def _create_using(self, model, **kwds):
"""
Force all variables to lie in the nonnegative orthant.
Required arguments:
model The model to transform.
Optional keyword arguments:
pos_suffix The suffix applied to the 'positive' component of
converted variables. Default is '_plus'.
neg_suffix The suffix applied to the 'positive' component of
converted variables. Default is '_minus'.
"""
#
# Optional naming schemes
#
pos_suffix = kwds.pop("pos_suffix", "_plus")
neg_suffix = kwds.pop("neg_suffix", "_minus")
#
# We first perform an abstract problem transformation. Then, if model
# data is available, we instantiate the new model. If not, we construct
# a mapping that can later be used to populate the new model.
#
nonneg = model.clone()
components = collectAbstractComponents(nonneg)
# Map from variable base names to a {index, rule} map
constraint_rules = {}
# Map from variable base names to a rule defining the domains for that
# variable
domain_rules = {}
# Map from variable base names to its set of indices
var_indices = {}
# Map from fully qualified variable names to replacement expressions.
# For now, it is actually a map from a variable name to a closure that
# must later be evaulated with a model containing the replacement
# variables.
var_map = {}
#
# Get the constraints that enforce the bounds and domains of each
# variable
#
for var_name in components["Var"]:
var = nonneg.__getattribute__(var_name)
# Individual bounds and domains
orig_bounds = {}
orig_domain = {}
# New indices
indices = set()
# Map from constraint names to a constraint rule.
constraints = {}
# Map from variable indices to a domain
domains = {}
for ndx in var:
# Fully qualified variable name
vname = create_name(str(var_name), ndx)
# We convert each index to a string to avoid difficult issues
# regarding appending a suffix to tuples.
#
# If the index is None, this casts the index to a string,
# which doesn't match up with how Pyomo treats None indices
# internally. Replace with "" to be consistent.
if ndx is None:
v_ndx = ""
else:
v_ndx = str(ndx)
# Get the variable bounds
lb = value(var[ndx].lb)
ub = value(var[ndx].ub)
orig_bounds[ndx] = (lb, ub)
# Get the variable domain
if var[ndx].domain is not None:
orig_domain[ndx] = var[ndx].domain
else:
orig_domain[ndx] = var.domain
# Determine the replacement expression. Either a new single
# variable with the same attributes, or a sum of two new
# variables.
#
# If both the bounds and domain allow for negative values,
# replace the variable with the sum of nonnegative ones.
bounds_neg = (orig_bounds[ndx] == (None, None)
or orig_bounds[ndx][0] is None
or orig_bounds[ndx][0] < 0)
domain_neg = (orig_domain[ndx] is None
or orig_domain[ndx].bounds()[0] is None
or orig_domain[ndx].bounds()[0] < 0)
if bounds_neg and domain_neg:
# Make two new variables.
posVarSuffix = "%s%s" % (v_ndx, pos_suffix)
negVarSuffix = "%s%s" % (v_ndx, neg_suffix)
new_indices = (posVarSuffix, negVarSuffix)
# Replace the original variable with a sum expression
expr_dict = {posVarSuffix: 1, negVarSuffix: -1}
else:
# Add the new index. Lie if is 'None', since Pyomo treats
# 'None' specially as a key.
#
# More lies: don't let a blank index exist. Replace it with
# '_'. I don't actually have a justification for this other
# than that allowing "" as a key will eventually almost
# certainly lead to a strange bug.
if v_ndx is None:
t_ndx = "None"
elif v_ndx == "":
t_ndx = "_"
else:
t_ndx = v_ndx
new_indices = (t_ndx, )
# Replace the original variable with a sum expression
expr_dict = {t_ndx: 1}
# Add the new indices
for x in new_indices:
indices.add(x)
# Replace the original variable with an expression
var_map[vname] = partial(self.sumRule, var_name, expr_dict)
# Enforce bounds as constraints
if orig_bounds[ndx] != (None, None):
cname = "%s_%s" % (vname, "bounds")
tmp = orig_bounds[ndx]
constraints[cname] = partial(self.boundsConstraintRule,
tmp[0], tmp[1], var_name,
expr_dict)
# Enforce the bounds of the domain as constraints
if orig_domain[ndx] != None:
cname = "%s_%s" % (vname, "domain_bounds")
tmp = orig_domain[ndx].bounds()
constraints[cname] = partial(self.boundsConstraintRule,
tmp[0], tmp[1], var_name,
expr_dict)
# Domain will either be NonNegativeReals, NonNegativeIntegers,
# or Binary. We consider Binary because some solvers may
# optimize over binary variables.
if var[ndx].is_continuous():
for x in new_indices:
domains[x] = NonNegativeReals
elif var[ndx].is_binary():
for x in new_indices:
domains[x] = Binary
elif var[ndx].is_integer():
for x in new_indices:
domains[x] = NonNegativeIntegers
else:
logger.warning("Warning: domain '%s' not recognized, "
"defaulting to 'NonNegativeReals'" %
(var.domain, ))
for x in new_indices:
domains[x] = NonNegativeReals
constraint_rules[var_name] = constraints
domain_rules[var_name] = partial(self.exprMapRule, domains)
var_indices[var_name] = indices
# Remove all existing variables.
toRemove = []
for (attr_name, attr) in nonneg.__dict__.items():
if isinstance(attr, Var):
toRemove.append(attr_name)
for attr_name in toRemove:
nonneg.__delattr__(attr_name)
# Add the sets defining the variables, then the variables
for (k, v) in var_indices.items():
sname = "%s_indices" % k
nonneg.__setattr__(sname, Set(initialize=v))
nonneg.__setattr__(
k,
Var(nonneg.__getattribute__(sname),
domain=domain_rules[k],
bounds=(0, None)))
# Construct the model to get the variables and their indices
# recognized in the model
##nonneg = nonneg.create()
# Safe to evaluate the modifiedVars mapping
for var in var_map:
var_map[var] = var_map[var](nonneg)
# Map from constraint base names to maps from indices to expressions
constraintExprs = {}
#
# Convert all modified variables in all constraints in the original
# problem
#
for conName in components["Constraint"]:
con = nonneg.__getattribute__(conName)
# Map from constraint indices to a corrected expression
exprMap = {}
for (ndx, cdata) in con._data.items():
lower = _walk_expr(cdata.lower, var_map)
body = _walk_expr(cdata.body, var_map)
upper = _walk_expr(cdata.upper, var_map)
# Lie if ndx is None. Pyomo treats 'None' indices specially.
if ndx is None:
ndx = "None"
# Cast indices to strings, otherwise tuples ruin everything
exprMap[str(ndx)] = (lower, body, upper)
# Add to list of expression maps
constraintExprs[conName] = exprMap
# Map from constraint base names to maps from indices to expressions
objectiveExprs = {}
#
# Convert all modified variables in all objectives in the original
# problem
#
for objName in components["Objective"]:
obj = nonneg.__getattribute__(objName)
# Map from objective indices to a corrected expression
exprMap = {}
for (ndx, odata) in obj._data.items():
exprMap[ndx] = _walk_expr(odata.expr, var_map)
# Add to list of expression maps
objectiveExprs[objName] = exprMap
# Make the modified original constraints
for (conName, ruleMap) in constraintExprs.items():
# Make the set of indices
sname = conName + "_indices"
_set = Set(initialize=ruleMap.keys())
nonneg.__setattr__(sname, _set)
_set.construct()
# Define the constraint
_con = Constraint(nonneg.__getattribute__(sname),
rule=partial(self.exprMapRule, ruleMap))
nonneg.__setattr__(conName, _con)
_con.construct()
# Make the bounds constraints
for (varName, ruleMap) in constraint_rules.items():
conName = varName + "_constraints"
# Make the set of indices
sname = conName + "_indices"
_set = Set(initialize=ruleMap.keys())
nonneg.__setattr__(sname, _set)
_set.construct()
# Define the constraint
_con = Constraint(nonneg.__getattribute__(sname),
rule=partial(self.delayedExprMapRule, ruleMap))
nonneg.__setattr__(conName, _con)
_con.construct()
# Make the objectives
for (objName, ruleMap) in objectiveExprs.items():
# Make the set of indices
sname = objName + "_indices"
_set = Set(initialize=ruleMap.keys())
nonneg.__setattr__(sname, _set)
_set.construct()
# Define the constraint
_obj = Objective(nonneg.__getattribute__(sname),
rule=partial(self.exprMapRule, ruleMap))
nonneg.__setattr__(objName, _obj)
_obj.construct()
return nonneg
def form_linearized_objective_constraints(instance_name,
instance,
scenario_tree,
linearize_nonbinary_penalty_terms,
breakpoint_strategy,
tolerance):
# keep track and return what was added to the instance, so
# it can be cleaned up if necessary.
new_instance_attributes = []
linearization_index_set_name = "PH_LINEARIZATION_INDEX_SET"
linearization_index_set = instance.find_component(linearization_index_set_name)
if linearization_index_set is None:
linearization_index_set = Set(initialize=range(0, linearize_nonbinary_penalty_terms*2), dimen=1, name=linearization_index_set_name)
instance.add_component(linearization_index_set_name, linearization_index_set)
scenario = scenario_tree.get_scenario(instance_name)
nodeid_to_vardata_map = instance._ScenarioTreeSymbolMap.bySymbol
for tree_node in scenario._node_list[:-1]:
xbar_dict = tree_node._xbars
# if linearizing, then we have previously defined a variable
# associated with the result of the linearized approximation
# of the penalty term - this is simply added to the objective
# function.
linearized_cost_variable_name = "PHQUADPENALTY_"+str(tree_node._name)
linearized_cost_variable = instance.find_component(linearized_cost_variable_name)
# grab the linearization constraint associated with the
# linearized cost variable, if it exists. otherwise, create it
# - but an empty variety. the constraints are stage-specific -
# we could index by constraint, but we don't know if that is
# really worth the additional effort.
linearization_constraint_name = "PH_LINEARIZATION_"+str(tree_node._name)
linearization_constraint = instance.find_component(linearization_constraint_name)
if linearization_constraint is not None:
# clear whatever constraint components are there - there
# may be fewer breakpoints, due to tolerances, and we
# don't want to the old pieces laying around.
linearization_constraint.clear()
else:
# this is the first time the constraint is being added -
# add it to the list of PH-specific constraints for this
# instance.
new_instance_attributes.append(linearization_constraint_name)
nodal_index_set_name = "PHINDEX_"+str(tree_node._name)
nodal_index_set = instance.find_component(nodal_index_set_name)
assert nodal_index_set is not None
linearization_constraint = \
Constraint(nodal_index_set,
linearization_index_set,
name=linearization_constraint_name)
linearization_constraint.construct()
instance.add_component(linearization_constraint_name, linearization_constraint)
for variable_id in tree_node._variable_ids:
# don't add weight terms for derived variables at the tree
# node.
if variable_id in tree_node._derived_variable_ids:
continue
if variable_id not in tree_node._minimums:
variable_name, index = tree_node._variable_ids[variable_id]
raise RuntimeError("No minimum value statistic found for variable=%s "
"on tree node=%s; cannot form linearized PH objective"
% (variable_name+indexToString(index),
tree_node._name))
if variable_id not in tree_node._maximums:
variable_name, index = tree_node._variable_ids[variable_id]
raise RuntimeError("No maximums value statistic found for "
"variable=%s on tree node=%s; cannot "
"form linearized PH objective"
% (variable_name+indexToString(index),
tree_node._name))
xbar = xbar_dict[variable_id]
node_min = tree_node._minimums[variable_id]
node_max = tree_node._maximums[variable_id]
instance_vardata = nodeid_to_vardata_map[variable_id]
if (instance_vardata.stale is False) and (instance_vardata.fixed is False):
# binaries have already been dealt with in the process of PH objective function formation.
if isinstance(instance_vardata.domain, BooleanSet) is False:
x = instance_vardata
if x.lb is None or x.ub is None:
msg = "Missing bound for variable '%s'\n" \
'Both lower and upper bounds required when' \
' piece-wise approximating quadratic ' \
'penalty terms'
raise ValueError(msg % instance_vardata.name)
lb = value(x.lb)
ub = value(x.ub)
# compute the breakpoint sequence according to the specified strategy.
try:
strategy = (compute_uniform_breakpoints,
compute_uniform_between_nodestat_breakpoints,
compute_uniform_between_woodruff_breakpoints,
compute_exponential_from_mean_breakpoints,
)[ breakpoint_strategy ]
args = ( lb, node_min, xbar, node_max, ub, \
linearize_nonbinary_penalty_terms, tolerance )
breakpoints = strategy( *args )
except ValueError:
e = sys.exc_info()[1]
msg = 'A breakpoint distribution strategy (%s) ' \
'is currently not supported within PH!'
raise ValueError(msg % breakpoint_strategy)
for i in xrange(len(breakpoints)-1):
this_lb = breakpoints[i]
this_ub = breakpoints[i+1]
segment_tuple = create_piecewise_constraint_tuple(this_lb,
this_ub,
x,
xbar,
linearized_cost_variable[variable_id],
tolerance)
linearization_constraint.add((variable_id,i), segment_tuple)
return new_instance_attributes
