sum(cvSlope[var] * (var - var.value)
for var in vars_in_constr
) + cvStart <= ub_int
aff_cuts.add(expr=concave_cut)
aff_cuts.add(expr=convex_cut)
counter += 2
config.logger.info("Added %s affine cuts" % counter)
def solve_feasibility_subproblem(solve_data, config):
"""Solves a feasibility NLP if the fixed_nlp problem is infeasible.
Args:
solve_data (MindtPySolveData): data container that holds solve-instance data.
config (ConfigBlock): the specific configurations for MindtPy.
Returns:
feas_subproblem (Pyomo model): feasibility NLP from the model.
feas_soln (SolverResults): results from solving the feasibility NLP.
"""
feas_subproblem = solve_data.working_model.clone()
add_feas_slacks(feas_subproblem, config)
MindtPy = feas_subproblem.MindtPy_utils
if MindtPy.find_component('objective_value') is not None:
MindtPy.objective_value.value = 0
next(feas_subproblem.component_data_objects(
Objective, active=True)).deactivate()
for constr in feas_subproblem.MindtPy_utils.nonlinear_constraint_list:
constr.deactivate()
MindtPy.feas_opt.activate()
if config.feasibility_norm == 'L1':
MindtPy.feas_obj = Objective(
expr=sum(s for s in MindtPy.feas_opt.slack_var[...]),
sense=minimize)
elif config.feasibility_norm == 'L2':
MindtPy.feas_obj = Objective(
expr=sum(s*s for s in MindtPy.feas_opt.slack_var[...]),
sense=minimize)
else:
MindtPy.feas_obj = Objective(
expr=MindtPy.feas_opt.slack_var,
sense=minimize)
TransformationFactory('core.fix_integer_vars').apply_to(feas_subproblem)
nlpopt = SolverFactory(config.nlp_solver)
nlp_args = dict(config.nlp_solver_args)
set_solver_options(nlpopt, solve_data, config, solver_type='nlp')
with SuppressInfeasibleWarning():
try:
with time_code(solve_data.timing, 'feasibility subproblem'):
feas_soln = nlpopt.solve(
feas_subproblem, tee=config.nlp_solver_tee, **nlp_args)
except (ValueError, OverflowError) as error:
for nlp_var, orig_val in zip(
MindtPy.variable_list,
solve_data.initial_var_values):
if not nlp_var.fixed and not nlp_var.is_binary():
nlp_var.set_value(orig_val, skip_validation=True)
with time_code(solve_data.timing, 'feasibility subproblem'):
feas_soln = nlpopt.solve(
feas_subproblem, tee=config.nlp_solver_tee, **nlp_args)
handle_feasibility_subproblem_tc(
feas_soln.solver.termination_condition, MindtPy, solve_data, config)
return feas_subproblem, feas_soln
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
This function performs all of the GDPopt solver setup and problem
validation. It then calls upon helper functions to construct the
initial master approximation and iteration loop.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}), preserve_implicit=True)
config.set_value(kwds)
if config.strategy is None:
msg = 'Please specify solution strategy. Options are: \n'
msg += ' LOA: Logic-based Outer Approximation\n'
msg += ' GLOA: Global Logic-based Outer Approximation\n'
msg += ' LBB: Logic-based Branch and Bound\n'
msg += ' RIC: Relaxation with Integer Cuts'
raise ValueError(msg)
with setup_solver_environment(model, config) as solve_data:
self._log_solver_intro_message(config)
solve_data.results.solver.name = 'GDPopt %s - %s' % (str(
self.version()), config.strategy)
# Verify that objective has correct form
process_objective(solve_data, config)
# Presolve LP or NLP problems using subsolvers
presolved, presolve_results = presolve_lp_nlp(solve_data, config)
if presolved:
# TODO merge the solver results
return presolve_results # problem presolved
if solve_data.active_strategy in {'LOA', 'GLOA', 'RIC'}:
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
GDPopt_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
GDPopt_iteration_loop(solve_data, config)
elif solve_data.active_strategy == 'LBB':
_perform_branch_and_bound(solve_data)
else:
raise ValueError('Unrecognized strategy: ' + config.strategy)
return solve_data.results
def GDPopt_iteration_loop(solve_data, config):
"""Algorithm main loop.
Returns True if successful convergence is obtained. False otherwise.
"""
while solve_data.master_iteration < config.iterlim:
# Set iteration counters for new master iteration.
solve_data.master_iteration += 1
solve_data.mip_iteration = 0
solve_data.nlp_iteration = 0
# print line for visual display
config.logger.info(
'---GDPopt Master Iteration %s---'
% solve_data.master_iteration)
# solve linear master problem
with time_code(solve_data.timing, 'mip'):
mip_result = solve_LOA_master(solve_data, config)
# Check termination conditions
if algorithm_should_terminate(solve_data, config):
break
# Solve NLP subproblem
if solve_data.active_strategy == 'LOA':
with time_code(solve_data.timing, 'nlp'):
nlp_result = solve_local_subproblem(mip_result, solve_data, config)
if nlp_result.feasible:
add_outer_approximation_cuts(nlp_result, solve_data, config)
elif solve_data.active_strategy == 'GLOA':
with time_code(solve_data.timing, 'nlp'):
nlp_result = solve_global_subproblem(mip_result, solve_data, config)
if nlp_result.feasible:
add_affine_cuts(nlp_result, solve_data, config)
elif solve_data.active_strategy == 'RIC':
with time_code(solve_data.timing, 'nlp'):
nlp_result = solve_local_subproblem(mip_result, solve_data, config)
else:
raise ValueError('Unrecognized strategy: ' + solve_data.active_strategy)
# Add integer cut
add_integer_cut(
mip_result.var_values, solve_data.linear_GDP, solve_data, config,
feasible=nlp_result.feasible)
# Check termination conditions
if algorithm_should_terminate(solve_data, config):
break
def add_lazy_no_good_cuts(self,
var_values,
solve_data,
config,
opt,
feasible=False):
"""Adds no-good cuts.
Add the no-good cuts through Cplex inherent function self.add().
Args:
var_values (list): values of the current variables, used to generate the cut.
solve_data (MindtPySolveData): data container that holds solve-instance data.
config (ConfigBlock): the specific configurations for MindtPy.
opt (SolverFactory): cplex_persistent.
feasible (bool, optional): whether the integer combination yields a feasible or infeasible NLP. Defaults to False.
Raises:
ValueError: binary variable is not 0 or 1
"""
if not config.add_no_good_cuts:
return
config.logger.info('Adding no-good cuts')
with time_code(solve_data.timing, 'No-good cut generation'):
m = solve_data.mip
MindtPy = m.MindtPy_utils
int_tol = config.integer_tolerance
binary_vars = [v for v in MindtPy.variable_list if v.is_binary()]
# copy variable values over
for var, val in zip(MindtPy.variable_list, var_values):
if not var.is_binary():
continue
var.set_value(val, skip_validation=True)
# check to make sure that binary variables are all 0 or 1
for v in binary_vars:
if value(abs(v - 1)) > int_tol and value(abs(v)) > int_tol:
raise ValueError('Binary {} = {} is not 0 or 1'.format(
v.name, value(v)))
if not binary_vars: # if no binary variables, skip
return
pyomo_no_good_cut = sum(
1 - v
for v in binary_vars if value(abs(v - 1)) <= int_tol) + sum(
v for v in binary_vars if value(abs(v)) <= int_tol)
cplex_no_good_rhs = generate_standard_repn(
pyomo_no_good_cut).constant
cplex_no_good_cut, _ = opt._get_expr_from_pyomo_expr(
pyomo_no_good_cut)
self.add(constraint=cplex.SparsePair(
ind=cplex_no_good_cut.variables,
val=cplex_no_good_cut.coefficients),
sense='G',
rhs=1 - cplex_no_good_rhs)
def GDPopt_iteration_loop(solve_data, config):
"""Algorithm main loop.
Returns True if successful convergence is obtained. False otherwise.
"""
while solve_data.master_iteration < config.iterlim:
# Set iteration counters for new master iteration.
solve_data.master_iteration += 1
solve_data.mip_iteration = 0
solve_data.nlp_iteration = 0
# print line for visual display
config.logger.info(
'---GDPopt Master Iteration %s---'
% solve_data.master_iteration)
# solve linear master problem
with time_code(solve_data.timing, 'mip'):
mip_result = solve_LOA_master(solve_data, config)
# Check termination conditions
if algorithm_should_terminate(solve_data, config):
break
# Solve NLP subproblem
if solve_data.current_strategy == 'LOA':
with time_code(solve_data.timing, 'nlp'):
nlp_result = solve_local_subproblem(mip_result, solve_data, config)
if nlp_result.feasible:
add_outer_approximation_cuts(nlp_result, solve_data, config)
elif solve_data.current_strategy == 'GLOA':
with time_code(solve_data.timing, 'nlp'):
nlp_result = solve_global_subproblem(mip_result, solve_data, config)
if nlp_result.feasible:
add_affine_cuts(nlp_result, solve_data, config)
# Add integer cut
add_integer_cut(
mip_result.var_values, solve_data.linear_GDP, solve_data, config,
feasible=nlp_result.feasible)
# Check termination conditions
if algorithm_should_terminate(solve_data, config):
break
def add_integer_cut(var_values, target_model, solve_data, config, feasible=False):
"""Add an integer cut to the target GDP model."""
with time_code(solve_data.timing, 'integer cut generation'):
m = target_model
GDPopt = m.GDPopt_utils
var_value_is_one = ComponentSet()
var_value_is_zero = ComponentSet()
for var, val in zip(GDPopt.variable_list, var_values):
if not var.is_binary():
continue
if var.fixed:
if val is not None and var.value != val:
# val needs to be None or match var.value. Otherwise, we have a
# contradiction.
raise ValueError(
"Fixed variable %s has value %s != "
"provided value of %s." % (var.name, var.value, val))
val = var.value
if not config.force_subproblem_nlp:
# Skip indicator variables
# TODO we should implement this as a check among Disjuncts instead
if not (var.local_name == 'indicator_var' and var.parent_block().type() == Disjunct):
continue
if fabs(val - 1) <= config.integer_tolerance:
var_value_is_one.add(var)
elif fabs(val) <= config.integer_tolerance:
var_value_is_zero.add(var)
else:
raise ValueError(
'Binary %s = %s is not 0 or 1' % (var.name, val))
if not (var_value_is_one or var_value_is_zero):
# if no remaining binary variables, then terminate algorithm.
config.logger.info(
'Adding integer cut to a model without discrete variables. '
'Model is now infeasible.')
if solve_data.objective_sense == minimize:
solve_data.LB = float('inf')
else:
solve_data.UB = float('-inf')
return False
int_cut = (sum(1 - v for v in var_value_is_one) +
sum(v for v in var_value_is_zero)) >= 1
if not feasible:
config.logger.info('Adding integer cut')
GDPopt.integer_cuts.add(expr=int_cut)
else:
backtracking_enabled = (
"disabled" if GDPopt.no_backtracking.active else "allowed")
config.logger.info(
'Registering explored configuration. '
'Backtracking is currently %s.' % backtracking_enabled)
GDPopt.no_backtracking.add(expr=int_cut)
def add_integer_cut(var_values, target_model, solve_data, config, feasible=False):
"""Add an integer cut to the target GDP model."""
with time_code(solve_data.timing, 'integer cut generation'):
m = target_model
GDPopt = m.GDPopt_utils
var_value_is_one = ComponentSet()
var_value_is_zero = ComponentSet()
for var, val in zip(GDPopt.variable_list, var_values):
if not var.is_binary():
continue
if var.fixed:
if val is not None and var.value != val:
# val needs to be None or match var.value. Otherwise, we have a
# contradiction.
raise ValueError(
"Fixed variable %s has value %s != "
"provided value of %s." % (var.name, var.value, val))
val = var.value
if not config.force_subproblem_nlp:
# Skip indicator variables
# TODO we should implement this as a check among Disjuncts instead
if not (var.local_name == 'indicator_var' and var.parent_block().type() == Disjunct):
continue
if fabs(val - 1) <= config.integer_tolerance:
var_value_is_one.add(var)
elif fabs(val) <= config.integer_tolerance:
var_value_is_zero.add(var)
else:
raise ValueError(
'Binary %s = %s is not 0 or 1' % (var.name, val))
if not (var_value_is_one or var_value_is_zero):
# if no remaining binary variables, then terminate algorithm.
config.logger.info(
'Adding integer cut to a model without discrete variables. '
'Model is now infeasible.')
if solve_data.objective_sense == minimize:
solve_data.LB = float('inf')
else:
solve_data.UB = float('-inf')
return False
int_cut = (sum(1 - v for v in var_value_is_one) +
sum(v for v in var_value_is_zero)) >= 1
if not feasible:
config.logger.info('Adding integer cut')
GDPopt.integer_cuts.add(expr=int_cut)
else:
backtracking_enabled = (
"disabled" if GDPopt.no_backtracking.active else "allowed")
config.logger.info(
'Registering explored configuration. '
'Backtracking is currently %s.' % backtracking_enabled)
GDPopt.no_backtracking.add(expr=int_cut)
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
This function performs all of the GDPopt solver setup and problem
validation. It then calls upon helper functions to construct the
initial master approximation and iteration loop.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
with setup_solver_environment(model, config) as solve_data:
self._log_solver_intro_message(config)
solve_data.results.solver.name = 'GDPopt %s - %s' % (str(
self.version()), config.strategy)
# Verify that objective has correct form
process_objective(solve_data, config)
# Presolve LP or NLP problems using subsolvers
presolved, presolve_results = presolve_lp_nlp(solve_data, config)
if presolved:
# TODO merge the solver results
return presolve_results # problem presolved
if solve_data.active_strategy in {'LOA', 'GLOA'}:
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
GDPopt_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
GDPopt_iteration_loop(solve_data, config)
elif solve_data.active_strategy == 'LBB':
_perform_branch_and_bound(solve_data)
return solve_data.results
def add_no_good_cuts(var_values, solve_data, config):
"""Adds no-good cuts.
This adds an no-good cuts to the no_good_cuts ConstraintList, which is not activated by default.
However, it may be activated as needed in certain situations or for certain values of option flags.
Parameters
----------
var_values : list
Variable values of the current solution, used to generate the cut.
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
Raises
------
ValueError
The value of binary variable is not 0 or 1.
"""
if not config.add_no_good_cuts:
return
with time_code(solve_data.timing, 'no_good cut generation'):
config.logger.debug('Adding no-good cuts')
m = solve_data.mip
MindtPy = m.MindtPy_utils
int_tol = config.integer_tolerance
binary_vars = [v for v in MindtPy.variable_list if v.is_binary()]
# copy variable values over
for var, val in zip(MindtPy.variable_list, var_values):
if not var.is_binary():
continue
var.set_value(val, skip_validation=True)
# check to make sure that binary variables are all 0 or 1
for v in binary_vars:
if value(abs(v - 1)) > int_tol and value(abs(v)) > int_tol:
raise ValueError(
'Binary {} = {} is not 0 or 1'.format(v.name, value(v)))
if not binary_vars: # if no binary variables, skip
return
int_cut = (sum(1 - v for v in binary_vars
if value(abs(v - 1)) <= int_tol) +
sum(v for v in binary_vars
if value(abs(v)) <= int_tol) >= 1)
MindtPy.cuts.no_good_cuts.add(expr=int_cut)
def add_no_good_cuts(var_values, solve_data, config, feasible=False):
"""
Adds no-good cuts; modifies the model to include no-good cuts
This adds an no-good cuts to the no_good_cuts ConstraintList, which is not activated by default.
However, it may be activated as needed in certain situations or for certain values of option flags.
Parameters
----------
var_values: list
values of the current variables, used to generate the cut
solve_data: MindtPy Data Container
data container that holds solve-instance data
config: ConfigBlock
contains the specific configurations for the algorithm
feasible: bool, optional
boolean indicating if integer combination yields a feasible or infeasible NLP
"""
if not config.add_no_good_cuts:
return
with time_code(solve_data.timing, 'no_good cut generation'):
config.logger.info('Adding no_good cuts')
m = solve_data.mip
MindtPy = m.MindtPy_utils
int_tol = config.integer_tolerance
binary_vars = [v for v in MindtPy.variable_list if v.is_binary()]
# copy variable values over
for var, val in zip(MindtPy.variable_list, var_values):
if not var.is_binary():
continue
var.value = val
# check to make sure that binary variables are all 0 or 1
for v in binary_vars:
if value(abs(v - 1)) > int_tol and value(abs(v)) > int_tol:
raise ValueError('Binary {} = {} is not 0 or 1'.format(
v.name, value(v)))
if not binary_vars: # if no binary variables, skip
return
int_cut = (sum(1 - v
for v in binary_vars if value(abs(v - 1)) <= int_tol) +
sum(v
for v in binary_vars if value(abs(v)) <= int_tol) >= 1)
MindtPy.cuts.no_good_cuts.add(expr=int_cut)
def test_solve_linear_GDP_unbounded(self):
m = ConcreteModel()
m.GDPopt_utils = Block()
m.x = Var(bounds=(-1, 10))
m.y = Var(bounds=(2, 3))
m.z = Var()
m.d = Disjunction(expr=[
[m.x + m.y >= 5], [m.x - m.y <= 3]
])
m.o = Objective(expr=m.z)
m.GDPopt_utils.variable_list = [m.x, m.y, m.z]
m.GDPopt_utils.disjunct_list = [m.d._autodisjuncts[0], m.d._autodisjuncts[1]]
output = StringIO()
with LoggingIntercept(output, 'pyomo.contrib.gdpopt', logging.WARNING):
solver_data = GDPoptSolveData()
solver_data.timing = Bunch()
with time_code(solver_data.timing, 'main', is_main_timer=True):
solve_linear_GDP(m, solver_data, GDPoptSolver.CONFIG(dict(mip_solver=mip_solver, strategy='LOA')))
self.assertIn("Linear GDP was unbounded. Resolving with arbitrary bound values",
output.getvalue().strip())
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
This function performs all of the GDPopt solver setup and problem
validation. It then calls upon helper functions to construct the
initial master approximation and iteration loop.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
solve_data = GDPoptSolveData()
solve_data.results = SolverResults()
solve_data.timing = Container()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total', is_main_timer=True), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPopt_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPopt version %s using %s algorithm"
% (".".join(map(str, self.version())), config.strategy)
)
config.logger.info(
"""
If you use this software, you may cite the following:
- Implementation:
Chen, Q; Johnson, ES; Siirola, JD; Grossmann, IE.
Pyomo.GDP: Disjunctive Models in Python.
Proc. of the 13th Intl. Symposium on Process Systems Eng.
San Diego, 2018.
- LOA algorithm:
Türkay, M; Grossmann, IE.
Logic-based MINLP algorithms for the optimal synthesis of process networks.
Comp. and Chem. Eng. 1996, 20(8), 959–978.
DOI: 10.1016/0098-1354(95)00219-7.
- GLOA algorithm:
Lee, S; Grossmann, IE.
A Global Optimization Algorithm for Nonconvex Generalized Disjunctive Programming and Applications to Process Systems
Comp. and Chem. Eng. 2001, 25, 1675-1697.
DOI: 10.1016/S0098-1354(01)00732-3
""".strip()
)
solve_data.results.solver.name = 'GDPopt %s - %s' % (
str(self.version()), config.strategy)
solve_data.original_model = model
solve_data.working_model = model.clone()
GDPopt = solve_data.working_model.GDPopt_utils
setup_results_object(solve_data, config)
solve_data.current_strategy = config.strategy
# Verify that objective has correct form
process_objective(solve_data, config)
# Save model initial values. These are used later to initialize NLP
# subproblems.
solve_data.initial_var_values = list(
v.value for v in GDPopt.variable_list)
solve_data.best_solution_found = None
# Validate the model to ensure that GDPopt is able to solve it.
if not model_is_valid(solve_data, config):
return
# Integer cuts exclude particular discrete decisions
GDPopt.integer_cuts = ConstraintList(doc='integer cuts')
# Feasible integer cuts exclude discrete realizations that have
# been explored via an NLP subproblem. Depending on model
# characteristics, the user may wish to revisit NLP subproblems
# (with a different initialization, for example). Therefore, these
# cuts are not enabled by default, unless the initial model has no
# discrete decisions.
# Note: these cuts will only exclude integer realizations that are
# not already in the primary GDPopt_integer_cuts ConstraintList.
GDPopt.no_backtracking = ConstraintList(
doc='explored integer cuts')
# Set up iteration counters
solve_data.master_iteration = 0
solve_data.mip_iteration = 0
solve_data.nlp_iteration = 0
# set up bounds
solve_data.LB = float('-inf')
solve_data.UB = float('inf')
solve_data.iteration_log = {}
# Flag indicating whether the solution improved in the past
# iteration or not
solve_data.feasible_solution_improved = False
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
GDPopt_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
GDPopt_iteration_loop(solve_data, config)
if solve_data.best_solution_found is not None:
# Update values in working model
copy_var_list_values(
from_list=solve_data.best_solution_found.GDPopt_utils.variable_list,
to_list=GDPopt.variable_list,
config=config)
# Update values in original model
copy_var_list_values(
GDPopt.variable_list,
solve_data.original_model.GDPopt_utils.variable_list,
config)
solve_data.results.problem.lower_bound = solve_data.LB
solve_data.results.problem.upper_bound = solve_data.UB
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.user_time = solve_data.timing.total
solve_data.results.solver.wallclock_time = solve_data.timing.total
solve_data.results.solver.iterations = solve_data.master_iteration
return solve_data.results
def add_ecp_cuts(target_model, solve_data, config,
linearize_active=True,
linearize_violated=True):
"""Linearizes nonlinear constraints. Adds the cuts for the ECP method.
Parameters
----------
target_model : Pyomo model
The relaxed linear model.
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
linearize_active : bool, optional
Whether to linearize the active nonlinear constraints, by default True.
linearize_violated : bool, optional
Whether to linearize the violated nonlinear constraints, by default True.
"""
with time_code(solve_data.timing, 'ECP cut generation'):
for constr in target_model.MindtPy_utils.nonlinear_constraint_list:
constr_vars = list(identify_variables(constr.body))
jacs = solve_data.jacobians
if constr.has_lb() and constr.has_ub():
config.logger.warning(
'constraint {} has both a lower '
'and upper bound.'
'\n'.format(
constr))
continue
if constr.has_ub():
try:
upper_slack = constr.uslack()
except (ValueError, OverflowError):
config.logger.warning(
'constraint {} has caused either a '
'ValueError or OverflowError.'
'\n'.format(
constr))
continue
if (linearize_active and abs(upper_slack) < config.ecp_tolerance) \
or (linearize_violated and upper_slack < 0) \
or (config.linearize_inactive and upper_slack > 0):
if config.add_slack:
slack_var = target_model.MindtPy_utils.cuts.slack_vars.add()
target_model.MindtPy_utils.cuts.ecp_cuts.add(
expr=(sum(value(jacs[constr][var])*(var - var.value)
for var in constr_vars)
- (slack_var if config.add_slack else 0)
<= upper_slack)
)
if constr.has_lb():
try:
lower_slack = constr.lslack()
except (ValueError, OverflowError):
config.logger.warning(
'constraint {} has caused either a '
'ValueError or OverflowError.'
'\n'.format(
constr))
continue
if (linearize_active and abs(lower_slack) < config.ecp_tolerance) \
or (linearize_violated and lower_slack < 0) \
or (config.linearize_inactive and lower_slack > 0):
if config.add_slack:
slack_var = target_model.MindtPy_utils.cuts.slack_vars.add()
target_model.MindtPy_utils.cuts.ecp_cuts.add(
expr=(sum(value(jacs[constr][var])*(var - var.value)
for var in constr_vars)
+ (slack_var if config.add_slack else 0)
>= -lower_slack)
)
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
Warning: at this point in time, if you try to use PSC or GBD with
anything other than IPOPT as the NLP solver, bad things will happen.
This is because the suffixes are not in place to extract dual values
from the variable bounds for any other solver.
TODO: fix needed with the GBD implementation.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
solve_data = MindtPySolveData()
solve_data.results = SolverResults()
solve_data.timing = Container()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total'), \
restore_logger_level(config.logger), \
create_utility_block(model, 'MindtPy_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info("---Starting MindtPy---")
solve_data.original_model = model
solve_data.working_model = model.clone()
MindtPy = solve_data.working_model.MindtPy_utils
setup_results_object(solve_data, config)
process_objective(solve_data, config)
# Save model initial values.
solve_data.initial_var_values = list(
v.value for v in MindtPy.variable_list)
# Store the initial model state as the best solution found. If we
# find no better solution, then we will restore from this copy.
solve_data.best_solution_found = None
# Record solver name
solve_data.results.solver.name = 'MindtPy' + str(config.strategy)
# Validate the model to ensure that MindtPy is able to solve it.
if not model_is_valid(solve_data, config):
return
# Create a model block in which to store the generated feasibility
# slack constraints. Do not leave the constraints on by default.
feas = MindtPy.MindtPy_feas = Block()
feas.deactivate()
feas.feas_constraints = ConstraintList(
doc='Feasibility Problem Constraints')
# Create a model block in which to store the generated linear
# constraints. Do not leave the constraints on by default.
lin = MindtPy.MindtPy_linear_cuts = Block()
lin.deactivate()
# Integer cuts exclude particular discrete decisions
lin.integer_cuts = ConstraintList(doc='integer cuts')
# Feasible integer cuts exclude discrete realizations that have
# been explored via an NLP subproblem. Depending on model
# characteristics, the user may wish to revisit NLP subproblems
# (with a different initialization, for example). Therefore, these
# cuts are not enabled by default.
#
# Note: these cuts will only exclude integer realizations that are
# not already in the primary integer_cuts ConstraintList.
lin.feasible_integer_cuts = ConstraintList(
doc='explored integer cuts')
lin.feasible_integer_cuts.deactivate()
# Set up iteration counters
solve_data.nlp_iter = 0
solve_data.mip_iter = 0
solve_data.mip_subiter = 0
# set up bounds
solve_data.LB = float('-inf')
solve_data.UB = float('inf')
solve_data.LB_progress = [solve_data.LB]
solve_data.UB_progress = [solve_data.UB]
# Set of NLP iterations for which cuts were generated
lin.nlp_iters = Set(dimen=1)
# Set of MIP iterations for which cuts were generated in ECP
lin.mip_iters = Set(dimen=1)
nonlinear_constraints = [c for c in MindtPy.constraint_list if
c.body.polynomial_degree() not in (1, 0)]
lin.nl_constraint_set = RangeSet(
len(nonlinear_constraints),
doc="Integer index set over the nonlinear constraints")
feas.constraint_set = RangeSet(
len(MindtPy.constraint_list),
doc="integer index set over the constraints")
# # Mapping Constraint -> integer index
# MindtPy.feas_map = {}
# # Mapping integer index -> Constraint
# MindtPy.feas_inverse_map = {}
# # Generate the two maps. These maps may be helpful for later
# # interpreting indices on the slack variables or generated cuts.
# for c, n in zip(MindtPy.constraint_list, feas.constraint_set):
# MindtPy.feas_map[c] = n
# MindtPy.feas_inverse_map[n] = c
# Create slack variables for OA cuts
lin.slack_vars = VarList(bounds=(0, config.max_slack), initialize=0, domain=NonNegativeReals)
# Create slack variables for feasibility problem
feas.slack_var = Var(feas.constraint_set,
domain=NonNegativeReals, initialize=1)
# Flag indicating whether the solution improved in the past
# iteration or not
solve_data.solution_improved = False
if not hasattr(solve_data.working_model, 'ipopt_zL_out'):
solve_data.working_model.ipopt_zL_out = Suffix(
direction=Suffix.IMPORT)
if not hasattr(solve_data.working_model, 'ipopt_zU_out'):
solve_data.working_model.ipopt_zU_out = Suffix(
direction=Suffix.IMPORT)
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
MindtPy_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
MindtPy_iteration_loop(solve_data, config)
if solve_data.best_solution_found is not None:
# Update values in original model
copy_var_list_values(
from_list=solve_data.best_solution_found.MindtPy_utils.variable_list,
to_list=MindtPy.variable_list,
config=config)
# MindtPy.objective_value.set_value(
# value(solve_data.working_objective_expr, exception=False))
copy_var_list_values(
MindtPy.variable_list,
solve_data.original_model.MindtPy_utils.variable_list,
config)
solve_data.results.problem.lower_bound = solve_data.LB
solve_data.results.problem.upper_bound = solve_data.UB
def solve_main(solve_data, config, fp=False, regularization_problem=False):
"""
This function solves the MIP main problem
Parameters
----------
solve_data: MindtPy Data Container
data container that holds solve-instance data
config: ConfigBlock
contains the specific configurations for the algorithm
Returns
-------
solve_data.mip: Pyomo model
the MIP stored in solve_data
main_mip_results: Pyomo results object
result from solving the main MIP
fp: Bool
generate the feasibility pump regularization main problem
regularization_problem: Bool
generate the ROA regularization main problem
"""
if fp:
config.logger.info('FP-MIP %s: Solve main problem.' %
(solve_data.fp_iter,))
elif regularization_problem:
config.logger.info('Regularization-MIP %s: Solve main regularization problem.' %
(solve_data.mip_iter,))
else:
solve_data.mip_iter += 1
config.logger.info('MIP %s: Solve main problem.' %
(solve_data.mip_iter,))
# setup main problem
setup_main(solve_data, config, fp, regularization_problem)
mainopt = setup_mip_solver(solve_data, config, regularization_problem)
mip_args = dict(config.mip_solver_args)
if config.mip_solver in {'cplex', 'cplex_persistent', 'gurobi', 'gurobi_persistent'}:
mip_args['warmstart'] = True
set_solver_options(mainopt, solve_data, config,
solver_type='mip', regularization=regularization_problem)
try:
with time_code(solve_data.timing, 'regularization main' if regularization_problem else ('fp main' if fp else 'main')):
main_mip_results = mainopt.solve(solve_data.mip,
tee=config.mip_solver_tee, **mip_args)
except (ValueError, AttributeError):
if config.single_tree:
config.logger.warning('Single tree terminate.')
if get_main_elapsed_time(solve_data.timing) >= config.time_limit - 2:
config.logger.warning('due to the timelimit.')
solve_data.results.solver.termination_condition = tc.maxTimeLimit
if config.strategy == 'GOA' or config.add_no_good_cuts:
config.logger.warning('ValueError: Cannot load a SolverResults object with bad status: error. '
'MIP solver failed. This usually happens in the single-tree GOA algorithm. '
"No-good cuts are added and GOA algorithm doesn't converge within the time limit. "
'No integer solution is found, so the cplex solver will report an error status. ')
return None, None
if main_mip_results.solver.termination_condition is tc.optimal:
if config.single_tree and not config.add_no_good_cuts and not regularization_problem:
if solve_data.objective_sense == minimize:
solve_data.LB = max(
main_mip_results.problem.lower_bound, solve_data.LB)
solve_data.bound_improved = solve_data.LB > solve_data.LB_progress[-1]
solve_data.LB_progress.append(solve_data.LB)
else:
solve_data.UB = min(
main_mip_results.problem.upper_bound, solve_data.UB)
solve_data.bound_improved = solve_data.UB < solve_data.UB_progress[-1]
solve_data.UB_progress.append(solve_data.UB)
elif main_mip_results.solver.termination_condition is tc.infeasibleOrUnbounded:
# Linear solvers will sometimes tell me that it's infeasible or
# unbounded during presolve, but fails to distinguish. We need to
# resolve with a solver option flag on.
main_mip_results, _ = distinguish_mip_infeasible_or_unbounded(
solve_data.mip, config)
return solve_data.mip, main_mip_results
if regularization_problem:
solve_data.mip.MindtPy_utils.objective_constr.deactivate()
solve_data.mip.MindtPy_utils.del_component('loa_proj_mip_obj')
solve_data.mip.MindtPy_utils.cuts.del_component('obj_reg_estimate')
if config.add_regularization == 'level_L1':
solve_data.mip.MindtPy_utils.del_component('L1_obj')
elif config.add_regularization == 'level_L_infinity':
solve_data.mip.MindtPy_utils.del_component(
'L_infinity_obj')
return solve_data.mip, main_mip_results
def solve(self, model, **kwds):
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
return SolverFactory('gdpopt').solve(
model,
strategy='LBB',
minlp_solver=config.solver,
minlp_solver_args=config.solver_args,
tee=config.tee,
check_sat=config.check_sat,
logger=config.logger,
time_limit=config.time_limit)
# Validate model to be used with gdpbb
self.validate_model(model)
# Set solver as an MINLP
solve_data = GDPbbSolveData()
solve_data.timing = Container()
solve_data.original_model = model
solve_data.results = SolverResults()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total', is_main_timer=True), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPbb_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPbb version %s using %s as subsolver" %
(".".join(map(str, self.version())), config.solver))
# Setup results
solve_data.results.solver.name = 'GDPbb - %s' % (str(
config.solver))
setup_results_object(solve_data, config)
# clone original model for root node of branch and bound
root = solve_data.working_model = solve_data.original_model.clone()
# get objective sense
process_objective(solve_data, config)
objectives = solve_data.original_model.component_data_objects(
Objective, active=True)
obj = next(objectives, None)
solve_data.results.problem.sense = obj.sense
# set up lists to keep track of which disjunctions have been covered.
# this list keeps track of the relaxed disjunctions
root.GDPbb_utils.unenforced_disjunctions = list(
disjunction
for disjunction in root.GDPbb_utils.disjunction_list
if disjunction.active)
root.GDPbb_utils.deactivated_constraints = ComponentSet([
constr
for disjunction in root.GDPbb_utils.unenforced_disjunctions
for disjunct in disjunction.disjuncts
for constr in disjunct.component_data_objects(ctype=Constraint,
active=True)
if constr.body.polynomial_degree() not in (1, 0)
])
# Deactivate nonlinear constraints in unenforced disjunctions
for constr in root.GDPbb_utils.deactivated_constraints:
constr.deactivate()
# Add the BigM suffix if it does not already exist. Used later during nonlinear constraint activation.
if not hasattr(root, 'BigM'):
root.BigM = Suffix()
# Pre-screen that none of the disjunctions are already predetermined due to the disjuncts being fixed
# to True/False values.
# TODO this should also be done within the loop, but we aren't handling it right now.
# Should affect efficiency, but not correctness.
root.GDPbb_utils.disjuncts_fixed_True = ComponentSet()
# Only find top-level (non-nested) disjunctions
for disjunction in root.component_data_objects(Disjunction,
active=True):
fixed_true_disjuncts = [
disjunct for disjunct in disjunction.disjuncts
if disjunct.indicator_var.fixed
and disjunct.indicator_var.value == 1
]
fixed_false_disjuncts = [
disjunct for disjunct in disjunction.disjuncts
if disjunct.indicator_var.fixed
and disjunct.indicator_var.value == 0
]
for disjunct in fixed_false_disjuncts:
disjunct.deactivate()
if len(fixed_false_disjuncts) == len(
disjunction.disjuncts) - 1:
# all but one disjunct in the disjunction is fixed to False. Remaining one must be true.
if not fixed_true_disjuncts:
fixed_true_disjuncts = [
disjunct for disjunct in disjunction.disjuncts
if disjunct not in fixed_false_disjuncts
]
# Reactivate the fixed-true disjuncts
for disjunct in fixed_true_disjuncts:
newly_activated = ComponentSet()
for constr in disjunct.component_data_objects(Constraint):
if constr in root.GDPbb_utils.deactivated_constraints:
newly_activated.add(constr)
constr.activate()
# Set the big M value for the constraint
root.BigM[constr] = 1
# Note: we use a default big M value of 1
# because all non-selected disjuncts should be deactivated.
# Therefore, none of the big M transformed nonlinear constraints will need to be relaxed.
# The default M value should therefore be irrelevant.
root.GDPbb_utils.deactivated_constraints -= newly_activated
root.GDPbb_utils.disjuncts_fixed_True.add(disjunct)
if fixed_true_disjuncts:
assert disjunction.xor, "GDPbb only handles disjunctions in which one term can be selected. " \
"%s violates this assumption." % (disjunction.name, )
root.GDPbb_utils.unenforced_disjunctions.remove(
disjunction)
# Check satisfiability
if config.check_sat and satisfiable(root, config.logger) is False:
# Problem is not satisfiable. Problem is infeasible.
obj_value = obj_sign * float('inf')
else:
# solve the root node
config.logger.info("Solving the root node.")
obj_value, result, var_values = self.subproblem_solve(
root, config)
if obj_sign * obj_value == float('inf'):
config.logger.info(
"Model was found to be infeasible at the root node. Elapsed %.2f seconds."
% get_main_elapsed_time(solve_data.timing))
if solve_data.results.problem.sense == minimize:
solve_data.results.problem.lower_bound = float('inf')
solve_data.results.problem.upper_bound = None
else:
solve_data.results.problem.lower_bound = None
solve_data.results.problem.upper_bound = float('-inf')
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = 0
solve_data.results.solver.termination_condition = tc.infeasible
return solve_data.results
# initialize minheap for Branch and Bound algorithm
# Heap structure: (ordering tuple, model)
# Ordering tuple: (objective value, disjunctions_left, -total_nodes_counter)
# - select solutions with lower objective value,
# then fewer disjunctions left to explore (depth first),
# then more recently encountered (tiebreaker)
heap = []
total_nodes_counter = 0
disjunctions_left = len(root.GDPbb_utils.unenforced_disjunctions)
heapq.heappush(heap,
((obj_sign * obj_value, disjunctions_left,
-total_nodes_counter), root, result, var_values))
# loop to branch through the tree
while len(heap) > 0:
# pop best model off of heap
sort_tuple, incumbent_model, incumbent_results, incumbent_var_values = heapq.heappop(
heap)
incumbent_obj_value, disjunctions_left, _ = sort_tuple
config.logger.info(
"Exploring node with LB %.10g and %s inactive disjunctions."
% (incumbent_obj_value, disjunctions_left))
# if all the originally active disjunctions are active, solve and
# return solution
if disjunctions_left == 0:
config.logger.info("Model solved.")
# Model is solved. Copy over solution values.
original_model = solve_data.original_model
for orig_var, val in zip(
original_model.GDPbb_utils.variable_list,
incumbent_var_values):
orig_var.value = val
solve_data.results.problem.lower_bound = incumbent_results.problem.lower_bound
solve_data.results.problem.upper_bound = incumbent_results.problem.upper_bound
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = total_nodes_counter
solve_data.results.solver.termination_condition = incumbent_results.solver.termination_condition
return solve_data.results
# Pick the next disjunction to branch on
next_disjunction = incumbent_model.GDPbb_utils.unenforced_disjunctions[
0]
config.logger.info("Branching on disjunction %s" %
next_disjunction.name)
assert next_disjunction.xor, "GDPbb only handles disjunctions in which one term can be selected. " \
"%s violates this assumption." % (next_disjunction.name, )
new_nodes_counter = 0
for i, disjunct in enumerate(next_disjunction.disjuncts):
# Create one branch for each of the disjuncts on the disjunction
if any(disj.indicator_var.fixed
and disj.indicator_var.value == 1
for disj in next_disjunction.disjuncts
if disj is not disjunct):
# If any other disjunct is fixed to 1 and an xor relationship applies,
# then this disjunct cannot be activated.
continue
# Check time limit
if get_main_elapsed_time(
solve_data.timing) >= config.time_limit:
if solve_data.results.problem.sense == minimize:
solve_data.results.problem.lower_bound = incumbent_obj_value
solve_data.results.problem.upper_bound = float(
'inf')
else:
solve_data.results.problem.lower_bound = float(
'-inf')
solve_data.results.problem.upper_bound = incumbent_obj_value
config.logger.info('GDPopt unable to converge bounds '
'before time limit of {} seconds. '
'Elapsed: {} seconds'.format(
config.time_limit,
get_main_elapsed_time(
solve_data.timing)))
config.logger.info(
'Final bound values: LB: {} UB: {}'.format(
solve_data.results.problem.lower_bound,
solve_data.results.problem.upper_bound))
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = total_nodes_counter
solve_data.results.solver.termination_condition = tc.maxTimeLimit
return solve_data.results
# Branch on the disjunct
child = incumbent_model.clone()
# TODO I am leaving the old branching system in place, but there should be
# something better, ideally that deals with nested disjunctions as well.
disjunction_to_branch = child.GDPbb_utils.unenforced_disjunctions.pop(
0)
child_disjunct = disjunction_to_branch.disjuncts[i]
child_disjunct.indicator_var.fix(1)
# Deactivate (and fix to 0) other disjuncts on the disjunction
for disj in disjunction_to_branch.disjuncts:
if disj is not child_disjunct:
disj.deactivate()
# Activate nonlinear constraints on the newly fixed child disjunct
newly_activated = ComponentSet()
for constr in child_disjunct.component_data_objects(
Constraint):
if constr in child.GDPbb_utils.deactivated_constraints:
newly_activated.add(constr)
constr.activate()
# Set the big M value for the constraint
child.BigM[constr] = 1
# Note: we use a default big M value of 1
# because all non-selected disjuncts should be deactivated.
# Therefore, none of the big M transformed nonlinear constraints will need to be relaxed.
# The default M value should therefore be irrelevant.
child.GDPbb_utils.deactivated_constraints -= newly_activated
child.GDPbb_utils.disjuncts_fixed_True.add(child_disjunct)
if disjunct in incumbent_model.GDPbb_utils.disjuncts_fixed_True:
# If the disjunct was already branched to True from a parent disjunct branching, just pass
# through the incumbent value without resolving. The solution should be the same as the parent.
total_nodes_counter += 1
ordering_tuple = (obj_sign * incumbent_obj_value,
disjunctions_left - 1,
-total_nodes_counter)
heapq.heappush(heap, (ordering_tuple, child, result,
incumbent_var_values))
new_nodes_counter += 1
continue
if config.check_sat and satisfiable(
child, config.logger) is False:
# Problem is not satisfiable. Skip this disjunct.
continue
obj_value, result, var_values = self.subproblem_solve(
child, config)
total_nodes_counter += 1
ordering_tuple = (obj_sign * obj_value,
disjunctions_left - 1,
-total_nodes_counter)
heapq.heappush(heap,
(ordering_tuple, child, result, var_values))
new_nodes_counter += 1
config.logger.info(
"Added %s new nodes with %s relaxed disjunctions to the heap. Size now %s."
% (new_nodes_counter, disjunctions_left - 1, len(heap)))
def add_integer_cut(var_values, target_model, solve_data, config, feasible=False):
"""Add an integer cut to the target GDP model."""
with time_code(solve_data.timing, 'integer cut generation'):
m = target_model
GDPopt = m.GDPopt_utils
var_value_is_one = ComponentSet()
var_value_is_zero = ComponentSet()
indicator_vars = ComponentSet(disj.indicator_var for disj in GDPopt.disjunct_list)
for var, val in zip(GDPopt.variable_list, var_values):
if not var.is_binary():
continue
if var.fixed:
# if val is not None and var.value != val:
# # val needs to be None or match var.value. Otherwise, we have a
# # contradiction.
# raise ValueError(
# "Fixed variable %s has value %s != "
# "provided value of %s." % (var.name, var.value, val))
# Note: FBBT may cause some disjuncts to be fathomed, which can cause
# a fixed variable to be different than the subproblem value.
# In this case, we simply construct the integer cut as usual with
# the subproblem value rather than its fixed value.
if val is None:
val = var.value
if not config.force_subproblem_nlp:
# By default (config.force_subproblem_nlp = False), we only want
# the integer cuts to be over disjunct indicator vars.
if var not in indicator_vars:
continue
if fabs(val - 1) <= config.integer_tolerance:
var_value_is_one.add(var)
elif fabs(val) <= config.integer_tolerance:
var_value_is_zero.add(var)
else:
raise ValueError(
'Binary %s = %s is not 0 or 1' % (var.name, val))
if not (var_value_is_one or var_value_is_zero):
# if no remaining binary variables, then terminate algorithm.
config.logger.info(
'Adding integer cut to a model without discrete variables. '
'Model is now infeasible.')
if solve_data.objective_sense == minimize:
solve_data.LB = float('inf')
else:
solve_data.UB = float('-inf')
return False
int_cut = (sum(1 - v for v in var_value_is_one) +
sum(v for v in var_value_is_zero)) >= 1
# Exclude the current binary combination
config.logger.info('Adding integer cut')
GDPopt.integer_cuts.add(expr=int_cut)
if config.calc_disjunctive_bounds:
with time_code(solve_data.timing, "disjunctive variable bounding"):
TransformationFactory('contrib.compute_disj_var_bounds').apply_to(
m,
solver=config.mip_solver if config.obbt_disjunctive_bounds else None
)
def add_outer_approximation_cuts(nlp_result, solve_data, config):
"""Add outer approximation cuts to the linear GDP model."""
with time_code(solve_data.timing, 'OA cut generation'):
m = solve_data.linear_GDP
GDPopt = m.GDPopt_utils
sign_adjust = -1 if solve_data.objective_sense == minimize else 1
# copy values over
for var, val in zip(GDPopt.variable_list, nlp_result.var_values):
if val is not None and not var.fixed:
var.value = val
# TODO some kind of special handling if the dual is phenomenally small?
config.logger.debug('Adding OA cuts.')
counter = 0
if not hasattr(GDPopt, 'jacobians'):
GDPopt.jacobians = ComponentMap()
for constr, dual_value in zip(GDPopt.constraint_list,
nlp_result.dual_values):
if dual_value is None or constr.body.polynomial_degree() in (1, 0):
continue
# Determine if the user pre-specified that OA cuts should not be
# generated for the given constraint.
parent_block = constr.parent_block()
ignore_set = getattr(parent_block, 'GDPopt_ignore_OA', None)
config.logger.debug('Ignore_set %s' % ignore_set)
if (ignore_set and (constr in ignore_set
or constr.parent_component() in ignore_set)):
config.logger.debug(
'OA cut addition for %s skipped because it is in '
'the ignore set.' % constr.name)
continue
config.logger.debug("Adding OA cut for %s with dual value %s" %
(constr.name, dual_value))
# Cache jacobian
jacobian = GDPopt.jacobians.get(constr, None)
if jacobian is None:
constr_vars = list(
identify_variables(constr.body, include_fixed=False))
if len(constr_vars) >= MAX_SYMBOLIC_DERIV_SIZE:
mode = differentiate.Modes.reverse_numeric
else:
mode = differentiate.Modes.sympy
try:
jac_list = differentiate(constr.body,
wrt_list=constr_vars,
mode=mode)
jac_map = ComponentMap(zip(constr_vars, jac_list))
except:
if mode is differentiate.Modes.reverse_numeric:
raise
mode = differentiate.Modes.reverse_numeric
jac_map = ComponentMap()
jacobian = JacInfo(mode=mode, vars=constr_vars, jac=jac_map)
GDPopt.jacobians[constr] = jacobian
# Recompute numeric derivatives
if not jacobian.jac:
jac_list = differentiate(constr.body,
wrt_list=jacobian.vars,
mode=jacobian.mode)
jacobian.jac.update(zip(jacobian.vars, jac_list))
# Create a block on which to put outer approximation cuts.
oa_utils = parent_block.component('GDPopt_OA')
if oa_utils is None:
oa_utils = parent_block.GDPopt_OA = Block(
doc="Block holding outer approximation cuts "
"and associated data.")
oa_utils.GDPopt_OA_cuts = ConstraintList()
oa_utils.GDPopt_OA_slacks = VarList(bounds=(0,
config.max_slack),
domain=NonNegativeReals,
initialize=0)
oa_cuts = oa_utils.GDPopt_OA_cuts
slack_var = oa_utils.GDPopt_OA_slacks.add()
rhs = value(constr.lower) if constr.has_lb() else value(
constr.upper)
try:
new_oa_cut = (copysign(1, sign_adjust * dual_value) *
(value(constr.body) - rhs + sum(
value(jac) * (var - value(var))
for var, jac in iteritems(jacobian.jac))) -
slack_var <= 0)
if new_oa_cut.polynomial_degree() not in (1, 0):
for var, jac in iteritems(jacobian.jac):
print(var.name, value(jac))
oa_cuts.add(expr=new_oa_cut)
counter += 1
except ZeroDivisionError:
config.logger.warning(
"Zero division occured attempting to generate OA cut for constraint %s.\n"
"Skipping OA cut generation for this constraint." %
(constr.name, ))
# Simply continue on to the next constraint.
# Clear out the numeric Jacobian values
if jacobian.mode is differentiate.Modes.reverse_numeric:
jacobian.jac.clear()
config.logger.info('Added %s OA cuts' % counter)
def add_oa_cuts(target_model, dual_values, solve_data, config,
cb_opt=None,
linearize_active=True,
linearize_violated=True):
"""Adds OA cuts.
Generates and adds OA cuts (linearizes nonlinear constraints).
For nonconvex problems, turn on 'config.add_slack'.
Slack variables will always be used for nonlinear equality constraints.
Parameters
----------
target_model : Pyomo model
The relaxed linear model.
dual_values : list
The value of the duals for each constraint.
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
cb_opt : SolverFactory, optional
Gurobi_persistent solver, by default None.
linearize_active : bool, optional
Whether to linearize the active nonlinear constraints, by default True.
linearize_violated : bool, optional
Whether to linearize the violated nonlinear constraints, by default True.
"""
with time_code(solve_data.timing, 'OA cut generation'):
for index, constr in enumerate(target_model.MindtPy_utils.constraint_list):
# TODO: here the index is correlated to the duals, try if this can be fixed when temp duals are removed.
if constr.body.polynomial_degree() in {0, 1}:
continue
constr_vars = list(identify_variables(constr.body))
jacs = solve_data.jacobians
# Equality constraint (makes the problem nonconvex)
if constr.has_ub() and constr.has_lb() and value(constr.lower) == value(constr.upper) and config.equality_relaxation:
sign_adjust = -1 if solve_data.objective_sense == minimize else 1
rhs = constr.lower
if config.add_slack:
slack_var = target_model.MindtPy_utils.cuts.slack_vars.add()
target_model.MindtPy_utils.cuts.oa_cuts.add(
expr=copysign(1, sign_adjust * dual_values[index])
* (sum(value(jacs[constr][var]) * (var - value(var))
for var in EXPR.identify_variables(constr.body))
+ value(constr.body) - rhs)
- (slack_var if config.add_slack else 0) <= 0)
if config.single_tree and config.mip_solver == 'gurobi_persistent' and solve_data.mip_iter > 0 and cb_opt is not None:
cb_opt.cbLazy(
target_model.MindtPy_utils.cuts.oa_cuts[len(target_model.MindtPy_utils.cuts.oa_cuts)])
else: # Inequality constraint (possibly two-sided)
if (constr.has_ub()
and (linearize_active and abs(constr.uslack()) < config.zero_tolerance)
or (linearize_violated and constr.uslack() < 0)
or (config.linearize_inactive and constr.uslack() > 0)) or ('MindtPy_utils.objective_constr' in constr.name and constr.has_ub()):
# always add the linearization for the epigraph of the objective
if config.add_slack:
slack_var = target_model.MindtPy_utils.cuts.slack_vars.add()
target_model.MindtPy_utils.cuts.oa_cuts.add(
expr=(sum(value(jacs[constr][var])*(var - var.value)
for var in constr_vars) + value(constr.body)
- (slack_var if config.add_slack else 0)
<= value(constr.upper))
)
if config.single_tree and config.mip_solver == 'gurobi_persistent' and solve_data.mip_iter > 0 and cb_opt is not None:
cb_opt.cbLazy(
target_model.MindtPy_utils.cuts.oa_cuts[len(target_model.MindtPy_utils.cuts.oa_cuts)])
if (constr.has_lb()
and (linearize_active and abs(constr.lslack()) < config.zero_tolerance)
or (linearize_violated and constr.lslack() < 0)
or (config.linearize_inactive and constr.lslack() > 0)) or ('MindtPy_utils.objective_constr' in constr.name and constr.has_lb()):
if config.add_slack:
slack_var = target_model.MindtPy_utils.cuts.slack_vars.add()
target_model.MindtPy_utils.cuts.oa_cuts.add(
expr=(sum(value(jacs[constr][var])*(var - var.value)
for var in constr_vars) + value(constr.body)
+ (slack_var if config.add_slack else 0)
>= value(constr.lower))
)
if config.single_tree and config.mip_solver == 'gurobi_persistent' and solve_data.mip_iter > 0 and cb_opt is not None:
cb_opt.cbLazy(
target_model.MindtPy_utils.cuts.oa_cuts[len(target_model.MindtPy_utils.cuts.oa_cuts)])
def solve(self, model, **kwds):
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
# Validate model to be used with gdpbb
self.validate_model(model)
# Set solver as an MINLP
solver = SolverFactory(config.solver)
solve_data = GDPbbSolveData()
solve_data.timing = Container()
solve_data.original_model = model
solve_data.results = SolverResults()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total'), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPbb_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPbb version %s using %s as subsolver" %
(".".join(map(str, self.version())), config.solver))
# Setup results
solve_data.results.solver.name = 'GDPbb - %s' % (str(
config.solver))
setup_results_object(solve_data, config)
# Initialize list containing indicator vars for reupdating model after solving
indicator_list_name = unique_component_name(
model, "_indicator_list")
indicator_vars = []
for disjunction in model.component_data_objects(ctype=Disjunction,
active=True):
for disjunct in disjunction.disjuncts:
indicator_vars.append(disjunct.indicator_var)
setattr(model, indicator_list_name, indicator_vars)
# get objective sense
objectives = model.component_data_objects(Objective, active=True)
obj = next(objectives, None)
obj_sign = 1 if obj.sense == minimize else -1
solve_data.results.problem.sense = obj.sense
# clone original model for root node of branch and bound
root = model.clone()
# set up lists to keep track of which disjunctions have been covered.
# this list keeps track of the original disjunctions that were active and are soon to be inactive
root.GDPbb_utils.unenforced_disjunctions = list(
disjunction
for disjunction in root.GDPbb_utils.disjunction_list
if disjunction.active)
# this list keeps track of the disjunctions that have been activated by the branch and bound
root.GDPbb_utils.curr_active_disjunctions = []
# deactivate all disjunctions in the model
# self.indicate(root)
for djn in root.GDPbb_utils.unenforced_disjunctions:
djn.deactivate()
# Deactivate all disjuncts in model. To be reactivated when disjunction
# is reactivated.
for disj in root.component_data_objects(Disjunct, active=True):
disj._deactivate_without_fixing_indicator()
# Satisfiability check would go here
# solve the root node
config.logger.info("Solving the root node.")
obj_value, result, _ = self.subproblem_solve(root, solver, config)
# initialize minheap for Branch and Bound algorithm
# Heap structure: (ordering tuple, model)
# Ordering tuple: (objective value, disjunctions_left, -counter)
# - select solutions with lower objective value,
# then fewer disjunctions left to explore (depth first),
# then more recently encountered (tiebreaker)
heap = []
counter = 0
disjunctions_left = len(root.GDPbb_utils.unenforced_disjunctions)
heapq.heappush(
heap, ((obj_sign * obj_value, disjunctions_left, -counter),
root, result, root.GDPbb_utils.variable_list))
# loop to branch through the tree
while len(heap) > 0:
# pop best model off of heap
sort_tup, mdl, mdl_results, vars = heapq.heappop(heap)
old_obj_val, disjunctions_left, _ = sort_tup
config.logger.info(
"Exploring node with LB %.10g and %s inactive disjunctions."
% (old_obj_val, disjunctions_left))
# if all the originally active disjunctions are active, solve and
# return solution
if disjunctions_left == 0:
config.logger.info("Model solved.")
# Model is solved. Copy over solution values.
for orig_var, soln_var in zip(
model.GDPbb_utils.variable_list, vars):
orig_var.value = soln_var.value
solve_data.results.problem.lower_bound = mdl_results.problem.lower_bound
solve_data.results.problem.upper_bound = mdl_results.problem.upper_bound
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.termination_condition = mdl_results.solver.termination_condition
return solve_data.results
next_disjunction = mdl.GDPbb_utils.unenforced_disjunctions.pop(
0)
config.logger.info("Activating disjunction %s" %
next_disjunction.name)
next_disjunction.activate()
mdl.GDPbb_utils.curr_active_disjunctions.append(
next_disjunction)
djn_left = len(mdl.GDPbb_utils.unenforced_disjunctions)
for disj in next_disjunction.disjuncts:
disj._activate_without_unfixing_indicator()
if not disj.indicator_var.fixed:
disj.indicator_var = 0 # initially set all indicator vars to zero
added_disj_counter = 0
for disj in next_disjunction.disjuncts:
if not disj.indicator_var.fixed:
disj.indicator_var = 1
mnew = mdl.clone()
if not disj.indicator_var.fixed:
disj.indicator_var = 0
# Check feasibility
if config.check_sat and satisfiable(
mnew, config.logger) is False:
# problem is not satisfiable. Skip this disjunct.
continue
obj_value, result, vars = self.subproblem_solve(
mnew, solver, config)
counter += 1
ordering_tuple = (obj_sign * obj_value, djn_left, -counter)
heapq.heappush(heap, (ordering_tuple, mnew, result, vars))
added_disj_counter = added_disj_counter + 1
config.logger.info(
"Added %s new nodes with %s relaxed disjunctions to the heap. Size now %s."
% (added_disj_counter, djn_left, len(heap)))
def solve_fp_subproblem(solve_data, config):
"""
Solves the feasibility pump NLP
This function sets up the 'fp_nlp' by relax integer variables.
precomputes dual values, deactivates trivial constraints, and then solves NLP model.
Parameters
----------
solve_data: MindtPy Data Container
data container that holds solve-instance data
config: ConfigBlock
contains the specific configurations for the algorithm
Returns
-------
fp_nlp: Pyomo model
Fixed-NLP from the model
results: Pyomo results object
result from solving the Fixed-NLP
"""
fp_nlp = solve_data.working_model.clone()
MindtPy = fp_nlp.MindtPy_utils
config.logger.info('FP-NLP %s: Solve feasibility pump NLP subproblem.'
% (solve_data.fp_iter,))
# Set up NLP
fp_nlp.MindtPy_utils.objective_list[-1].deactivate()
if solve_data.objective_sense == minimize:
fp_nlp.improving_objective_cut = Constraint(
expr=fp_nlp.MindtPy_utils.objective_value <= solve_data.UB)
else:
fp_nlp.improving_objective_cut = Constraint(
expr=fp_nlp.MindtPy_utils.objective_value >= solve_data.LB)
# Add norm_constraint, which guarantees the monotonicity of the norm objective value sequence of all iterations
# Ref: Paper 'A storm of feasibility pumps for nonconvex MINLP'
# the norm type is consistant with the norm obj of the FP-main problem.
if config.fp_norm_constraint:
if config.fp_main_norm == 'L1':
# TODO: check if we can access the block defined in FP-main problem
generate_norm1_norm_constraint(
fp_nlp, solve_data.mip, config, discrete_only=True)
elif config.fp_main_norm == 'L2':
fp_nlp.norm_constraint = Constraint(expr=sum((nlp_var - mip_var.value)**2 - config.fp_norm_constraint_coef*(nlp_var.value - mip_var.value)**2
for nlp_var, mip_var in zip(fp_nlp.MindtPy_utils.discrete_variable_list, solve_data.mip.MindtPy_utils.discrete_variable_list)) <= 0)
elif config.fp_main_norm == 'L_infinity':
fp_nlp.norm_constraint = ConstraintList()
rhs = config.fp_norm_constraint_coef * max(nlp_var.value - mip_var.value for nlp_var, mip_var in zip(
fp_nlp.MindtPy_utils.discrete_variable_list, solve_data.mip.MindtPy_utils.discrete_variable_list))
for nlp_var, mip_var in zip(fp_nlp.MindtPy_utils.discrete_variable_list, solve_data.mip.MindtPy_utils.discrete_variable_list):
fp_nlp.norm_constraint.add(nlp_var - mip_var.value <= rhs)
MindtPy.fp_nlp_obj = generate_norm2sq_objective_function(
fp_nlp, solve_data.mip, discrete_only=config.fp_discrete_only)
MindtPy.cuts.deactivate()
TransformationFactory('core.relax_integer_vars').apply_to(fp_nlp)
try:
TransformationFactory('contrib.deactivate_trivial_constraints').apply_to(
fp_nlp, tmp=True, ignore_infeasible=False, tolerance=config.constraint_tolerance)
except ValueError:
config.logger.warning(
'infeasibility detected in deactivate_trivial_constraints')
results = SolverResults()
results.solver.termination_condition = tc.infeasible
return fp_nlp, results
# Solve the NLP
nlpopt = SolverFactory(config.nlp_solver)
nlp_args = dict(config.nlp_solver_args)
set_solver_options(nlpopt, solve_data, config, solver_type='nlp')
with SuppressInfeasibleWarning():
with time_code(solve_data.timing, 'fp subproblem'):
results = nlpopt.solve(
fp_nlp, tee=config.nlp_solver_tee, **nlp_args)
return fp_nlp, results
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
Warning: at this point in time, if you try to use PSC or GBD with
anything other than IPOPT as the NLP solver, bad things will happen.
This is because the suffixes are not in place to extract dual values
from the variable bounds for any other solver.
TODO: fix needed with the GBD implementation.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
solve_data = MindtPySolveData()
solve_data.results = SolverResults()
solve_data.timing = Container()
solve_data.original_model = model
solve_data.working_model = model.clone()
if config.integer_to_binary:
TransformationFactory('contrib.integer_to_binary'). \
apply_to(solve_data.working_model)
new_logging_level = logging.INFO if config.tee else None
with time_code(solve_data.timing, 'total', is_main_timer=True), \
lower_logger_level_to(config.logger, new_logging_level), \
create_utility_block(solve_data.working_model, 'MindtPy_utils', solve_data):
config.logger.info("---Starting MindtPy---")
MindtPy = solve_data.working_model.MindtPy_utils
setup_results_object(solve_data, config)
process_objective(solve_data, config)
# Save model initial values.
solve_data.initial_var_values = list(
v.value for v in MindtPy.variable_list)
# Store the initial model state as the best solution found. If we
# find no better solution, then we will restore from this copy.
solve_data.best_solution_found = None
# Record solver name
solve_data.results.solver.name = 'MindtPy' + str(config.strategy)
# Validate the model to ensure that MindtPy is able to solve it.
if not model_is_valid(solve_data, config):
return
# Create a model block in which to store the generated feasibility
# slack constraints. Do not leave the constraints on by default.
feas = MindtPy.MindtPy_feas = Block()
feas.deactivate()
feas.feas_constraints = ConstraintList(
doc='Feasibility Problem Constraints')
# Create a model block in which to store the generated linear
# constraints. Do not leave the constraints on by default.
lin = MindtPy.MindtPy_linear_cuts = Block()
lin.deactivate()
# Integer cuts exclude particular discrete decisions
lin.integer_cuts = ConstraintList(doc='integer cuts')
# Feasible integer cuts exclude discrete realizations that have
# been explored via an NLP subproblem. Depending on model
# characteristics, the user may wish to revisit NLP subproblems
# (with a different initialization, for example). Therefore, these
# cuts are not enabled by default.
#
# Note: these cuts will only exclude integer realizations that are
# not already in the primary integer_cuts ConstraintList.
lin.feasible_integer_cuts = ConstraintList(
doc='explored integer cuts')
lin.feasible_integer_cuts.deactivate()
# Set up iteration counters
solve_data.nlp_iter = 0
solve_data.mip_iter = 0
solve_data.mip_subiter = 0
# set up bounds
solve_data.LB = float('-inf')
solve_data.UB = float('inf')
solve_data.LB_progress = [solve_data.LB]
solve_data.UB_progress = [solve_data.UB]
# Set of NLP iterations for which cuts were generated
lin.nlp_iters = Set(dimen=1)
# Set of MIP iterations for which cuts were generated in ECP
lin.mip_iters = Set(dimen=1)
nonlinear_constraints = [
c for c in MindtPy.constraint_list
if c.body.polynomial_degree() not in (1, 0)
]
lin.nl_constraint_set = RangeSet(
len(nonlinear_constraints),
doc="Integer index set over the nonlinear constraints")
feas.constraint_set = RangeSet(
len(MindtPy.constraint_list),
doc="integer index set over the constraints")
# # Mapping Constraint -> integer index
# MindtPy.feas_map = {}
# # Mapping integer index -> Constraint
# MindtPy.feas_inverse_map = {}
# # Generate the two maps. These maps may be helpful for later
# # interpreting indices on the slack variables or generated cuts.
# for c, n in zip(MindtPy.constraint_list, feas.constraint_set):
# MindtPy.feas_map[c] = n
# MindtPy.feas_inverse_map[n] = c
# Create slack variables for OA cuts
lin.slack_vars = VarList(bounds=(0, config.max_slack),
initialize=0,
domain=NonNegativeReals)
# Create slack variables for feasibility problem
feas.slack_var = Var(feas.constraint_set,
domain=NonNegativeReals,
initialize=1)
# Flag indicating whether the solution improved in the past
# iteration or not
solve_data.solution_improved = False
if not hasattr(solve_data.working_model, 'ipopt_zL_out'):
solve_data.working_model.ipopt_zL_out = Suffix(
direction=Suffix.IMPORT)
if not hasattr(solve_data.working_model, 'ipopt_zU_out'):
solve_data.working_model.ipopt_zU_out = Suffix(
direction=Suffix.IMPORT)
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
MindtPy_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
MindtPy_iteration_loop(solve_data, config)
if solve_data.best_solution_found is not None:
# Update values in original model
copy_var_list_values(from_list=solve_data.best_solution_found.
MindtPy_utils.variable_list,
to_list=MindtPy.variable_list,
config=config)
# MindtPy.objective_value.set_value(
# value(solve_data.working_objective_expr, exception=False))
copy_var_list_values(
MindtPy.variable_list,
solve_data.original_model.component_data_objects(Var),
config)
solve_data.results.problem.lower_bound = solve_data.LB
solve_data.results.problem.upper_bound = solve_data.UB
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.user_time = solve_data.timing.total
solve_data.results.solver.wallclock_time = solve_data.timing.total
solve_data.results.solver.iterations = solve_data.mip_iter
return solve_data.results
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
This function performs all of the GDPopt solver setup and problem
validation. It then calls upon helper functions to construct the
initial master approximation and iteration loop.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
solve_data = GDPoptSolveData()
solve_data.results = SolverResults()
solve_data.timing = Container()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total'), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPopt_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPopt version %s using %s algorithm"
% (".".join(map(str, self.version())), config.strategy)
)
config.logger.info(
"""
If you use this software, you may cite the following:
- Implementation:
Chen, Q; Johnson, ES; Siirola, JD; Grossmann, IE.
Pyomo.GDP: Disjunctive Models in Python.
Proc. of the 13th Intl. Symposium on Process Systems Eng.
San Diego, 2018.
- LOA algorithm:
Türkay, M; Grossmann, IE.
Logic-based MINLP algorithms for the optimal synthesis of process networks.
Comp. and Chem. Eng. 1996, 20(8), 959–978.
DOI: 10.1016/0098-1354(95)00219-7.
- GLOA algorithm:
Lee, S; Grossmann, IE.
A Global Optimization Algorithm for Nonconvex Generalized Disjunctive Programming and Applications to Process Systems
Comp. and Chem. Eng. 2001, 25, 1675-1697.
DOI: 10.1016/S0098-1354(01)00732-3
""".strip()
)
solve_data.results.solver.name = 'GDPopt %s - %s' % (
str(self.version()), config.strategy)
solve_data.original_model = model
solve_data.working_model = model.clone()
GDPopt = solve_data.working_model.GDPopt_utils
setup_results_object(solve_data, config)
solve_data.current_strategy = config.strategy
# Verify that objective has correct form
process_objective(solve_data, config)
# Save model initial values. These are used later to initialize NLP
# subproblems.
solve_data.initial_var_values = list(
v.value for v in GDPopt.variable_list)
solve_data.best_solution_found = None
# Validate the model to ensure that GDPopt is able to solve it.
if not model_is_valid(solve_data, config):
return
# Integer cuts exclude particular discrete decisions
GDPopt.integer_cuts = ConstraintList(doc='integer cuts')
# Feasible integer cuts exclude discrete realizations that have
# been explored via an NLP subproblem. Depending on model
# characteristics, the user may wish to revisit NLP subproblems
# (with a different initialization, for example). Therefore, these
# cuts are not enabled by default, unless the initial model has no
# discrete decisions.
# Note: these cuts will only exclude integer realizations that are
# not already in the primary GDPopt_integer_cuts ConstraintList.
GDPopt.no_backtracking = ConstraintList(
doc='explored integer cuts')
# Set up iteration counters
solve_data.master_iteration = 0
solve_data.mip_iteration = 0
solve_data.nlp_iteration = 0
# set up bounds
solve_data.LB = float('-inf')
solve_data.UB = float('inf')
solve_data.iteration_log = {}
# Flag indicating whether the solution improved in the past
# iteration or not
solve_data.feasible_solution_improved = False
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
GDPopt_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
GDPopt_iteration_loop(solve_data, config)
if solve_data.best_solution_found is not None:
# Update values in working model
copy_var_list_values(
from_list=solve_data.best_solution_found.GDPopt_utils.variable_list,
to_list=GDPopt.variable_list,
config=config)
# Update values in original model
copy_var_list_values(
GDPopt.variable_list,
solve_data.original_model.GDPopt_utils.variable_list,
config)
solve_data.results.problem.lower_bound = solve_data.LB
solve_data.results.problem.upper_bound = solve_data.UB
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.user_time = solve_data.timing.total
solve_data.results.solver.wallclock_time = solve_data.timing.total
solve_data.results.solver.iterations = solve_data.master_iteration
return solve_data.results
def solve_main(solve_data, config, fp=False, regularization_problem=False):
"""This function solves the MIP main problem.
Args:
solve_data (MindtPySolveData): data container that holds solve-instance data.
config (ConfigBlock): the specific configurations for MindtPy.
fp (bool, optional): whether it is in the loop of feasibility pump. Defaults to False.
regularization_problem (bool, optional): whether it is solving a regularization problem. Defaults to False.
Returns:
solve_data.mip (Pyomo model): the MIP stored in solve_data.
main_mip_results (SolverResults): results from solving the main MIP.
"""
if not fp and not regularization_problem:
solve_data.mip_iter += 1
# setup main problem
setup_main(solve_data, config, fp, regularization_problem)
mainopt = set_up_mip_solver(solve_data, config, regularization_problem)
mip_args = dict(config.mip_solver_args)
if config.mip_solver in {
'cplex', 'cplex_persistent', 'gurobi', 'gurobi_persistent'
}:
mip_args['warmstart'] = True
set_solver_options(mainopt,
solve_data,
config,
solver_type='mip',
regularization=regularization_problem)
try:
with time_code(
solve_data.timing,
'regularization main' if regularization_problem else
('fp main' if fp else 'main')):
main_mip_results = mainopt.solve(solve_data.mip,
tee=config.mip_solver_tee,
**mip_args)
except (ValueError, AttributeError):
if config.single_tree:
config.logger.warning('Single tree terminate.')
if get_main_elapsed_time(
solve_data.timing) >= config.time_limit - 2:
config.logger.warning('due to the timelimit.')
solve_data.results.solver.termination_condition = tc.maxTimeLimit
if config.strategy == 'GOA' or config.add_no_good_cuts:
config.logger.warning(
'ValueError: Cannot load a SolverResults object with bad status: error. '
'MIP solver failed. This usually happens in the single-tree GOA algorithm. '
"No-good cuts are added and GOA algorithm doesn't converge within the time limit. "
'No integer solution is found, so the cplex solver will report an error status. '
)
return None, None
if config.solution_pool:
main_mip_results._solver_model = mainopt._solver_model
main_mip_results._pyomo_var_to_solver_var_map = mainopt._pyomo_var_to_solver_var_map
if main_mip_results.solver.termination_condition is tc.optimal:
if config.single_tree and not config.add_no_good_cuts and not regularization_problem:
uptade_suboptimal_dual_bound(solve_data, main_mip_results)
if regularization_problem:
config.logger.info(
solve_data.log_formatter.format(
solve_data.mip_iter,
'Reg ' + solve_data.regularization_mip_type,
value(solve_data.mip.MindtPy_utils.loa_proj_mip_obj),
solve_data.LB, solve_data.UB, solve_data.rel_gap,
get_main_elapsed_time(solve_data.timing)))
elif main_mip_results.solver.termination_condition is tc.infeasibleOrUnbounded:
# Linear solvers will sometimes tell me that it's infeasible or
# unbounded during presolve, but fails to distinguish. We need to
# resolve with a solver option flag on.
main_mip_results, _ = distinguish_mip_infeasible_or_unbounded(
solve_data.mip, config)
return solve_data.mip, main_mip_results
if regularization_problem:
solve_data.mip.MindtPy_utils.objective_constr.deactivate()
solve_data.mip.MindtPy_utils.del_component('loa_proj_mip_obj')
solve_data.mip.MindtPy_utils.cuts.del_component('obj_reg_estimate')
if config.add_regularization == 'level_L1':
solve_data.mip.MindtPy_utils.del_component('L1_obj')
elif config.add_regularization == 'level_L_infinity':
solve_data.mip.MindtPy_utils.del_component('L_infinity_obj')
return solve_data.mip, main_mip_results
def add_affine_cuts(solve_data, config):
"""Adds affine cuts using MCPP.
Parameters
----------
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
"""
with time_code(solve_data.timing, 'Affine cut generation'):
m = solve_data.mip
config.logger.debug('Adding affine cuts')
counter = 0
for constr in m.MindtPy_utils.nonlinear_constraint_list:
vars_in_constr = list(
identify_variables(constr.body))
if any(var.value is None for var in vars_in_constr):
continue # a variable has no values
# mcpp stuff
try:
mc_eqn = mc(constr.body)
except MCPP_Error as e:
config.logger.debug(
'Skipping constraint %s due to MCPP error %s' % (constr.name, str(e)))
continue # skip to the next constraint
ccSlope = mc_eqn.subcc()
cvSlope = mc_eqn.subcv()
ccStart = mc_eqn.concave()
cvStart = mc_eqn.convex()
# check if the value of ccSlope and cvSlope is not Nan or inf. If so, we skip this.
concave_cut_valid = True
convex_cut_valid = True
for var in vars_in_constr:
if not var.fixed:
if ccSlope[var] == float('nan') or ccSlope[var] == float('inf'):
concave_cut_valid = False
if cvSlope[var] == float('nan') or cvSlope[var] == float('inf'):
convex_cut_valid = False
# check if the value of ccSlope and cvSlope all equals zero. if so, we skip this.
if not any(list(ccSlope.values())):
concave_cut_valid = False
if not any(list(cvSlope.values())):
convex_cut_valid = False
if ccStart == float('nan') or ccStart == float('inf'):
concave_cut_valid = False
if cvStart == float('nan') or cvStart == float('inf'):
convex_cut_valid = False
if not (concave_cut_valid or convex_cut_valid):
continue
ub_int = min(value(constr.upper), mc_eqn.upper()
) if constr.has_ub() else mc_eqn.upper()
lb_int = max(value(constr.lower), mc_eqn.lower()
) if constr.has_lb() else mc_eqn.lower()
aff_cuts = m.MindtPy_utils.cuts.aff_cuts
if concave_cut_valid:
concave_cut = sum(ccSlope[var] * (var - var.value)
for var in vars_in_constr
if not var.fixed) + ccStart >= lb_int
aff_cuts.add(expr=concave_cut)
counter += 1
if convex_cut_valid:
convex_cut = sum(cvSlope[var] * (var - var.value)
for var in vars_in_constr
if not var.fixed) + cvStart <= ub_int
aff_cuts.add(expr=convex_cut)
counter += 1
config.logger.debug('Added %s affine cuts' % counter)
def add_lazy_no_good_cuts(self,
var_values,
solve_data,
config,
opt,
feasible=False):
"""
Adds no-good cuts; add the no-good cuts through Cplex inherent function self.add()
Parameters
----------
var_values: list
values of the current variables, used to generate the cut
solve_data: MindtPy Data Container
data container that holds solve-instance data
config: ConfigBlock
contains the specific configurations for the algorithm
feasible: bool, optional
boolean indicating if integer combination yields a feasible or infeasible NLP
opt: SolverFactory
the mip solver
"""
if not config.add_no_good_cuts:
return
config.logger.info('Adding no-good cuts')
with time_code(solve_data.timing, 'No-good cut generation'):
m = solve_data.mip
MindtPy = m.MindtPy_utils
int_tol = config.integer_tolerance
binary_vars = [v for v in MindtPy.variable_list if v.is_binary()]
# copy variable values over
for var, val in zip(MindtPy.variable_list, var_values):
if not var.is_binary():
continue
var.value = val
# check to make sure that binary variables are all 0 or 1
for v in binary_vars:
if value(abs(v - 1)) > int_tol and value(abs(v)) > int_tol:
raise ValueError('Binary {} = {} is not 0 or 1'.format(
v.name, value(v)))
if not binary_vars: # if no binary variables, skip
return
pyomo_no_good_cut = sum(
1 - v
for v in binary_vars if value(abs(v - 1)) <= int_tol) + sum(
v for v in binary_vars if value(abs(v)) <= int_tol)
cplex_no_good_rhs = generate_standard_repn(
pyomo_no_good_cut).constant
cplex_no_good_cut, _ = opt._get_expr_from_pyomo_expr(
pyomo_no_good_cut)
self.add(constraint=cplex.SparsePair(
ind=cplex_no_good_cut.variables,
val=cplex_no_good_cut.coefficients),
sense='G',
rhs=1 - cplex_no_good_rhs)
def solve(self, model, **kwds):
"""Solve the model.
Warning: this solver is still in beta. Keyword arguments subject to
change. Undocumented keyword arguments definitely subject to change.
Warning: at this point in time, if you try to use PSC or GBD with
anything other than IPOPT as the NLP solver, bad things will happen.
This is because the suffixes are not in place to extract dual values
from the variable bounds for any other solver.
TODO: fix needed with the GBD implementation.
Args:
model (Block): a Pyomo model or block to be solved
"""
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
# configuration confirmation
if config.single_tree:
config.iteration_limit = 1
config.add_slack = False
config.add_nogood_cuts = False
config.mip_solver = 'cplex_persistent'
config.logger.info(
"Single tree implementation is activated. The defalt MIP solver is 'cplex_persistent'"
)
# if the slacks fix to zero, just don't add them
if config.max_slack == 0.0:
config.add_slack = False
if config.strategy == "GOA":
config.add_nogood_cuts = True
config.add_slack = True
config.use_mcpp = True
config.integer_to_binary = True
config.use_dual = False
config.use_fbbt = True
if config.nlp_solver == "baron":
config.use_dual = False
# if ecp tolerance is not provided use bound tolerance
if config.ecp_tolerance is None:
config.ecp_tolerance = config.bound_tolerance
# if the objective function is a constant, dual bound constraint is not added.
obj = next(model.component_data_objects(ctype=Objective, active=True))
if obj.expr.polynomial_degree() == 0:
config.use_dual_bound = False
solve_data = MindtPySolveData()
solve_data.results = SolverResults()
solve_data.timing = Container()
solve_data.curr_int_sol = []
solve_data.prev_int_sol = []
if config.use_fbbt:
fbbt(model)
config.logger.info(
"Use the fbbt to tighten the bounds of variables")
solve_data.original_model = model
solve_data.working_model = model.clone()
if config.integer_to_binary:
TransformationFactory('contrib.integer_to_binary'). \
apply_to(solve_data.working_model)
new_logging_level = logging.INFO if config.tee else None
with time_code(solve_data.timing, 'total', is_main_timer=True), \
lower_logger_level_to(config.logger, new_logging_level), \
create_utility_block(solve_data.working_model, 'MindtPy_utils', solve_data):
config.logger.info("---Starting MindtPy---")
MindtPy = solve_data.working_model.MindtPy_utils
setup_results_object(solve_data, config)
process_objective(solve_data, config, use_mcpp=config.use_mcpp)
# Save model initial values.
solve_data.initial_var_values = list(
v.value for v in MindtPy.variable_list)
# Store the initial model state as the best solution found. If we
# find no better solution, then we will restore from this copy.
solve_data.best_solution_found = None
solve_data.best_solution_found_time = None
# Record solver name
solve_data.results.solver.name = 'MindtPy' + str(config.strategy)
# Validate the model to ensure that MindtPy is able to solve it.
if not model_is_valid(solve_data, config):
return
# Create a model block in which to store the generated feasibility
# slack constraints. Do not leave the constraints on by default.
feas = MindtPy.MindtPy_feas = Block()
feas.deactivate()
feas.feas_constraints = ConstraintList(
doc='Feasibility Problem Constraints')
# Create a model block in which to store the generated linear
# constraints. Do not leave the constraints on by default.
lin = MindtPy.MindtPy_linear_cuts = Block()
lin.deactivate()
# Integer cuts exclude particular discrete decisions
lin.integer_cuts = ConstraintList(doc='integer cuts')
# Feasible integer cuts exclude discrete realizations that have
# been explored via an NLP subproblem. Depending on model
# characteristics, the user may wish to revisit NLP subproblems
# (with a different initialization, for example). Therefore, these
# cuts are not enabled by default.
#
# Note: these cuts will only exclude integer realizations that are
# not already in the primary integer_cuts ConstraintList.
lin.feasible_integer_cuts = ConstraintList(
doc='explored integer cuts')
lin.feasible_integer_cuts.deactivate()
# Set up iteration counters
solve_data.nlp_iter = 0
solve_data.mip_iter = 0
solve_data.mip_subiter = 0
# set up bounds
solve_data.LB = float('-inf')
solve_data.UB = float('inf')
solve_data.LB_progress = [solve_data.LB]
solve_data.UB_progress = [solve_data.UB]
if config.single_tree and config.add_nogood_cuts:
solve_data.stored_bound = {}
if config.strategy == 'GOA' and config.add_nogood_cuts:
solve_data.num_no_good_cuts_added = {}
# Set of NLP iterations for which cuts were generated
lin.nlp_iters = Set(dimen=1)
# Set of MIP iterations for which cuts were generated in ECP
lin.mip_iters = Set(dimen=1)
if config.feasibility_norm == 'L1' or config.feasibility_norm == 'L2':
feas.nl_constraint_set = Set(
initialize=[
i
for i, constr in enumerate(MindtPy.constraint_list, 1)
if constr.body.polynomial_degree() not in (1, 0)
],
doc="Integer index set over the nonlinear constraints."
"The set corresponds to the index of nonlinear constraint in constraint_set"
)
# Create slack variables for feasibility problem
feas.slack_var = Var(feas.nl_constraint_set,
domain=NonNegativeReals,
initialize=1)
else:
feas.slack_var = Var(domain=NonNegativeReals, initialize=1)
# Create slack variables for OA cuts
if config.add_slack:
lin.slack_vars = VarList(bounds=(0, config.max_slack),
initialize=0,
domain=NonNegativeReals)
# Flag indicating whether the solution improved in the past
# iteration or not
solve_data.solution_improved = False
if config.nlp_solver == 'ipopt':
if not hasattr(solve_data.working_model, 'ipopt_zL_out'):
solve_data.working_model.ipopt_zL_out = Suffix(
direction=Suffix.IMPORT)
if not hasattr(solve_data.working_model, 'ipopt_zU_out'):
solve_data.working_model.ipopt_zU_out = Suffix(
direction=Suffix.IMPORT)
# Initialize the master problem
with time_code(solve_data.timing, 'initialization'):
MindtPy_initialize_master(solve_data, config)
# Algorithm main loop
with time_code(solve_data.timing, 'main loop'):
MindtPy_iteration_loop(solve_data, config)
if solve_data.best_solution_found is not None:
# Update values in original model
copy_var_list_values(from_list=solve_data.best_solution_found.
MindtPy_utils.variable_list,
to_list=MindtPy.variable_list,
config=config)
# MindtPy.objective_value.set_value(
# value(solve_data.working_objective_expr, exception=False))
copy_var_list_values(
MindtPy.variable_list,
solve_data.original_model.component_data_objects(Var),
config)
solve_data.results.problem.lower_bound = solve_data.LB
solve_data.results.problem.upper_bound = solve_data.UB
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.user_time = solve_data.timing.total
solve_data.results.solver.wallclock_time = solve_data.timing.total
solve_data.results.solver.iterations = solve_data.mip_iter
solve_data.results.solver.best_solution_found_time = solve_data.best_solution_found_time
if config.single_tree:
solve_data.results.solver.num_nodes = solve_data.nlp_iter - \
(1 if config.init_strategy == 'rNLP' else 0)
return solve_data.results
def add_outer_approximation_cuts(nlp_result, solve_data, config):
"""Add outer approximation cuts to the linear GDP model."""
with time_code(solve_data.timing, 'OA cut generation'):
m = solve_data.linear_GDP
GDPopt = m.GDPopt_utils
sign_adjust = -1 if solve_data.objective_sense == minimize else 1
# copy values over
for var, val in zip(GDPopt.variable_list, nlp_result.var_values):
if val is not None and not var.fixed:
var.value = val
# TODO some kind of special handling if the dual is phenomenally small?
config.logger.debug('Adding OA cuts.')
counter = 0
if not hasattr(GDPopt, 'jacobians'):
GDPopt.jacobians = ComponentMap()
for constr, dual_value in zip(GDPopt.constraint_list,
nlp_result.dual_values):
if dual_value is None or constr.body.polynomial_degree() in (1, 0):
continue
# Determine if the user pre-specified that OA cuts should not be
# generated for the given constraint.
parent_block = constr.parent_block()
ignore_set = getattr(parent_block, 'GDPopt_ignore_OA', None)
config.logger.debug('Ignore_set %s' % ignore_set)
if (ignore_set and (constr in ignore_set
or constr.parent_component() in ignore_set)):
config.logger.debug(
'OA cut addition for %s skipped because it is in '
'the ignore set.' % constr.name)
continue
config.logger.debug("Adding OA cut for %s with dual value %s" %
(constr.name, dual_value))
# Cache jacobians
jacobians = GDPopt.jacobians.get(constr, None)
if jacobians is None:
constr_vars = list(identify_variables(constr.body))
jac_list = differentiate(constr.body, wrt_list=constr_vars)
jacobians = ComponentMap(zip(constr_vars, jac_list))
GDPopt.jacobians[constr] = jacobians
# Create a block on which to put outer approximation cuts.
oa_utils = parent_block.component('GDPopt_OA')
if oa_utils is None:
oa_utils = parent_block.GDPopt_OA = Block(
doc="Block holding outer approximation cuts "
"and associated data.")
oa_utils.GDPopt_OA_cuts = ConstraintList()
oa_utils.GDPopt_OA_slacks = VarList(bounds=(0,
config.max_slack),
domain=NonNegativeReals,
initialize=0)
oa_cuts = oa_utils.GDPopt_OA_cuts
slack_var = oa_utils.GDPopt_OA_slacks.add()
rhs = value(constr.lower) if constr.has_lb() else value(
constr.upper)
oa_cuts.add(expr=copysign(1, sign_adjust * dual_value) *
(value(constr.body) - rhs + sum(
value(jacobians[var]) * (var - value(var))
for var in jacobians)) - slack_var <= 0)
counter += 1
config.logger.info('Added %s OA cuts' % counter)
def test_handle_termination_condition(self):
"""Test the outer approximation decomposition algorithm."""
model = SimpleMINLP()
config = _get_MindtPy_config()
solve_data = set_up_solve_data(model, config)
with time_code(solve_data.timing, 'total', is_main_timer=True), \
create_utility_block(solve_data.working_model, 'MindtPy_utils', solve_data):
MindtPy = solve_data.working_model.MindtPy_utils
MindtPy = solve_data.working_model.MindtPy_utils
setup_results_object(solve_data, config)
process_objective(
solve_data,
config,
move_linear_objective=(config.init_strategy == 'FP' or
config.add_regularization is not None),
use_mcpp=config.use_mcpp,
update_var_con_list=config.add_regularization is None)
feas = MindtPy.feas_opt = Block()
feas.deactivate()
feas.feas_constraints = ConstraintList(
doc='Feasibility Problem Constraints')
lin = MindtPy.cuts = Block()
lin.deactivate()
if config.feasibility_norm == 'L1' or config.feasibility_norm == 'L2':
feas.nl_constraint_set = RangeSet(
len(MindtPy.nonlinear_constraint_list),
doc='Integer index set over the nonlinear constraints.')
# Create slack variables for feasibility problem
feas.slack_var = Var(feas.nl_constraint_set,
domain=NonNegativeReals,
initialize=1)
else:
feas.slack_var = Var(domain=NonNegativeReals, initialize=1)
# no-good cuts exclude particular discrete decisions
lin.no_good_cuts = ConstraintList(doc='no-good cuts')
fixed_nlp = solve_data.working_model.clone()
TransformationFactory('core.fix_integer_vars').apply_to(fixed_nlp)
MindtPy_initialize_main(solve_data, config)
# test handle_subproblem_other_termination
termination_condition = tc.maxIterations
config.add_no_good_cuts = True
handle_subproblem_other_termination(fixed_nlp,
termination_condition,
solve_data, config)
self.assertEqual(
len(solve_data.mip.MindtPy_utils.cuts.no_good_cuts), 1)
# test handle_main_other_conditions
main_mip, main_mip_results = solve_main(solve_data, config)
main_mip_results.solver.termination_condition = tc.infeasible
handle_main_other_conditions(solve_data.mip, main_mip_results,
solve_data, config)
self.assertIs(solve_data.results.solver.termination_condition,
tc.feasible)
main_mip_results.solver.termination_condition = tc.unbounded
handle_main_other_conditions(solve_data.mip, main_mip_results,
solve_data, config)
self.assertIn(main_mip.MindtPy_utils.objective_bound,
main_mip.component_data_objects(ctype=Constraint))
main_mip.MindtPy_utils.del_component('objective_bound')
main_mip_results.solver.termination_condition = tc.infeasibleOrUnbounded
handle_main_other_conditions(solve_data.mip, main_mip_results,
solve_data, config)
self.assertIn(main_mip.MindtPy_utils.objective_bound,
main_mip.component_data_objects(ctype=Constraint))
main_mip_results.solver.termination_condition = tc.maxTimeLimit
handle_main_other_conditions(solve_data.mip, main_mip_results,
solve_data, config)
self.assertIs(solve_data.results.solver.termination_condition,
tc.maxTimeLimit)
main_mip_results.solver.termination_condition = tc.other
main_mip_results.solution.status = SolutionStatus.feasible
handle_main_other_conditions(solve_data.mip, main_mip_results,
solve_data, config)
for v1, v2 in zip(
main_mip.MindtPy_utils.variable_list,
solve_data.working_model.MindtPy_utils.variable_list):
self.assertEqual(v1.value, v2.value)
# test handle_feasibility_subproblem_tc
feas_subproblem = solve_data.working_model.clone()
add_feas_slacks(feas_subproblem, config)
MindtPy = feas_subproblem.MindtPy_utils
MindtPy.feas_opt.activate()
if config.feasibility_norm == 'L1':
MindtPy.feas_obj = Objective(expr=sum(
s for s in MindtPy.feas_opt.slack_var[...]),
sense=minimize)
elif config.feasibility_norm == 'L2':
MindtPy.feas_obj = Objective(expr=sum(
s * s for s in MindtPy.feas_opt.slack_var[...]),
sense=minimize)
else:
MindtPy.feas_obj = Objective(expr=MindtPy.feas_opt.slack_var,
sense=minimize)
handle_feasibility_subproblem_tc(tc.optimal, MindtPy, solve_data,
config)
handle_feasibility_subproblem_tc(tc.infeasible, MindtPy,
solve_data, config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.status, SolverStatus.error)
solve_data.should_terminate = False
solve_data.results.solver.status = None
handle_feasibility_subproblem_tc(tc.maxIterations, MindtPy,
solve_data, config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.status, SolverStatus.error)
solve_data.should_terminate = False
solve_data.results.solver.status = None
handle_feasibility_subproblem_tc(tc.solverFailure, MindtPy,
solve_data, config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.status, SolverStatus.error)
# test NLP subproblem infeasible
solve_data.working_model.Y[1].value = 0
solve_data.working_model.Y[2].value = 0
solve_data.working_model.Y[3].value = 0
fixed_nlp, fixed_nlp_results = solve_subproblem(solve_data, config)
solve_data.working_model.Y[1].value = None
solve_data.working_model.Y[2].value = None
solve_data.working_model.Y[3].value = None
# test handle_nlp_subproblem_tc
fixed_nlp_results.solver.termination_condition = tc.maxTimeLimit
handle_nlp_subproblem_tc(fixed_nlp, fixed_nlp_results, solve_data,
config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.maxTimeLimit)
fixed_nlp_results.solver.termination_condition = tc.maxEvaluations
handle_nlp_subproblem_tc(fixed_nlp, fixed_nlp_results, solve_data,
config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.maxEvaluations)
fixed_nlp_results.solver.termination_condition = tc.maxIterations
handle_nlp_subproblem_tc(fixed_nlp, fixed_nlp_results, solve_data,
config)
self.assertIs(solve_data.should_terminate, True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.maxEvaluations)
# test handle_fp_main_tc
config.init_strategy = 'FP'
solve_data.fp_iter = 1
init_rNLP(solve_data, config)
feas_main, feas_main_results = solve_main(solve_data,
config,
fp=True)
feas_main_results.solver.termination_condition = tc.optimal
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, False)
feas_main_results.solver.termination_condition = tc.maxTimeLimit
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.maxTimeLimit)
feas_main_results.solver.termination_condition = tc.infeasible
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, True)
feas_main_results.solver.termination_condition = tc.unbounded
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, True)
feas_main_results.solver.termination_condition = tc.other
feas_main_results.solution.status = SolutionStatus.feasible
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, False)
feas_main_results.solver.termination_condition = tc.solverFailure
fp_should_terminate = handle_fp_main_tc(feas_main_results,
solve_data, config)
self.assertIs(fp_should_terminate, True)
# test generate_norm_constraint
fp_nlp = solve_data.working_model.clone()
config.fp_main_norm = 'L1'
generate_norm_constraint(fp_nlp, solve_data, config)
self.assertIsNotNone(
fp_nlp.MindtPy_utils.find_component('L1_norm_constraint'))
config.fp_main_norm = 'L2'
generate_norm_constraint(fp_nlp, solve_data, config)
self.assertIsNotNone(fp_nlp.find_component('norm_constraint'))
fp_nlp.del_component('norm_constraint')
config.fp_main_norm = 'L_infinity'
generate_norm_constraint(fp_nlp, solve_data, config)
self.assertIsNotNone(fp_nlp.find_component('norm_constraint'))
# test set_solver_options
config.mip_solver = 'gams'
config.threads = 1
opt = SolverFactory(config.mip_solver)
set_solver_options(opt,
solve_data,
config,
'mip',
regularization=False)
config.mip_solver = 'gurobi'
config.mip_regularization_solver = 'gurobi'
config.regularization_mip_threads = 1
opt = SolverFactory(config.mip_solver)
set_solver_options(opt,
solve_data,
config,
'mip',
regularization=True)
config.nlp_solver = 'gams'
config.nlp_solver_args['solver'] = 'ipopt'
set_solver_options(opt,
solve_data,
config,
'nlp',
regularization=False)
config.nlp_solver_args['solver'] = 'ipopth'
set_solver_options(opt,
solve_data,
config,
'nlp',
regularization=False)
config.nlp_solver_args['solver'] = 'conopt'
set_solver_options(opt,
solve_data,
config,
'nlp',
regularization=False)
config.nlp_solver_args['solver'] = 'msnlp'
set_solver_options(opt,
solve_data,
config,
'nlp',
regularization=False)
config.nlp_solver_args['solver'] = 'baron'
set_solver_options(opt,
solve_data,
config,
'nlp',
regularization=False)
# test algorithm_should_terminate
solve_data.should_terminate = True
solve_data.UB = float('inf')
self.assertIs(
algorithm_should_terminate(solve_data,
config,
check_cycling=False), True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.noSolution)
solve_data.UB = 100
self.assertIs(
algorithm_should_terminate(solve_data,
config,
check_cycling=False), True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.feasible)
solve_data.objective_sense = maximize
solve_data.LB = float('-inf')
self.assertIs(
algorithm_should_terminate(solve_data,
config,
check_cycling=False), True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.noSolution)
solve_data.LB = 100
self.assertIs(
algorithm_should_terminate(solve_data,
config,
check_cycling=False), True)
self.assertIs(solve_data.results.solver.termination_condition,
tc.feasible)
def add_outer_approximation_cuts(nlp_result, solve_data, config):
"""Add outer approximation cuts to the linear GDP model."""
with time_code(solve_data.timing, 'OA cut generation'):
m = solve_data.linear_GDP
GDPopt = m.GDPopt_utils
sign_adjust = -1 if solve_data.objective_sense == minimize else 1
# copy values over
for var, val in zip(GDPopt.variable_list, nlp_result.var_values):
if val is not None and not var.fixed:
var.value = val
# TODO some kind of special handling if the dual is phenomenally small?
config.logger.debug('Adding OA cuts.')
counter = 0
if not hasattr(GDPopt, 'jacobians'):
GDPopt.jacobians = ComponentMap()
for constr, dual_value in zip(GDPopt.constraint_list,
nlp_result.dual_values):
if dual_value is None or constr.body.polynomial_degree() in (1, 0):
continue
# Determine if the user pre-specified that OA cuts should not be
# generated for the given constraint.
parent_block = constr.parent_block()
ignore_set = getattr(parent_block, 'GDPopt_ignore_OA', None)
config.logger.debug('Ignore_set %s' % ignore_set)
if (ignore_set and (constr in ignore_set or
constr.parent_component() in ignore_set)):
config.logger.debug(
'OA cut addition for %s skipped because it is in '
'the ignore set.' % constr.name)
continue
config.logger.debug(
"Adding OA cut for %s with dual value %s"
% (constr.name, dual_value))
# Cache jacobians
jacobians = GDPopt.jacobians.get(constr, None)
if jacobians is None:
constr_vars = list(identify_variables(constr.body))
jac_list = differentiate(constr.body, wrt_list=constr_vars)
jacobians = ComponentMap(zip(constr_vars, jac_list))
GDPopt.jacobians[constr] = jacobians
# Create a block on which to put outer approximation cuts.
oa_utils = parent_block.component('GDPopt_OA')
if oa_utils is None:
oa_utils = parent_block.GDPopt_OA = Block(
doc="Block holding outer approximation cuts "
"and associated data.")
oa_utils.GDPopt_OA_cuts = ConstraintList()
oa_utils.GDPopt_OA_slacks = VarList(
bounds=(0, config.max_slack),
domain=NonNegativeReals, initialize=0)
oa_cuts = oa_utils.GDPopt_OA_cuts
slack_var = oa_utils.GDPopt_OA_slacks.add()
rhs = value(constr.lower) if constr.has_lb() else value(constr.upper)
oa_cuts.add(
expr=copysign(1, sign_adjust * dual_value) * (
value(constr.body) - rhs + sum(
value(jacobians[var]) * (var - value(var))
for var in jacobians)) - slack_var <= 0)
counter += 1
config.logger.info('Added %s OA cuts' % counter)
def solve(self, model, **kwds):
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
# Validate model to be used with gdpbb
self.validate_model(model)
# Set solver as an MINLP
solve_data = GDPbbSolveData()
solve_data.timing = Container()
solve_data.original_model = model
solve_data.results = SolverResults()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total', is_main_timer=True), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPbb_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPbb version %s using %s as subsolver"
% (".".join(map(str, self.version())), config.solver)
)
# Setup results
solve_data.results.solver.name = 'GDPbb - %s' % (str(config.solver))
setup_results_object(solve_data, config)
# clone original model for root node of branch and bound
root = solve_data.working_model = solve_data.original_model.clone()
# get objective sense
process_objective(solve_data, config)
objectives = solve_data.original_model.component_data_objects(Objective, active=True)
obj = next(objectives, None)
obj_sign = 1 if obj.sense == minimize else -1
solve_data.results.problem.sense = obj.sense
# set up lists to keep track of which disjunctions have been covered.
# this list keeps track of the relaxed disjunctions
root.GDPbb_utils.unenforced_disjunctions = list(
disjunction for disjunction in root.GDPbb_utils.disjunction_list if disjunction.active
)
root.GDPbb_utils.deactivated_constraints = ComponentSet([
constr for disjunction in root.GDPbb_utils.unenforced_disjunctions
for disjunct in disjunction.disjuncts
for constr in disjunct.component_data_objects(ctype=Constraint, active=True)
if constr.body.polynomial_degree() not in (1, 0)
])
# Deactivate nonlinear constraints in unenforced disjunctions
for constr in root.GDPbb_utils.deactivated_constraints:
constr.deactivate()
# Add the BigM suffix if it does not already exist. Used later during nonlinear constraint activation.
if not hasattr(root, 'BigM'):
root.BigM = Suffix()
# Pre-screen that none of the disjunctions are already predetermined due to the disjuncts being fixed
# to True/False values.
# TODO this should also be done within the loop, but we aren't handling it right now.
# Should affect efficiency, but not correctness.
root.GDPbb_utils.disjuncts_fixed_True = ComponentSet()
# Only find top-level (non-nested) disjunctions
for disjunction in root.component_data_objects(Disjunction, active=True):
fixed_true_disjuncts = [disjunct for disjunct in disjunction.disjuncts
if disjunct.indicator_var.fixed
and disjunct.indicator_var.value == 1]
fixed_false_disjuncts = [disjunct for disjunct in disjunction.disjuncts
if disjunct.indicator_var.fixed
and disjunct.indicator_var.value == 0]
for disjunct in fixed_false_disjuncts:
disjunct.deactivate()
if len(fixed_false_disjuncts) == len(disjunction.disjuncts) - 1:
# all but one disjunct in the disjunction is fixed to False. Remaining one must be true.
if not fixed_true_disjuncts:
fixed_true_disjuncts = [disjunct for disjunct in disjunction.disjuncts
if disjunct not in fixed_false_disjuncts]
# Reactivate the fixed-true disjuncts
for disjunct in fixed_true_disjuncts:
newly_activated = ComponentSet()
for constr in disjunct.component_data_objects(Constraint):
if constr in root.GDPbb_utils.deactivated_constraints:
newly_activated.add(constr)
constr.activate()
# Set the big M value for the constraint
root.BigM[constr] = 1
# Note: we use a default big M value of 1
# because all non-selected disjuncts should be deactivated.
# Therefore, none of the big M transformed nonlinear constraints will need to be relaxed.
# The default M value should therefore be irrelevant.
root.GDPbb_utils.deactivated_constraints -= newly_activated
root.GDPbb_utils.disjuncts_fixed_True.add(disjunct)
if fixed_true_disjuncts:
assert disjunction.xor, "GDPbb only handles disjunctions in which one term can be selected. " \
"%s violates this assumption." % (disjunction.name, )
root.GDPbb_utils.unenforced_disjunctions.remove(disjunction)
# Check satisfiability
if config.check_sat and satisfiable(root, config.logger) is False:
# Problem is not satisfiable. Problem is infeasible.
obj_value = obj_sign * float('inf')
else:
# solve the root node
config.logger.info("Solving the root node.")
obj_value, result, var_values = self.subproblem_solve(root, config)
if obj_sign * obj_value == float('inf'):
config.logger.info("Model was found to be infeasible at the root node. Elapsed %.2f seconds."
% get_main_elapsed_time(solve_data.timing))
if solve_data.results.problem.sense == minimize:
solve_data.results.problem.lower_bound = float('inf')
solve_data.results.problem.upper_bound = None
else:
solve_data.results.problem.lower_bound = None
solve_data.results.problem.upper_bound = float('-inf')
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = 0
solve_data.results.solver.termination_condition = tc.infeasible
return solve_data.results
# initialize minheap for Branch and Bound algorithm
# Heap structure: (ordering tuple, model)
# Ordering tuple: (objective value, disjunctions_left, -total_nodes_counter)
# - select solutions with lower objective value,
# then fewer disjunctions left to explore (depth first),
# then more recently encountered (tiebreaker)
heap = []
total_nodes_counter = 0
disjunctions_left = len(root.GDPbb_utils.unenforced_disjunctions)
heapq.heappush(
heap, (
(obj_sign * obj_value, disjunctions_left, -total_nodes_counter),
root, result, var_values))
# loop to branch through the tree
while len(heap) > 0:
# pop best model off of heap
sort_tuple, incumbent_model, incumbent_results, incumbent_var_values = heapq.heappop(heap)
incumbent_obj_value, disjunctions_left, _ = sort_tuple
config.logger.info("Exploring node with LB %.10g and %s inactive disjunctions." % (
incumbent_obj_value, disjunctions_left
))
# if all the originally active disjunctions are active, solve and
# return solution
if disjunctions_left == 0:
config.logger.info("Model solved.")
# Model is solved. Copy over solution values.
original_model = solve_data.original_model
for orig_var, val in zip(original_model.GDPbb_utils.variable_list, incumbent_var_values):
orig_var.value = val
solve_data.results.problem.lower_bound = incumbent_results.problem.lower_bound
solve_data.results.problem.upper_bound = incumbent_results.problem.upper_bound
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = total_nodes_counter
solve_data.results.solver.termination_condition = incumbent_results.solver.termination_condition
return solve_data.results
# Pick the next disjunction to branch on
next_disjunction = incumbent_model.GDPbb_utils.unenforced_disjunctions[0]
config.logger.info("Branching on disjunction %s" % next_disjunction.name)
assert next_disjunction.xor, "GDPbb only handles disjunctions in which one term can be selected. " \
"%s violates this assumption." % (next_disjunction.name, )
new_nodes_counter = 0
for i, disjunct in enumerate(next_disjunction.disjuncts):
# Create one branch for each of the disjuncts on the disjunction
if any(disj.indicator_var.fixed and disj.indicator_var.value == 1
for disj in next_disjunction.disjuncts if disj is not disjunct):
# If any other disjunct is fixed to 1 and an xor relationship applies,
# then this disjunct cannot be activated.
continue
# Check time limit
if get_main_elapsed_time(solve_data.timing) >= config.time_limit:
if solve_data.results.problem.sense == minimize:
solve_data.results.problem.lower_bound = incumbent_obj_value
solve_data.results.problem.upper_bound = float('inf')
else:
solve_data.results.problem.lower_bound = float('-inf')
solve_data.results.problem.upper_bound = incumbent_obj_value
config.logger.info(
'GDPopt unable to converge bounds '
'before time limit of {} seconds. '
'Elapsed: {} seconds'
.format(config.time_limit, get_main_elapsed_time(solve_data.timing)))
config.logger.info(
'Final bound values: LB: {} UB: {}'.
format(solve_data.results.problem.lower_bound, solve_data.results.problem.upper_bound))
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.iterations = total_nodes_counter
solve_data.results.solver.termination_condition = tc.maxTimeLimit
return solve_data.results
# Branch on the disjunct
child = incumbent_model.clone()
# TODO I am leaving the old branching system in place, but there should be
# something better, ideally that deals with nested disjunctions as well.
disjunction_to_branch = child.GDPbb_utils.unenforced_disjunctions.pop(0)
child_disjunct = disjunction_to_branch.disjuncts[i]
child_disjunct.indicator_var.fix(1)
# Deactivate (and fix to 0) other disjuncts on the disjunction
for disj in disjunction_to_branch.disjuncts:
if disj is not child_disjunct:
disj.deactivate()
# Activate nonlinear constraints on the newly fixed child disjunct
newly_activated = ComponentSet()
for constr in child_disjunct.component_data_objects(Constraint):
if constr in child.GDPbb_utils.deactivated_constraints:
newly_activated.add(constr)
constr.activate()
# Set the big M value for the constraint
child.BigM[constr] = 1
# Note: we use a default big M value of 1
# because all non-selected disjuncts should be deactivated.
# Therefore, none of the big M transformed nonlinear constraints will need to be relaxed.
# The default M value should therefore be irrelevant.
child.GDPbb_utils.deactivated_constraints -= newly_activated
child.GDPbb_utils.disjuncts_fixed_True.add(child_disjunct)
if disjunct in incumbent_model.GDPbb_utils.disjuncts_fixed_True:
# If the disjunct was already branched to True from a parent disjunct branching, just pass
# through the incumbent value without resolving. The solution should be the same as the parent.
total_nodes_counter += 1
ordering_tuple = (obj_sign * incumbent_obj_value, disjunctions_left - 1, -total_nodes_counter)
heapq.heappush(heap, (ordering_tuple, child, result, incumbent_var_values))
new_nodes_counter += 1
continue
if config.check_sat and satisfiable(child, config.logger) is False:
# Problem is not satisfiable. Skip this disjunct.
continue
obj_value, result, var_values = self.subproblem_solve(child, config)
total_nodes_counter += 1
ordering_tuple = (obj_sign * obj_value, disjunctions_left - 1, -total_nodes_counter)
heapq.heappush(heap, (ordering_tuple, child, result, var_values))
new_nodes_counter += 1
config.logger.info("Added %s new nodes with %s relaxed disjunctions to the heap. Size now %s." % (
new_nodes_counter, disjunctions_left - 1, len(heap)))
def add_lazy_oa_cuts(self,
target_model,
dual_values,
solve_data,
config,
opt,
linearize_active=True,
linearize_violated=True):
"""Linearizes nonlinear constraints; add the OA cuts through Cplex inherent function self.add()
For nonconvex problems, turn on 'config.add_slack'. Slack variables will always be used for
nonlinear equality constraints.
Parameters
----------
target_model : Pyomo model
The MIP main problem.
dual_values : list
The value of the duals for each constraint.
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
opt : SolverFactory
The cplex_persistent solver.
linearize_active : bool, optional
Whether to linearize the active nonlinear constraints, by default True.
linearize_violated : bool, optional
Whether to linearize the violated nonlinear constraints, by default True.
"""
config.logger.debug('Adding OA cuts')
with time_code(solve_data.timing, 'OA cut generation'):
for index, constr in enumerate(
target_model.MindtPy_utils.constraint_list):
if constr.body.polynomial_degree(
) in solve_data.mip_constraint_polynomial_degree:
continue
constr_vars = list(identify_variables(constr.body))
jacs = solve_data.jacobians
# Equality constraint (makes the problem nonconvex)
if constr.has_ub() and constr.has_lb() and value(
constr.lower) == value(constr.upper):
sign_adjust = -1 if solve_data.objective_sense == minimize else 1
rhs = constr.lower
# since the cplex requires the lazy cuts in cplex type, we need to transform the pyomo expression into cplex expression
pyomo_expr = copysign(
1, sign_adjust *
dual_values[index]) * (sum(
value(jacs[constr][var]) * (var - value(var))
for var in EXPR.identify_variables(constr.body)) +
value(constr.body) - rhs)
cplex_expr, _ = opt._get_expr_from_pyomo_expr(pyomo_expr)
cplex_rhs = -generate_standard_repn(pyomo_expr).constant
self.add(constraint=cplex.SparsePair(
ind=cplex_expr.variables, val=cplex_expr.coefficients),
sense='L',
rhs=cplex_rhs)
else: # Inequality constraint (possibly two-sided)
if (constr.has_ub() and
(linearize_active
and abs(constr.uslack()) < config.zero_tolerance) or
(linearize_violated and constr.uslack() < 0) or
(config.linearize_inactive and constr.uslack() > 0)
) or ('MindtPy_utils.objective_constr' in constr.name
and constr.has_ub()):
pyomo_expr = sum(
value(jacs[constr][var]) * (var - var.value)
for var in constr_vars) + value(constr.body)
cplex_rhs = - \
generate_standard_repn(pyomo_expr).constant
cplex_expr, _ = opt._get_expr_from_pyomo_expr(
pyomo_expr)
self.add(constraint=cplex.SparsePair(
ind=cplex_expr.variables,
val=cplex_expr.coefficients),
sense='L',
rhs=value(constr.upper) + cplex_rhs)
if (constr.has_lb() and
(linearize_active
and abs(constr.lslack()) < config.zero_tolerance) or
(linearize_violated and constr.lslack() < 0) or
(config.linearize_inactive and constr.lslack() > 0)
) or ('MindtPy_utils.objective_constr' in constr.name
and constr.has_lb()):
pyomo_expr = sum(
value(jacs[constr][var]) * (var - self.get_values(
opt._pyomo_var_to_solver_var_map[var]))
for var in constr_vars) + value(constr.body)
cplex_rhs = - \
generate_standard_repn(pyomo_expr).constant
cplex_expr, _ = opt._get_expr_from_pyomo_expr(
pyomo_expr)
self.add(constraint=cplex.SparsePair(
ind=cplex_expr.variables,
val=cplex_expr.coefficients),
sense='G',
rhs=value(constr.lower) + cplex_rhs)
def solve(self, model, **kwds):
config = self.CONFIG(kwds.pop('options', {}))
config.set_value(kwds)
# Validate model to be used with gdpbb
self.validate_model(model)
# Set solver as an MINLP
solver = SolverFactory(config.solver)
solve_data = GDPbbSolveData()
solve_data.timing = Container()
solve_data.original_model = model
solve_data.results = SolverResults()
old_logger_level = config.logger.getEffectiveLevel()
with time_code(solve_data.timing, 'total'), \
restore_logger_level(config.logger), \
create_utility_block(model, 'GDPbb_utils', solve_data):
if config.tee and old_logger_level > logging.INFO:
# If the logger does not already include INFO, include it.
config.logger.setLevel(logging.INFO)
config.logger.info(
"Starting GDPbb version %s using %s as subsolver"
% (".".join(map(str, self.version())), config.solver)
)
# Setup results
solve_data.results.solver.name = 'GDPbb - %s' % (str(config.solver))
setup_results_object(solve_data, config)
# Initialize list containing indicator vars for reupdating model after solving
indicator_list_name = unique_component_name(model, "_indicator_list")
indicator_vars = []
for disjunction in model.component_data_objects(
ctype=Disjunction, active=True):
for disjunct in disjunction.disjuncts:
indicator_vars.append(disjunct.indicator_var)
setattr(model, indicator_list_name, indicator_vars)
# get objective sense
objectives = model.component_data_objects(Objective, active=True)
obj = next(objectives, None)
obj_sign = 1 if obj.sense == minimize else -1
solve_data.results.problem.sense = obj.sense
# clone original model for root node of branch and bound
root = model.clone()
# set up lists to keep track of which disjunctions have been covered.
# this list keeps track of the original disjunctions that were active and are soon to be inactive
root.GDPbb_utils.unenforced_disjunctions = list(
disjunction for disjunction in root.GDPbb_utils.disjunction_list if disjunction.active
)
# this list keeps track of the disjunctions that have been activated by the branch and bound
root.GDPbb_utils.curr_active_disjunctions = []
# deactivate all disjunctions in the model
# self.indicate(root)
for djn in root.GDPbb_utils.unenforced_disjunctions:
djn.deactivate()
# Deactivate all disjuncts in model. To be reactivated when disjunction
# is reactivated.
for disj in root.component_data_objects(Disjunct, active=True):
disj._deactivate_without_fixing_indicator()
# Satisfiability check would go here
# solve the root node
config.logger.info("Solving the root node.")
obj_value, result, _ = self.subproblem_solve(root, solver, config)
# initialize minheap for Branch and Bound algorithm
# Heap structure: (ordering tuple, model)
# Ordering tuple: (objective value, disjunctions_left, -counter)
# - select solutions with lower objective value,
# then fewer disjunctions left to explore (depth first),
# then more recently encountered (tiebreaker)
heap = []
counter = 0
disjunctions_left = len(root.GDPbb_utils.unenforced_disjunctions)
heapq.heappush(heap,
((obj_sign * obj_value, disjunctions_left, -counter), root,
result, root.GDPbb_utils.variable_list))
# loop to branch through the tree
while len(heap) > 0:
# pop best model off of heap
sort_tup, mdl, mdl_results, vars = heapq.heappop(heap)
old_obj_val, disjunctions_left, _ = sort_tup
config.logger.info("Exploring node with LB %.10g and %s inactive disjunctions." % (
old_obj_val, disjunctions_left
))
# if all the originally active disjunctions are active, solve and
# return solution
if disjunctions_left == 0:
config.logger.info("Model solved.")
# Model is solved. Copy over solution values.
for orig_var, soln_var in zip(model.GDPbb_utils.variable_list, vars):
orig_var.value = soln_var.value
solve_data.results.problem.lower_bound = mdl_results.problem.lower_bound
solve_data.results.problem.upper_bound = mdl_results.problem.upper_bound
solve_data.results.solver.timing = solve_data.timing
solve_data.results.solver.termination_condition = mdl_results.solver.termination_condition
return solve_data.results
next_disjunction = mdl.GDPbb_utils.unenforced_disjunctions.pop(0)
config.logger.info("Activating disjunction %s" % next_disjunction.name)
next_disjunction.activate()
mdl.GDPbb_utils.curr_active_disjunctions.append(next_disjunction)
djn_left = len(mdl.GDPbb_utils.unenforced_disjunctions)
for disj in next_disjunction.disjuncts:
disj._activate_without_unfixing_indicator()
if not disj.indicator_var.fixed:
disj.indicator_var = 0 # initially set all indicator vars to zero
added_disj_counter = 0
for disj in next_disjunction.disjuncts:
if not disj.indicator_var.fixed:
disj.indicator_var = 1
mnew = mdl.clone()
if not disj.indicator_var.fixed:
disj.indicator_var = 0
# Check feasibility
if config.check_sat and satisfiable(mnew, config.logger) is False:
# problem is not satisfiable. Skip this disjunct.
continue
obj_value, result, vars = self.subproblem_solve(mnew, solver, config)
counter += 1
ordering_tuple = (obj_sign * obj_value, djn_left, -counter)
heapq.heappush(heap, (ordering_tuple, mnew, result, vars))
added_disj_counter = added_disj_counter + 1
config.logger.info("Added %s new nodes with %s relaxed disjunctions to the heap. Size now %s." % (
added_disj_counter, djn_left, len(heap)))
def generate_lag_objective_function(model,
setpoint_model,
config,
solve_data,
discrete_only=False):
"""The function generate taylor extension of the Lagrangean function.
Args:
model ([type]): [description]
setpoint_model ([type]): [description]
discrete_only (bool, optional): [description]. Defaults to False.
"""
temp_model = setpoint_model.clone()
for var in temp_model.MindtPy_utils.variable_list:
if var.is_integer():
var.unfix()
# objective_list[0] is the original objective function, not in MindtPy_utils block
temp_model.MindtPy_utils.objective_list[0].activate()
temp_model.MindtPy_utils.deactivate()
TransformationFactory('core.relax_integer_vars').apply_to(temp_model)
# Note: PyNumero does not support discrete variables
# So PyomoNLP should operate on setpoint_model
# Implementation 1
# First calculate Jacobian and Hessian without assigning variable and constraint sequence, then use get_primal_indices to get the indices.
with time_code(solve_data.timing, 'PyomoNLP'):
nlp = pyomo_nlp.PyomoNLP(temp_model)
lam = [
-temp_model.dual[constr]
if abs(temp_model.dual[constr]) > config.zero_tolerance else 0
for constr in nlp.get_pyomo_constraints()
]
nlp.set_duals(lam)
obj_grad = nlp.evaluate_grad_objective().reshape(-1, 1)
jac = nlp.evaluate_jacobian().toarray()
jac_lag = obj_grad + jac.transpose().dot(
numpy.array(lam).reshape(-1, 1))
jac_lag[abs(jac_lag) < config.zero_tolerance] = 0
# jac_lag of continuous variables should be zero
for var in temp_model.MindtPy_utils.continuous_variable_list[:-1]:
jac_lag[nlp.get_primal_indices([var])[0]] = 0
nlp_var = set([i.name for i in nlp.get_pyomo_variables()])
first_order_term = sum(
float(jac_lag[nlp.get_primal_indices([temp_var])[0]]) *
(var - temp_var.value) for var, temp_var in zip(
model.MindtPy_utils.variable_list[:-1],
temp_model.MindtPy_utils.variable_list[:-1])
if temp_var.name in nlp_var)
if config.add_regularization == 'grad_lag':
return Objective(expr=first_order_term, sense=minimize)
elif config.add_regularization in {'hess_lag', 'hess_only_lag'}:
# Implementation 1
hess_lag = nlp.evaluate_hessian_lag().toarray()
hess_lag[abs(hess_lag) < config.zero_tolerance] = 0
second_order_term = 0.5 * sum(
(var_i - temp_var_i.value) *
float(hess_lag[nlp.get_primal_indices([temp_var_i])[0]][
nlp.get_primal_indices([temp_var_j])[0]]) *
(var_j - temp_var_j.value) for var_i, temp_var_i in zip(
model.MindtPy_utils.variable_list[:-1],
temp_model.MindtPy_utils.variable_list[:-1])
for var_j, temp_var_j in zip(
model.MindtPy_utils.variable_list[:-1],
temp_model.MindtPy_utils.variable_list[:-1])
if (temp_var_i.name in nlp_var and temp_var_j.name in nlp_var))
if config.add_regularization == 'hess_lag':
return Objective(expr=first_order_term + second_order_term,
sense=minimize)
elif config.add_regularization == 'hess_only_lag':
return Objective(expr=second_order_term, sense=minimize)
elif config.add_regularization == 'sqp_lag':
var_filter = (lambda v: v[1].is_integer()) if discrete_only \
else (lambda v: v[1].name != 'MindtPy_utils.objective_value' and
'MindtPy_utils.feas_opt.slack_var' not in v[1].name)
model_vars, setpoint_vars = zip(*filter(
var_filter,
zip(model.component_data_objects(Var),
setpoint_model.component_data_objects(Var))))
assert len(model_vars) == len(
setpoint_vars
), 'Trying to generate Squared Norm2 objective function for models with different number of variables'
if config.sqp_lag_scaling_coef is None:
rho = 1
elif config.sqp_lag_scaling_coef == 'fixed':
r = 1
rho = numpy.linalg.norm(jac_lag / (2 * r))
elif config.sqp_lag_scaling_coef == 'variable_dependent':
r = numpy.sqrt(
len(temp_model.MindtPy_utils.discrete_variable_list))
rho = numpy.linalg.norm(jac_lag / (2 * r))
return Objective(
expr=first_order_term + rho *
sum([(model_var - setpoint_var.value)**2
for (model_var,
setpoint_var) in zip(model_vars, setpoint_vars)]))
def add_lazy_affine_cuts(self, solve_data, config, opt):
"""Adds affine cuts using MCPP.
Add affine cuts through Cplex inherent function self.add().
Parameters
----------
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
opt : SolverFactory
The cplex_persistent solver.
"""
with time_code(solve_data.timing, 'Affine cut generation'):
m = solve_data.mip
config.logger.debug('Adding affine cuts')
counter = 0
for constr in m.MindtPy_utils.nonlinear_constraint_list:
vars_in_constr = list(identify_variables(constr.body))
if any(var.value is None for var in vars_in_constr):
continue # a variable has no values
# mcpp stuff
try:
mc_eqn = mc(constr.body)
except MCPP_Error as e:
config.logger.debug(
'Skipping constraint %s due to MCPP error %s' %
(constr.name, str(e)))
continue # skip to the next constraint
# TODO: check if the value of ccSlope and cvSlope is not Nan or inf. If so, we skip this.
ccSlope = mc_eqn.subcc()
cvSlope = mc_eqn.subcv()
ccStart = mc_eqn.concave()
cvStart = mc_eqn.convex()
concave_cut_valid = True
convex_cut_valid = True
for var in vars_in_constr:
if not var.fixed:
if ccSlope[var] == float(
'nan') or ccSlope[var] == float('inf'):
concave_cut_valid = False
if cvSlope[var] == float(
'nan') or cvSlope[var] == float('inf'):
convex_cut_valid = False
if ccStart == float('nan') or ccStart == float('inf'):
concave_cut_valid = False
if cvStart == float('nan') or cvStart == float('inf'):
convex_cut_valid = False
# check if the value of ccSlope and cvSlope all equals zero. if so, we skip this.
if not any(ccSlope.values()):
concave_cut_valid = False
if not any(cvSlope.values()):
convex_cut_valid = False
if not (concave_cut_valid or convex_cut_valid):
continue
ub_int = min(
value(constr.upper),
mc_eqn.upper()) if constr.has_ub() else mc_eqn.upper()
lb_int = max(
value(constr.lower),
mc_eqn.lower()) if constr.has_lb() else mc_eqn.lower()
if concave_cut_valid:
pyomo_concave_cut = sum(ccSlope[var] * (var - var.value)
for var in vars_in_constr
if not var.fixed) + ccStart
cplex_concave_rhs = generate_standard_repn(
pyomo_concave_cut).constant
cplex_concave_cut, _ = opt._get_expr_from_pyomo_expr(
pyomo_concave_cut)
self.add(constraint=cplex.SparsePair(
ind=cplex_concave_cut.variables,
val=cplex_concave_cut.coefficients),
sense='G',
rhs=lb_int - cplex_concave_rhs)
counter += 1
if convex_cut_valid:
pyomo_convex_cut = sum(cvSlope[var] * (var - var.value)
for var in vars_in_constr
if not var.fixed) + cvStart
cplex_convex_rhs = generate_standard_repn(
pyomo_convex_cut).constant
cplex_convex_cut, _ = opt._get_expr_from_pyomo_expr(
pyomo_convex_cut)
self.add(constraint=cplex.SparsePair(
ind=cplex_convex_cut.variables,
val=cplex_convex_cut.coefficients),
sense='L',
rhs=ub_int - cplex_convex_rhs)
counter += 1
config.logger.info('Added %s affine cuts' % counter)
def add_lazy_no_good_cuts(self,
var_values,
solve_data,
config,
opt,
feasible=False):
"""Adds no-good cuts.
Add the no-good cuts through Cplex inherent function self.add().
Parameters
----------
var_values : list
The variable values of the incumbent solution, used to generate the cut.
solve_data : MindtPySolveData
Data container that holds solve-instance data.
config : ConfigBlock
The specific configurations for MindtPy.
opt : SolverFactory
The cplex_persistent solver.
feasible : bool, optional
Whether the integer combination yields a feasible or infeasible NLP, by default False.
Raises
------
ValueError
The value of binary variable is not 0 or 1.
"""
if not config.add_no_good_cuts:
return
config.logger.info('Adding no-good cuts')
with time_code(solve_data.timing, 'No-good cut generation'):
m = solve_data.mip
MindtPy = m.MindtPy_utils
int_tol = config.integer_tolerance
binary_vars = [v for v in MindtPy.variable_list if v.is_binary()]
# copy variable values over
for var, val in zip(MindtPy.variable_list, var_values):
if not var.is_binary():
continue
# We don't want to trigger the reset of the global stale
# indicator, so we will set this variable to be "stale",
# knowing that set_value will switch it back to "not
# stale"
var.stale = True
var.set_value(val, skip_validation=True)
# check to make sure that binary variables are all 0 or 1
for v in binary_vars:
if value(abs(v - 1)) > int_tol and value(abs(v)) > int_tol:
raise ValueError('Binary {} = {} is not 0 or 1'.format(
v.name, value(v)))
if not binary_vars: # if no binary variables, skip
return
pyomo_no_good_cut = sum(
1 - v
for v in binary_vars if value(abs(v - 1)) <= int_tol) + sum(
v for v in binary_vars if value(abs(v)) <= int_tol)
cplex_no_good_rhs = generate_standard_repn(
pyomo_no_good_cut).constant
cplex_no_good_cut, _ = opt._get_expr_from_pyomo_expr(
pyomo_no_good_cut)
self.add(constraint=cplex.SparsePair(
ind=cplex_no_good_cut.variables,
val=cplex_no_good_cut.coefficients),
sense='G',
rhs=1 - cplex_no_good_rhs)