def get_solution(self, m):
solution = Candidate()
# f = open('sol.txt', 'w')
for var_a in m.getVars():
vname = var_a.getAttr(GRB.Attr.VarName)
if self.objective_type == ObjectiveType.MAX_FLEX_UNCERTAINTY:
if vname.find("cl") == -1 and vname.find("cu") == -1:
continue
#f.write('%s %s\n' % (var_a.getAttr(GRB.Attr.VarName), var_a.getAttr(GRB.Attr.X)))
# if vname.find("l") != -1:
# print(var_a.getAttr(GRB.Attr.VarName), var_a.getAttr(GRB.Attr.X))
e = self.relax_e[vname]
new_relaxation = TemporalRelaxation(e)
if vname.find("cl") != -1:
new_relaxation.relaxed_lb = var_a.getAttr("X")
if new_relaxation.relaxed_lb != e.get_lower_bound():
solution.add_temporal_relaxation(new_relaxation)
else :
new_relaxation.relaxed_ub = var_a.getAttr("X")
if new_relaxation.relaxed_ub != e.get_upper_bound():
solution.add_temporal_relaxation(new_relaxation)
elif self.objective_type == ObjectiveType.MIN_COST:
if vname.find("vlr") == -1 and vname.find("vur") == -1:
continue;
e = self.relax_e[vname]
#print(vname,e.fro, e.to)
new_relaxation = TemporalRelaxation(e)
if vname.find("vlr") != -1:
new_relaxation.relaxed_lb = self.l[(e.fro,e.to)].getValue()
else:
new_relaxation.relaxed_ub = self.u[(e.fro, e.to)].getValue()
if var_a.getAttr("X") > 0:
solution.add_temporal_relaxation(new_relaxation)
# for e in self.network.temporal_constraints.values():
# if e.activated:
# print(e.fro, e.to, self.l[(e.fro, e.to)].getValue(), self.u[(e.fro, e.to)].getValue())
solution.utility = round(m.getObjective().getValue(), 6)
if solution.utility > 1000 and self.objective_type == ObjectiveType.MAX_FLEX_UNCERTAINTY:
solution.utility = 1000
# for node_a in range(1, self.num_nodes+1):
# for node_b in range(1, self.num_nodes+1):
# if node_a != node_b:
# f.write("%s->%s [%s, %s]\n"%(node_a,node_b,self.l[(node_a,node_b)].getValue(), self.u[(node_a,node_b)].getValue() ))
# # if solution.utility > 100:
# solution.utility = 100
# f.close()
#print a dot file
# f = open('sol.dot', 'w')
# f.write('digraph G {\n nodesep = .45; \n size = 30;\n label="CCTP";\n')
#
# for e in self.network.temporal_constraints.values():
# if e.activated:
# f.write('"%s"->"%s"'%(e.fro, e.to))
# tlb = self.l[(e.fro, e.to)].getValue()
# tub = self.u[(e.fro, e.to)].getValue()
# f.write('[ label = "[%s,%s]'%(tlb, tub))
# if not e.controllable:
# f.write(' type=dashed')
# f.write('"];\n')
# #%d"];'%(e.fro, e.to,e.get_lower_bound(),e.get_upper_bound(),e.controllable))
# f.write("}")
# f.close()
return solution
def generate_maxflex_relaxations(candidate,negative_cycle):
if len(negative_cycle.constraints) == 0:
return None,0
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# Solve using PuLP, TODO, incorporate interface to other solvers,
# especially for nonlinear objective functions
prob = LpProblem("MaxFlexConflictResolution", LpMaximize)
# Solve using PyOpt/Snopt,
# especially for nonlinear constraints/objective functions
# construct variables and constraints
lp_variables = {}
lp_objective = []
max_flex_variable = LpVariable("Max Flex",0, None)
lp_objective.append(max_flex_variable)
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1;
# We only consider the relaxation of the upper bound of
# uncontrollable duration.
# which means that we basically ignore the definition of
# 'relaxable bounds' in the network, and only consider the upper bounds
# of uncertain durations.
for cycle in all_cycles:
lp_constraint = []
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in lp_variables:
variable = lp_variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
variable = constraint.lower_bound
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if not constraint.controllable:
variable = LpVariable(constraint.id + "-UB",constraint.lower_bound, constraint.upper_bound)
lp_variables[(constraint,bound)] = variable
# add the variable to the objective function
max_flex_constraint = []
max_flex_constraint.append((variable-constraint.lower_bound))
max_flex_constraint.append(-1.0*max_flex_variable)
prob += sum(max_flex_constraint) >= 0
else:
variable = constraint.upper_bound
assert variable is not None
lp_constraint.append(variable*coefficient)
# add the constraint to the problem
# print(str(lp_constraint))
if sum(lp_constraint) >= 0:
# print(str(sum(lp_constraint)) + " >= 0")
prob += sum(lp_constraint) >= 0
else:
status = 0;
# this is not resolvable
# no need to proceed
if status > 0:
# Set the objective function
prob += sum(lp_objective)
# for c in prob.constraints:
# print("CON: ", prob.constraints[c])
# print("OBJ: ", prob.objective)
# Solve the problem
try:
import gurobipy
status = prob.solve(solvers.GUROBI(mip=False,msg=False))
except ImportError:
pass # Gurobi doesn't exist, use default Pulp solver.
status = prob.solve()
# exit(0);
# if no solution was found, do nothing
if status > 0:
# A solution has been bound
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
max_flex_value = value(max_flex_variable)
for constraint, bound in lp_variables.keys():
variable = lp_variables[(constraint, bound)]
relaxed_bound = value(variable)
if bound == 0:
# check if this constraint bound is relaxed
if relaxed_bound != constraint.lower_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
elif bound == 1:
# same for upper bound
if relaxed_bound != constraint.upper_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
if len(relaxations) > 0:
return relaxations,max_flex_value
return None,0
def generate_cc_relaxations(candidate,negative_cycle,feasibility_type,cc,pInfinity=1e6,nInfinity=-1e6):
global prob_means
global prob_stds
global prob_vars
global cc_var
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# print("CC Solving against " + str(len(all_cycles)) + " cycles")
numVar = 0
numConstraint = 0
# Solve using Snopt,
# especiallFy for nonlinear constraints/objective functions
# construct variables and constraints
variables = {}
variable_bounds = {}
initial_values = {}
constraints = []
objective = {}
list_iA = []
list_jA = []
list_A = []
neA_count = 0
list_iG = []
list_jG = []
neG_count = 0
prob_durations = set()
prob_means = []
prob_stds = []
prob_vars = []
cc_var = -1
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1;
# print("Cycles: " + str(len(all_cycles)))
for cycle in all_cycles:
lp_ncycle_constraint = {}
rhs = 0
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in variables:
variable = variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
# lower bound, the domain is less than the original LB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_lb or (not constraint.controllable and constraint.probabilistic):
# if constraint.relaxable_lb:
# print("LB cost " + str(constraint.relax_cost_lb) + "/" + constraint.name)
variable = numVar
variables[(constraint, bound)] = variable
initial_values[variable] = constraint.get_lower_bound()
numVar += 1
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
# add the variable to the objective function
variable_bounds[variable] = [nInfinity,constraint.lower_bound]
objective[variable] = [-1*constraint.relax_cost_lb,constraint.lower_bound*constraint.relax_cost_lb]
else:
variable_bounds[variable] = [constraint.lower_bound,constraint.upper_bound-0.001]
# We collect all probabilistic durations
# for encoding the chance constraint
if constraint.probabilistic:
prob_durations.add(constraint)
else:
objective[variable] = [constraint.relax_cost_lb,
-1 * constraint.lower_bound * constraint.relax_cost_lb]
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
# Only for non-probabilistic constraint
if (constraint, 1) in variables :
ub_variable = variables[(constraint, 1)]
constraint = {}
constraint[ub_variable] = 1
constraint[variable] = -1
constraint['lb'] = 0.001
constraint['ub'] = pInfinity
constraints.append(constraint)
list_iA.append(len(constraints))
list_jA.append(ub_variable + 1)
list_A.append(1)
neA_count += 1
list_iA.append(len(constraints))
list_jA.append(variable + 1)
list_A.append(-1)
neA_count += 1
lp_ncycle_constraint[variable] = coefficient
else:
variable = constraint.lower_bound
rhs -= constraint.lower_bound*coefficient
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_ub or (not constraint.controllable and constraint.probabilistic):
# if constraint.relaxable_ub:
# print("UB cost " + str(constraint.relax_cost_ub) + "/" + constraint.name)
variable = numVar
variables[(constraint, bound)] = variable
initial_values[variable] = constraint.get_upper_bound()
numVar += 1
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable_bounds[variable] = [constraint.upper_bound, pInfinity]
objective[variable] = [constraint.relax_cost_ub,-1*constraint.upper_bound*constraint.relax_cost_ub]
else:
variable_bounds[variable] = [constraint.lower_bound+0.001, constraint.upper_bound]
# We collection all probabilistic durations
# for encoding the chance constraint
if constraint.probabilistic:
prob_durations.add(constraint)
else:
objective[variable] = [-1 * constraint.relax_cost_ub,
constraint.upper_bound * constraint.relax_cost_ub]
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
# Only for non-probabilistic constraint
if (constraint, 0) in variables:
lb_variable = variables[(constraint, 0)]
constraint = {}
constraint[variable] = 1
constraint[lb_variable] = -1
constraint['lb'] = 0.001
constraint['ub'] = pInfinity
constraints.append(constraint)
list_iA.append(len(constraints))
list_jA.append(variable + 1)
list_A.append(1)
neA_count += 1
list_iA.append(len(constraints))
list_jA.append(lb_variable + 1)
list_A.append(-1)
neA_count += 1
lp_ncycle_constraint[variable] = coefficient
else:
variable = constraint.upper_bound
rhs -= constraint.upper_bound * coefficient
assert variable is not None
else:
lp_ncycle_constraint[variable] = coefficient
# print("Adding variable " + str(coefficient) + "*"+str(variable))
# add the ncycle constraint to the problem
# print(str(lp_ncycle_constraint))
constraints.append(lp_ncycle_constraint)
# print("C" + str(len(constraints)) + ": ", end="")
empty_constraint = True
for key in lp_ncycle_constraint:
if lp_ncycle_constraint[key] != 0:
list_iA.append(len(constraints))
list_jA.append(key + 1)
list_A.append(lp_ncycle_constraint[key])
neA_count += 1
empty_constraint = False
# constraint,bound = variables[key]
# print(str(lp_ncycle_constraint[key]) + "*" + "(" +str(key)+ ") ", end="")
lp_ncycle_constraint['lb'] = rhs+0.0001
lp_ncycle_constraint['ub'] = pInfinity
if empty_constraint and rhs > 0:
# This constraint cannot be met
return None,0
# print(">= " + str(rhs))
# Create the chance constraint
if len(prob_durations) > 0:
cc_constraint = {}
constraints.append(cc_constraint)
# Check if the chance constraint is relaxable
if cc.relaxable_bound:
cc_constraint['lb'] = -1
cc_constraint['ub'] = 0
# we need to encode an additional variable
# to represent cc
cc_var = numVar
variables[("CC", 1)] = cc_var
numVar += 1
variable_bounds[cc_var] = [cc.risk_bound, pInfinity]
list_iA.append(len(constraints))
list_jA.append(cc_var + 1)
list_A.append(-1)
neA_count += 1
# And add cc to the objective function
objective[cc_var] = [cc.relax_cost, -1 * cc.risk_bound * cc.relax_cost]
else:
cc_constraint['lb'] = 0
cc_constraint['ub'] = cc.risk_bound
# print("C" + str(len(constraints)) + " (CC): ", end="")
# print(str(cc_constraint['lb']) + "<= ", end="")
for constraint in prob_durations:
if (constraint, 0) in variables:
lb_var = variables[(constraint, 0)]
else:
lb_var = numVar
variables[(constraint, 0)] = lb_var
numVar += 1
if (constraint, 1) in variables:
ub_var = variables[(constraint, 1)]
else:
ub_var = numVar
variables[(constraint, 1)] = ub_var
numVar += 1
# override their domain since we are free to
# allocate any values for them
variable_bounds[lb_var] = [0, constraint.mean]
variable_bounds[ub_var] = [constraint.mean, pInfinity]
# initial_values[lb_var] = constraint.mean - 4*constraint.std
# initial_values[ub_var] = constraint.mean + 4*constraint.std
initial_values[lb_var] = constraint.get_lower_bound()
initial_values[ub_var] = constraint.get_upper_bound()
list_iG.append(len(constraints))
list_jG.append(lb_var + 1)
neG_count += 1
list_iG.append(len(constraints))
list_jG.append(ub_var + 1)
neG_count += 1
prob_vars.append(lb_var)
prob_vars.append(ub_var)
# print("Risk((" + str(lb_var) + ")--"+str(constraint.mean)+"--"+str(constraint.std)+"--("+str(ub_var)+")) ", end="")
prob_means.append(constraint.mean)
prob_stds.append(constraint.std)
# print("<= " + str(cc_constraint['ub']))
# Add the objective as the last constraint
constraints.append(objective)
for key in objective:
list_iA.append(len(constraints))
list_jA.append(key + 1)
list_A.append(objective[key][0])
if objective[key][0] != 0:
neA_count += 1
objective['lb'] = nInfinity
objective['ub'] = pInfinity
# Start Constructing the problem!!
# Solve using SNOPT
snopt.check_memory_compatibility()
# minrw, miniw, mincw defines how much character, integer and real
# storeage is neede to solve the problem.
# ! assigned by SNOPT
minrw = np.zeros((1), dtype=np.int32)
miniw = np.zeros((1), dtype=np.int32)
mincw = np.zeros((1), dtype=np.int32)
# Workspace for character, integer and real arrays.
# The plus 1 is for the objective function
rw = np.zeros((2000*(len(variables)+len(constraints)),), dtype=np.float64)
iw = np.zeros((1000*(len(variables)+len(constraints)),), dtype=np.int32)
cw = np.zeros((8 * 500,), dtype=np.character)
# rw = np.zeros((10000,), dtype=np.float64)
# iw = np.zeros((10000,), dtype=np.int32)
# cw = np.zeros((8 * 500,), dtype=np.character)
# The Start variable for SnoptA
Cold = np.array([0], dtype=np.int32)
Basis = np.array([1], dtype=np.int32)
Warm = np.array([2], dtype=np.int32)
# Variable definitions
# Initial values
x = np.zeros((len(variables),), dtype=np.float64)
# Lower and upper bounds
xlow = np.zeros((len(variables),), dtype=np.float64)
xupp = np.zeros((len(variables),), dtype=np.float64)
for key in variable_bounds:
if key in initial_values:
x[key] = initial_values[key]
bounds = variable_bounds[key]
xlow[key] = bounds[0]
xupp[key] = bounds[1]
# Initial values for x
xstate = np.zeros((len(variables),), dtype=np.int32)
# Vector of dual variables for the bound constraints
# ! assigned by SNOPT
xmul = np.zeros((len(variables),), dtype=np.float64)
# Initialize values for constraints
F = np.zeros((len(constraints),), dtype=np.float64)
# Lower and upper bounds for constraints
Flow = np.zeros((len(constraints),), dtype=np.float64)
Fupp = np.zeros((len(constraints),), dtype=np.float64)
for idx in range(0,len(constraints)):
constraint = constraints[idx]
Flow[idx] = constraint['lb']
Fupp[idx] = constraint['ub']
# Initial states
Fstate = np.zeros((len(constraints),), dtype=np.int32)
# Estimate of the vector of Lagrange multipliers.
# Since we know nothing about it at the moment. Set it to zero.
Fmul = np.zeros((len(constraints),), dtype=np.float64)
# Constant added to the objective row for
# printing purposes
ObjAdd = np.zeros((1,), dtype=np.float64)
ObjAdd[0] = 0
# The last row F is the objective function
ObjRow = np.zeros((1,), dtype=np.int32)
ObjRow[0] = len(constraints) # NOTE: We must add one to mesh with fortran */
# Reports the result of the call to snOptA
INFO = np.zeros((1,), dtype=np.int32)
# Set number of variables
n = np.zeros((1,), dtype=np.int32)
n[0] = len(variables)
# Set the number of functions (constraints and objectives)
# in F
neF = np.zeros((1,), dtype=np.int32)
neF[0] = len(constraints)
# dimension of arrays iAfun, jAvar and A
# They contain the nonzero elements of the linear part A
lenA = np.zeros((1,), dtype=np.int32)
lenA[0] = len(list_iA)
neA = np.zeros((1,), dtype=np.int32)
neA[0] = neA_count
# iAfun = np.zeros((lenA[0],), dtype=np.int32)
# jAvar = np.zeros((lenA[0],), dtype=np.int32)
# A = np.zeros((lenA[0],), dtype=np.float64)
if lenA[0] > 0:
iAfun = np.array(list_iA, dtype=np.int32)
jAvar = np.array(list_jA, dtype=np.int32)
A = np.array(list_A, dtype=np.float64)
else:
lenA[0] = 1
iAfun = np.zeros((lenA[0],), dtype=np.int32)
jAvar = np.zeros((lenA[0],), dtype=np.int32)
A = np.zeros((lenA[0],), dtype=np.float64)
# dimension of arrays iGfun and jGvar
# They contain the nonzero elements of the nonlinear part of the derivativeG
lenG = np.zeros((1,), dtype=np.int32)
lenG[0] = len(list_iG)
neG = np.zeros((1,), dtype=np.int32)
neG[0] = neG_count
if lenG[0] > 0:
iGfun = np.array(list_iG, dtype=np.int32)
jGvar = np.array(list_jG, dtype=np.int32)
else:
lenG[0] = 1
iGfun = np.zeros((lenG[0],), dtype=np.int32)
jGvar = np.zeros((lenG[0],), dtype=np.int32)
# Generate the usrf function for SNOPT
# names for variables and constraints
# By default no names provided (both set to 1)
nxname = np.zeros((1,), dtype=np.int32)
nFname = np.zeros((1,), dtype=np.int32)
nxname[0] = 1
nFname[0] = 1
# Names for variables and problem functions
# not used when nxname and nFname are set to 1
xnames = np.zeros((1 * 8,), dtype=np.character)
Fnames = np.zeros((1 * 8,), dtype=np.character)
Prob = np.zeros((200 * 8,), dtype=np.character)
# final number of superbasic variables
# ! assigned by SNOPT
nS = np.zeros((1,), dtype=np.int32)
# Number and sum of the infeasibilities of constraints
nInf = np.zeros((1,), dtype=np.int32)
sInf = np.zeros((1,), dtype=np.float64)
# Define Print and Summary files
# Must be called before any other SNOPT routing
# Unit number for summary file
iSumm = np.zeros((1,), dtype=np.int32)
iSumm[0] = 0
# Unit number for the print file
iPrint = np.zeros((1,), dtype=np.int32)
iPrint[0] = 0
# Unit number for the specs file
iSpecs = np.zeros((1,), dtype=np.int32)
iSpecs[0] = 4
# name for output print and spec files
printname = np.zeros((200 * 8,), dtype=np.character)
specname = np.zeros((200 * 8,), dtype=np.character)
# open output files using snfilewrappers.[ch] */
specn = "sntoya.spc"
printn = "sntoya.out"
specname[:len(specn)] = list(specn)
printname[:len(printn)] = list(printn)
# Open the print file, fortran style */
snopt.snopenappend(iPrint, printname, INFO)
# ================================================================== */
# First, sninit_ MUST be called to initialize optional parameters */
# to their default values. */
# ================================================================== */
snopt.sninit(iPrint, iSumm, cw, iw, rw)
# Set configurations
strOpt = np.zeros((200 * 8,), dtype=np.character)
DerOpt = np.zeros((1,), dtype=np.int32)
# set derivative options
strOpt_s = "Derivative option"
DerOpt[0] = 1
strOpt[:len(strOpt_s)] = list(strOpt_s)
snopt.snseti(strOpt, DerOpt, iPrint, iSumm, INFO, cw, iw, rw)
# set iteration limit
Major = np.array([250], dtype=np.int32)
strOpt_s = "Major Iteration limit"
strOpt[:len(strOpt_s)] = list(strOpt_s)
snopt.snseti(strOpt, Major, iPrint, iSumm, INFO, cw, iw, rw)
# SnoptA will compute the Jacobian by finite-differences. */
# The user has the option of calling snJac to define the */
# coordinate arrays (iAfun,jAvar,A) and (iGfun, jGvar). */
# ------------------------------------------------------------------ */
# Go for it, using a Cold start. */
# ------------------------------------------------------------------ */
# print("Computing relaxation")
snopt.snopta(Cold, neF, n, nxname, nFname,
ObjAdd, ObjRow, Prob, usrf,
iAfun, jAvar, lenA, neA, A,
iGfun, jGvar, lenG, neG,
xlow, xupp, xnames, Flow, Fupp, Fnames,
x, xstate, xmul, F, Fstate, Fmul,
INFO, mincw, miniw, minrw,
nS, nInf, sInf, cw, iw, rw, cw, iw, rw)
snopt.snclose(iPrint)
snopt.snclose(iSpecs)
# print("Done computing relaxation")
# print("Solution " + str(x))
if INFO[0] == 1:
# An optimal solution has been found
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
allocations = []
cc_relaxations = []
# print("Found relaxation: INFO=" + str(INFO[0]))
for constraint, bound in variables.keys():
variable = variables[(constraint, bound)]
relaxed_bound = x[variable]
# print(str(variable) + "==" + str(x[variable]))
if constraint == "CC":
# print("Relaxing CC from " + str(cc.risk_bound) + " to " + str(relaxed_bound))
if abs(relaxed_bound-cc.risk_bound) > 0.001:
cc_relaxation = ChanceConstraintRelaxation(cc)
cc_relaxation.relaxed_bound = relaxed_bound
cc_relaxations.append(cc_relaxation)
continue
if not constraint.controllable and constraint.probabilistic:
# This is allocation
# Both lb and ub variables must have been included in the calculation
if bound == 0:
lb_variable = variables[(constraint, 0)]
ub_variable = variables[(constraint, 1)]
allocated_lb = x[lb_variable]
allocated_ub = x[ub_variable]
allocation = TemporalAllocation(constraint)
allocation.allocated_lb = allocated_lb
allocation.allocated_ub = allocated_ub
# allocation.pretty_print()
allocations.append(allocation)
else:
# This is relaxation
if bound == 0:
# check if this constraint bound is relaxed
if abs(relaxed_bound - constraint.lower_bound) > 0.001:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
elif bound == 1:
# same for upper bound
if abs(relaxed_bound - constraint.upper_bound) > 0.001:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
# print("")
if len(relaxations) > 0 or len(allocations) > 0:
return relaxations, allocations, cc_relaxations, 0
return None, None, None, 0
def generate_mincost_relaxations(candidate, negative_cycle,
feasibility_type):
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# print("Solving against " + str(len(all_cycles)) + " cycles")
# Solve using PuLP, TODO, incorporate interface to other solvers,
# especially for nonlinear objective functions
prob = LpProblem("MinCostConflictResolution", LpMinimize)
# Solve using PyOpt/Snopt,
# especially for nonlinear constraints/objective functions
# construct variables and constraints
lp_variables = {}
lp_objective = []
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1
# print("Cycles: " + str(len(all_cycles)))
for cycle in all_cycles:
# cycle.pretty_print()
lp_ncycle_constraint = []
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in lp_variables:
variable = lp_variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
# lower bound, the domain is less than the original LB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_lb:
# print("LB cost " + str(constraint.relax_cost_lb) + "/" + constraint.name)
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable = LpVariable(constraint.id + "-LB",
None,
constraint.lower_bound)
# print("New LB VAR: " + str(variable) + " [" + str(None) + "," + str(constraint.lower_bound) + "]")
# add the variable to the objective function
lp_variables[(constraint, bound)] = variable
lp_objective.append(
(constraint.lower_bound - variable) *
constraint.relax_cost_lb)
else:
variable = LpVariable(
constraint.id + "-LB",
constraint.lower_bound,
constraint.upper_bound - 0.001)
# print("New LB-UC VAR: " + str(variable) + " [" + str(None) + "," + str(constraint.lower_bound) + "]")
# add the variable to the objective function
lp_variables[(constraint, bound)] = variable
lp_objective.append(
(variable - constraint.lower_bound) *
constraint.relax_cost_lb)
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
if (constraint, 1) in lp_variables:
ub_variable = lp_variables[(constraint, 1)]
uncertain_duration_constraint = []
uncertain_duration_constraint.append(
(ub_variable - variable))
prob += sum(
uncertain_duration_constraint) >= 0.001
# print("Adding domain constraint for " + constraint.name)
else:
variable = constraint.lower_bound
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_ub:
# print("UB cost " + str(constraint.relax_cost_ub) + "/" + constraint.name)
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable = LpVariable(constraint.id + "-UB",
constraint.upper_bound,
None)
# print("New UB VAR: " + str(variable) + " [" + str(constraint.upper_bound) + "," + str(None) + "]")
# add the variable to the objective function
lp_variables[(constraint, bound)] = variable
lp_objective.append(
(variable - constraint.upper_bound) *
constraint.relax_cost_ub)
else:
variable = LpVariable(
constraint.id + "-UB",
constraint.lower_bound + 0.001,
constraint.upper_bound)
# print("New UB-UC VAR: " + str(variable) + " [" + str(constraint.upper_bound) + "," + str(None) + "]")
lp_variables[(constraint, bound)] = variable
lp_objective.append(
(constraint.upper_bound - variable) *
constraint.relax_cost_ub)
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
if (constraint, 0) in lp_variables:
lb_variable = lp_variables[(constraint, 0)]
uncertain_duration_constraint = []
uncertain_duration_constraint.append(
(variable - lb_variable))
prob += sum(
uncertain_duration_constraint) >= 0.001
# print("Adding domain constraint for " + constraint.name)
else:
variable = constraint.upper_bound
assert variable is not None
# print("Adding variable " + str(coefficient) + "*"+str(variable))
lp_ncycle_constraint.append(variable * coefficient)
# add the constraint to the problem
# print(str(lp_ncycle_constraint))
if sum(lp_ncycle_constraint) >= 0:
# print(str(sum(lp_ncycle_constraint)) + " >= 0")
# Over relax a bit to account for precision issues
prob += sum(lp_ncycle_constraint) >= 0.0
else:
status = 0
# this is not resolvable
# no need to proceed
if status > 0:
# Set the objective function
prob += sum(lp_objective)
# for c in prob.constraints:
# print("CON: ", prob.constraints[c])
# print("OBJ: ", prob.objective)
# Solve the problem
try:
import gurobipy
status = prob.solve(pulp.GUROBI(mip=False, msg=False))
except ImportError:
# pass # Gurobi doesn't exist, use default Pulp solver.
status = prob.solve()
# exit(0);
# print("Computing relaxation")
# if no solution was found, do nothing
if status > 0:
# A solution has been found
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
# print("Found relaxation")
for constraint, bound in lp_variables.keys():
variable = lp_variables[(constraint, bound)]
relaxed_bound = value(variable)
if bound == 0:
# check if this constraint bound is relaxed
if relaxed_bound != constraint.lower_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
assert relaxation.relaxed_lb <= constraint.upper_bound
elif bound == 1:
# same for upper bound
if relaxed_bound != constraint.upper_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
assert relaxation.relaxed_ub >= constraint.lower_bound
# print("")
if len(relaxations) > 0:
return relaxations, 0
return None, 0
def generate_cc_relaxations(candidate,
negative_cycle,
feasibility_type,
cc,
pInfinity=1e6,
nInfinity=-1e6):
global prob_means
global prob_stds
global prob_vars
global cc_var
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# print("CC Solving against " + str(len(all_cycles)) + " cycles")
numVar = 0
numConstraint = 0
# Solve using Snopt,
# especiallFy for nonlinear constraints/objective functions
# construct variables and constraints
variables = {}
variable_bounds = {}
initial_values = {}
constraints = []
objective = {}
list_iA = []
list_jA = []
list_A = []
neA_count = 0
list_iG = []
list_jG = []
neG_count = 0
prob_durations = set()
prob_means = []
prob_stds = []
prob_vars = []
cc_var = -1
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1
# print("Cycles: " + str(len(all_cycles)))
for cycle in all_cycles:
lp_ncycle_constraint = {}
rhs = 0
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in variables:
variable = variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
# lower bound, the domain is less than the original LB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_lb or (
not constraint.controllable
and constraint.probabilistic):
# if constraint.relaxable_lb:
# print("LB cost " + str(constraint.relax_cost_lb) + "/" + constraint.name)
variable = numVar
variables[(constraint, bound)] = variable
initial_values[
variable] = constraint.get_lower_bound()
numVar += 1
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
# add the variable to the objective function
variable_bounds[variable] = [
nInfinity, constraint.lower_bound
]
objective[variable] = [
-1 * constraint.relax_cost_lb,
constraint.lower_bound *
constraint.relax_cost_lb
]
else:
variable_bounds[variable] = [
constraint.lower_bound,
constraint.upper_bound - 0.001
]
# We collect all probabilistic durations
# for encoding the chance constraint
if constraint.probabilistic:
prob_durations.add(constraint)
else:
objective[variable] = [
constraint.relax_cost_lb,
-1 * constraint.lower_bound *
constraint.relax_cost_lb
]
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
# Only for non-probabilistic constraint
if (constraint, 1) in variables:
ub_variable = variables[(constraint,
1)]
constraint = {}
constraint[ub_variable] = 1
constraint[variable] = -1
constraint['lb'] = 0.001
constraint['ub'] = pInfinity
constraints.append(constraint)
list_iA.append(len(constraints))
list_jA.append(ub_variable + 1)
list_A.append(1)
neA_count += 1
list_iA.append(len(constraints))
list_jA.append(variable + 1)
list_A.append(-1)
neA_count += 1
lp_ncycle_constraint[variable] = coefficient
else:
variable = constraint.lower_bound
rhs -= constraint.lower_bound * coefficient
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_ub or (
not constraint.controllable
and constraint.probabilistic):
# if constraint.relaxable_ub:
# print("UB cost " + str(constraint.relax_cost_ub) + "/" + constraint.name)
variable = numVar
variables[(constraint, bound)] = variable
initial_values[
variable] = constraint.get_upper_bound()
numVar += 1
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable_bounds[variable] = [
constraint.upper_bound, pInfinity
]
objective[variable] = [
constraint.relax_cost_ub,
-1 * constraint.upper_bound *
constraint.relax_cost_ub
]
else:
variable_bounds[variable] = [
constraint.lower_bound + 0.001,
constraint.upper_bound
]
# We collection all probabilistic durations
# for encoding the chance constraint
if constraint.probabilistic:
prob_durations.add(constraint)
else:
objective[variable] = [
-1 * constraint.relax_cost_ub,
constraint.upper_bound *
constraint.relax_cost_ub
]
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
# Only for non-probabilistic constraint
if (constraint, 0) in variables:
lb_variable = variables[(constraint,
0)]
constraint = {}
constraint[variable] = 1
constraint[lb_variable] = -1
constraint['lb'] = 0.001
constraint['ub'] = pInfinity
constraints.append(constraint)
list_iA.append(len(constraints))
list_jA.append(variable + 1)
list_A.append(1)
neA_count += 1
list_iA.append(len(constraints))
list_jA.append(lb_variable + 1)
list_A.append(-1)
neA_count += 1
lp_ncycle_constraint[variable] = coefficient
else:
variable = constraint.upper_bound
rhs -= constraint.upper_bound * coefficient
assert variable is not None
else:
lp_ncycle_constraint[variable] = coefficient
# print("Adding variable " + str(coefficient) + "*"+str(variable))
# add the ncycle constraint to the problem
# print(str(lp_ncycle_constraint))
constraints.append(lp_ncycle_constraint)
# print("C" + str(len(constraints)) + ": ", end="")
empty_constraint = True
for key in lp_ncycle_constraint:
if lp_ncycle_constraint[key] != 0:
list_iA.append(len(constraints))
list_jA.append(key + 1)
list_A.append(lp_ncycle_constraint[key])
neA_count += 1
empty_constraint = False
# constraint,bound = variables[key]
# print(str(lp_ncycle_constraint[key]) + "*" + "(" +str(key)+ ") ", end="")
lp_ncycle_constraint['lb'] = rhs + 0.0001
lp_ncycle_constraint['ub'] = pInfinity
if empty_constraint and rhs > 0:
# This constraint cannot be met
return None, 0
# print(">= " + str(rhs))
# Create the chance constraint
if len(prob_durations) > 0:
cc_constraint = {}
constraints.append(cc_constraint)
# Check if the chance constraint is relaxable
if cc.relaxable_bound:
cc_constraint['lb'] = -1
cc_constraint['ub'] = 0
# we need to encode an additional variable
# to represent cc
cc_var = numVar
variables[("CC", 1)] = cc_var
numVar += 1
variable_bounds[cc_var] = [cc.risk_bound, pInfinity]
list_iA.append(len(constraints))
list_jA.append(cc_var + 1)
list_A.append(-1)
neA_count += 1
# And add cc to the objective function
objective[cc_var] = [
cc.relax_cost, -1 * cc.risk_bound * cc.relax_cost
]
else:
cc_constraint['lb'] = 0
cc_constraint['ub'] = cc.risk_bound
# print("C" + str(len(constraints)) + " (CC): ", end="")
# print(str(cc_constraint['lb']) + "<= ", end="")
for constraint in prob_durations:
if (constraint, 0) in variables:
lb_var = variables[(constraint, 0)]
else:
lb_var = numVar
variables[(constraint, 0)] = lb_var
numVar += 1
if (constraint, 1) in variables:
ub_var = variables[(constraint, 1)]
else:
ub_var = numVar
variables[(constraint, 1)] = ub_var
numVar += 1
# override their domain since we are free to
# allocate any values for them
variable_bounds[lb_var] = [0, constraint.mean]
variable_bounds[ub_var] = [constraint.mean, pInfinity]
# initial_values[lb_var] = constraint.mean - 4*constraint.std
# initial_values[ub_var] = constraint.mean + 4*constraint.std
initial_values[lb_var] = constraint.get_lower_bound()
initial_values[ub_var] = constraint.get_upper_bound()
list_iG.append(len(constraints))
list_jG.append(lb_var + 1)
neG_count += 1
list_iG.append(len(constraints))
list_jG.append(ub_var + 1)
neG_count += 1
prob_vars.append(lb_var)
prob_vars.append(ub_var)
# print("Risk((" + str(lb_var) + ")--"+str(constraint.mean)+"--"+str(constraint.std)+"--("+str(ub_var)+")) ", end="")
prob_means.append(constraint.mean)
prob_stds.append(constraint.std)
# print("<= " + str(cc_constraint['ub']))
# Add the objective as the last constraint
constraints.append(objective)
for key in objective:
list_iA.append(len(constraints))
list_jA.append(key + 1)
list_A.append(objective[key][0])
if objective[key][0] != 0:
neA_count += 1
objective['lb'] = nInfinity
objective['ub'] = pInfinity
# Start Constructing the problem!!
# Solve using SNOPT
snopt.check_memory_compatibility()
# minrw, miniw, mincw defines how much character, integer and real
# storeage is neede to solve the problem.
# ! assigned by SNOPT
minrw = np.zeros((1), dtype=np.int32)
miniw = np.zeros((1), dtype=np.int32)
mincw = np.zeros((1), dtype=np.int32)
# Workspace for character, integer and real arrays.
# The plus 1 is for the objective function
rw = np.zeros((2000 * (len(variables) + len(constraints)), ),
dtype=np.float64)
iw = np.zeros((1000 * (len(variables) + len(constraints)), ),
dtype=np.int32)
cw = np.zeros((8 * 500, ), dtype=np.character)
# rw = np.zeros((10000,), dtype=np.float64)
# iw = np.zeros((10000,), dtype=np.int32)
# cw = np.zeros((8 * 500,), dtype=np.character)
# The Start variable for SnoptA
Cold = np.array([0], dtype=np.int32)
Basis = np.array([1], dtype=np.int32)
Warm = np.array([2], dtype=np.int32)
# Variable definitions
# Initial values
x = np.zeros((len(variables), ), dtype=np.float64)
# Lower and upper bounds
xlow = np.zeros((len(variables), ), dtype=np.float64)
xupp = np.zeros((len(variables), ), dtype=np.float64)
for key in variable_bounds:
if key in initial_values:
x[key] = initial_values[key]
bounds = variable_bounds[key]
xlow[key] = bounds[0]
xupp[key] = bounds[1]
# Initial values for x
xstate = np.zeros((len(variables), ), dtype=np.int32)
# Vector of dual variables for the bound constraints
# ! assigned by SNOPT
xmul = np.zeros((len(variables), ), dtype=np.float64)
# Initialize values for constraints
F = np.zeros((len(constraints), ), dtype=np.float64)
# Lower and upper bounds for constraints
Flow = np.zeros((len(constraints), ), dtype=np.float64)
Fupp = np.zeros((len(constraints), ), dtype=np.float64)
for idx in range(0, len(constraints)):
constraint = constraints[idx]
Flow[idx] = constraint['lb']
Fupp[idx] = constraint['ub']
# Initial states
Fstate = np.zeros((len(constraints), ), dtype=np.int32)
# Estimate of the vector of Lagrange multipliers.
# Since we know nothing about it at the moment. Set it to zero.
Fmul = np.zeros((len(constraints), ), dtype=np.float64)
# Constant added to the objective row for
# printing purposes
ObjAdd = np.zeros((1, ), dtype=np.float64)
ObjAdd[0] = 0
# The last row F is the objective function
ObjRow = np.zeros((1, ), dtype=np.int32)
ObjRow[0] = len(
constraints) # NOTE: We must add one to mesh with fortran */
# Reports the result of the call to snOptA
INFO = np.zeros((1, ), dtype=np.int32)
# Set number of variables
n = np.zeros((1, ), dtype=np.int32)
n[0] = len(variables)
# Set the number of functions (constraints and objectives)
# in F
neF = np.zeros((1, ), dtype=np.int32)
neF[0] = len(constraints)
# dimension of arrays iAfun, jAvar and A
# They contain the nonzero elements of the linear part A
lenA = np.zeros((1, ), dtype=np.int32)
lenA[0] = len(list_iA)
neA = np.zeros((1, ), dtype=np.int32)
neA[0] = neA_count
# iAfun = np.zeros((lenA[0],), dtype=np.int32)
# jAvar = np.zeros((lenA[0],), dtype=np.int32)
# A = np.zeros((lenA[0],), dtype=np.float64)
if lenA[0] > 0:
iAfun = np.array(list_iA, dtype=np.int32)
jAvar = np.array(list_jA, dtype=np.int32)
A = np.array(list_A, dtype=np.float64)
else:
lenA[0] = 1
iAfun = np.zeros((lenA[0], ), dtype=np.int32)
jAvar = np.zeros((lenA[0], ), dtype=np.int32)
A = np.zeros((lenA[0], ), dtype=np.float64)
# dimension of arrays iGfun and jGvar
# They contain the nonzero elements of the nonlinear part of the derivativeG
lenG = np.zeros((1, ), dtype=np.int32)
lenG[0] = len(list_iG)
neG = np.zeros((1, ), dtype=np.int32)
neG[0] = neG_count
if lenG[0] > 0:
iGfun = np.array(list_iG, dtype=np.int32)
jGvar = np.array(list_jG, dtype=np.int32)
else:
lenG[0] = 1
iGfun = np.zeros((lenG[0], ), dtype=np.int32)
jGvar = np.zeros((lenG[0], ), dtype=np.int32)
# Generate the usrf function for SNOPT
# names for variables and constraints
# By default no names provided (both set to 1)
nxname = np.zeros((1, ), dtype=np.int32)
nFname = np.zeros((1, ), dtype=np.int32)
nxname[0] = 1
nFname[0] = 1
# Names for variables and problem functions
# not used when nxname and nFname are set to 1
xnames = np.zeros((1 * 8, ), dtype=np.character)
Fnames = np.zeros((1 * 8, ), dtype=np.character)
Prob = np.zeros((200 * 8, ), dtype=np.character)
# final number of superbasic variables
# ! assigned by SNOPT
nS = np.zeros((1, ), dtype=np.int32)
# Number and sum of the infeasibilities of constraints
nInf = np.zeros((1, ), dtype=np.int32)
sInf = np.zeros((1, ), dtype=np.float64)
# Define Print and Summary files
# Must be called before any other SNOPT routing
# Unit number for summary file
iSumm = np.zeros((1, ), dtype=np.int32)
iSumm[0] = 0
# Unit number for the print file
iPrint = np.zeros((1, ), dtype=np.int32)
iPrint[0] = 0
# Unit number for the specs file
iSpecs = np.zeros((1, ), dtype=np.int32)
iSpecs[0] = 4
# name for output print and spec files
printname = np.zeros((200 * 8, ), dtype=np.character)
specname = np.zeros((200 * 8, ), dtype=np.character)
# open output files using snfilewrappers.[ch] */
specn = "sntoya.spc"
printn = "sntoya.out"
specname[:len(specn)] = list(specn)
printname[:len(printn)] = list(printn)
# Open the print file, fortran style */
snopt.snopenappend(iPrint, printname, INFO)
# ================================================================== */
# First, sninit_ MUST be called to initialize optional parameters */
# to their default values. */
# ================================================================== */
snopt.sninit(iPrint, iSumm, cw, iw, rw)
# Set configurations
strOpt = np.zeros((200 * 8, ), dtype=np.character)
DerOpt = np.zeros((1, ), dtype=np.int32)
# set derivative options
strOpt_s = "Derivative option"
DerOpt[0] = 1
strOpt[:len(strOpt_s)] = list(strOpt_s)
snopt.snseti(strOpt, DerOpt, iPrint, iSumm, INFO, cw, iw, rw)
# set iteration limit
Major = np.array([250], dtype=np.int32)
strOpt_s = "Major Iteration limit"
strOpt[:len(strOpt_s)] = list(strOpt_s)
snopt.snseti(strOpt, Major, iPrint, iSumm, INFO, cw, iw, rw)
# SnoptA will compute the Jacobian by finite-differences. */
# The user has the option of calling snJac to define the */
# coordinate arrays (iAfun,jAvar,A) and (iGfun, jGvar). */
# ------------------------------------------------------------------ */
# Go for it, using a Cold start. */
# ------------------------------------------------------------------ */
# print("Computing relaxation")
snopt.snopta(Cold, neF, n, nxname, nFname, ObjAdd, ObjRow, Prob, usrf,
iAfun, jAvar, lenA, neA, A, iGfun, jGvar, lenG, neG, xlow,
xupp, xnames, Flow, Fupp, Fnames, x, xstate, xmul, F,
Fstate, Fmul, INFO, mincw, miniw, minrw, nS, nInf, sInf,
cw, iw, rw, cw, iw, rw)
snopt.snclose(iPrint)
snopt.snclose(iSpecs)
# print("Done computing relaxation")
# print("Solution " + str(x))
if INFO[0] == 1:
# An optimal solution has been found
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
allocations = []
cc_relaxations = []
# print("Found relaxation: INFO=" + str(INFO[0]))
for constraint, bound in variables.keys():
variable = variables[(constraint, bound)]
relaxed_bound = x[variable]
# print(str(variable) + "==" + str(x[variable]))
if constraint == "CC":
# print("Relaxing CC from " + str(cc.risk_bound) + " to " + str(relaxed_bound))
if abs(relaxed_bound - cc.risk_bound) > 0.001:
cc_relaxation = ChanceConstraintRelaxation(cc)
cc_relaxation.relaxed_bound = relaxed_bound
cc_relaxations.append(cc_relaxation)
continue
if not constraint.controllable and constraint.probabilistic:
# This is allocation
# Both lb and ub variables must have been included in the calculation
if bound == 0:
lb_variable = variables[(constraint, 0)]
ub_variable = variables[(constraint, 1)]
allocated_lb = x[lb_variable]
allocated_ub = x[ub_variable]
allocation = TemporalAllocation(constraint)
allocation.allocated_lb = allocated_lb
allocation.allocated_ub = allocated_ub
# allocation.pretty_print()
allocations.append(allocation)
else:
# This is relaxation
if bound == 0:
# check if this constraint bound is relaxed
if abs(relaxed_bound - constraint.lower_bound) > 0.001:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
elif bound == 1:
# same for upper bound
if abs(relaxed_bound - constraint.upper_bound) > 0.001:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
# print("")
if len(relaxations) > 0 or len(allocations) > 0:
return relaxations, allocations, cc_relaxations, 0
return None, None, None, 0
def generate_mincost_relaxations(candidate,negative_cycle,feasibility_type):
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# print("Solving against " + str(len(all_cycles)) + " cycles")
# Solve using PuLP, TODO, incorporate interface to other solvers,
# especially for nonlinear objective functions
prob = LpProblem("MinCostConflictResolution", LpMinimize)
# Solve using PyOpt/Snopt,
# especially for nonlinear constraints/objective functions
# construct variables and constraints
lp_variables = {}
lp_objective = []
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1;
# print("Cycles: " + str(len(all_cycles)))
for cycle in all_cycles:
# cycle.pretty_print()
lp_ncycle_constraint = []
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in lp_variables:
variable = lp_variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
# lower bound, the domain is less than the original LB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_lb:
# print("LB cost " + str(constraint.relax_cost_lb) + "/" + constraint.name)
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable = LpVariable(constraint.id + "-LB",None,constraint.lower_bound)
# print("New LB VAR: " + str(variable) + " [" + str(None) + "," + str(constraint.lower_bound) + "]")
# add the variable to the objective function
lp_variables[(constraint,bound)] = variable
lp_objective.append((constraint.lower_bound - variable) * constraint.relax_cost_lb)
else:
variable = LpVariable(constraint.id + "-LB",constraint.lower_bound, constraint.upper_bound-0.001)
# print("New LB-UC VAR: " + str(variable) + " [" + str(None) + "," + str(constraint.lower_bound) + "]")
# add the variable to the objective function
lp_variables[(constraint,bound)] = variable
lp_objective.append((variable - constraint.lower_bound) * constraint.relax_cost_lb)
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
if (constraint,1) in lp_variables:
ub_variable = lp_variables[(constraint,1)]
uncertain_duration_constraint = []
uncertain_duration_constraint.append((ub_variable - variable))
prob += sum(uncertain_duration_constraint) >= 0.001
# print("Adding domain constraint for " + constraint.name)
else:
variable = constraint.lower_bound
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if constraint.relaxable_ub:
# print("UB cost " + str(constraint.relax_cost_ub) + "/" + constraint.name)
if constraint.controllable or feasibility_type == FeasibilityType.CONSISTENCY:
variable = LpVariable(constraint.id + "-UB",constraint.upper_bound, None)
# print("New UB VAR: " + str(variable) + " [" + str(constraint.upper_bound) + "," + str(None) + "]")
# add the variable to the objective function
lp_variables[(constraint,bound)] = variable
lp_objective.append((variable - constraint.upper_bound) * constraint.relax_cost_ub)
else:
variable = LpVariable(constraint.id + "-UB",constraint.lower_bound+0.001, constraint.upper_bound)
# print("New UB-UC VAR: " + str(variable) + " [" + str(constraint.upper_bound) + "," + str(None) + "]")
lp_variables[(constraint,bound)] = variable
lp_objective.append((constraint.upper_bound - variable) * constraint.relax_cost_ub)
# Add an additional constraint to make sure that
# the lower bound is smaller than the upper bound
if (constraint,0) in lp_variables:
lb_variable = lp_variables[(constraint,0)]
uncertain_duration_constraint = []
uncertain_duration_constraint.append((variable - lb_variable))
prob += sum(uncertain_duration_constraint) >= 0.001
# print("Adding domain constraint for " + constraint.name)
else:
variable = constraint.upper_bound
assert variable is not None
# print("Adding variable " + str(coefficient) + "*"+str(variable))
lp_ncycle_constraint.append(variable*coefficient)
# add the constraint to the problem
# print(str(lp_ncycle_constraint))
if sum(lp_ncycle_constraint) >= 0:
# print(str(sum(lp_ncycle_constraint)) + " >= 0")
# Over relax a bit to account for precision issues
prob += sum(lp_ncycle_constraint) >= 0.0
else:
status = 0;
# this is not resolvable
# no need to proceed
if status > 0:
# Set the objective function
prob += sum(lp_objective)
# for c in prob.constraints:
# print("CON: ", prob.constraints[c])
# print("OBJ: ", prob.objective)
# Solve the problem
try:
import gurobipy
status = prob.solve(solvers.GUROBI(mip=False,msg=False))
except ImportError:
# pass # Gurobi doesn't exist, use default Pulp solver.
status = prob.solve()
# exit(0);
# print("Computing relaxation")
# if no solution was found, do nothing
if status > 0:
# A solution has been found
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
# print("Found relaxation")
for constraint, bound in lp_variables.keys():
variable = lp_variables[(constraint, bound)]
relaxed_bound = value(variable)
if bound == 0:
# check if this constraint bound is relaxed
if relaxed_bound != constraint.lower_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
assert relaxation.relaxed_lb <= constraint.upper_bound
elif bound == 1:
# same for upper bound
if relaxed_bound != constraint.upper_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
assert relaxation.relaxed_ub >= constraint.lower_bound
# print("")
if len(relaxations) > 0:
return relaxations,0
return None,0
def generate_maxflex_relaxations(candidate, negative_cycle):
if len(negative_cycle.constraints) == 0:
return None, 0
all_cycles = candidate.continuously_resolved_cycles.copy()
all_cycles.add(negative_cycle)
# Solve using PuLP, TODO, incorporate interface to other solvers,
# especially for nonlinear objective functions
prob = LpProblem("MaxFlexConflictResolution", LpMaximize)
# Solve using PyOpt/Snopt,
# especially for nonlinear constraints/objective functions
# construct variables and constraints
lp_variables = {}
lp_objective = []
max_flex_variable = LpVariable("Max Flex", 0, None)
lp_objective.append(max_flex_variable)
# status indicating the feasibility of the relaxation problem
# 1 is feasible
# 0 is infeasible
# Note that this variable is shared by PuLP for its result
status = 1
# We only consider the relaxation of the upper bound of
# uncontrollable duration.
# which means that we basically ignore the definition of
# 'relaxable bounds' in the network, and only consider the upper bounds
# of uncertain durations.
for cycle in all_cycles:
lp_constraint = []
for constraint, bound in cycle.constraints.keys():
# The constraint is a pair (temporal_constraint,0/1)
# where 0 or 1 represent if it is the lower or upper bound
# first we define the variables
# which only come from relaxable bounds of constraints
# in other words, if no constraint in a negative cycle is
# relaxable, the LP is infeasible
# and we can stop here
# TODO: add handler for uncontrollable duration
variable = None
if (constraint, bound) in lp_variables:
variable = lp_variables[(constraint, bound)]
coefficient = cycle.constraints[(constraint, bound)]
if variable is None:
if bound == 0:
variable = constraint.lower_bound
elif bound == 1:
# upper bound, the domain is larger than the original UB
# if the constraint is not relaxable, fix its domain
if not constraint.controllable:
variable = LpVariable(constraint.id + "-UB",
constraint.lower_bound,
constraint.upper_bound)
lp_variables[(constraint, bound)] = variable
# add the variable to the objective function
max_flex_constraint = []
max_flex_constraint.append(
(variable - constraint.lower_bound))
max_flex_constraint.append(-1.0 *
max_flex_variable)
prob += sum(max_flex_constraint) >= 0
else:
variable = constraint.upper_bound
assert variable is not None
lp_constraint.append(variable * coefficient)
# add the constraint to the problem
# print(str(lp_constraint))
if sum(lp_constraint) >= 0:
# print(str(sum(lp_constraint)) + " >= 0")
prob += sum(lp_constraint) >= 0
else:
status = 0
# this is not resolvable
# no need to proceed
if status > 0:
# Set the objective function
prob += sum(lp_objective)
# for c in prob.constraints:
# print("CON: ", prob.constraints[c])
# print("OBJ: ", prob.objective)
# Solve the problem
try:
import gurobipy
status = prob.solve(pulp.GUROBI(mip=False, msg=False))
except ImportError:
pass # Gurobi doesn't exist, use default Pulp solver.
status = prob.solve()
# exit(0);
# if no solution was found, do nothing
if status > 0:
# A solution has been bound
# extract the result and store them into a set of relaxation
# the outcome is a set of relaxations
relaxations = []
max_flex_value = value(max_flex_variable)
for constraint, bound in lp_variables.keys():
variable = lp_variables[(constraint, bound)]
relaxed_bound = value(variable)
if bound == 0:
# check if this constraint bound is relaxed
if relaxed_bound != constraint.lower_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_lb = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
elif bound == 1:
# same for upper bound
if relaxed_bound != constraint.upper_bound:
# yes! create a new relaxation for it
relaxation = TemporalRelaxation(constraint)
relaxation.relaxed_ub = relaxed_bound
# relaxation.pretty_print()
relaxations.append(relaxation)
if len(relaxations) > 0:
return relaxations, max_flex_value
return None, 0
