class TestIsFixedCallCount(unittest.TestCase):
""" Tests for PR#1402 (669e7b2b) """
def setup(self, skip_trivial_constraints):
m = ConcreteModel()
m.x = Var()
m.y = Var()
m.c1 = Constraint(expr=m.x + m.y == 1)
m.c2 = Constraint(expr=m.x <= 1)
self.assertFalse(m.c2.has_lb())
self.assertTrue(m.c2.has_ub())
self._model = m
self._opt = SolverFactory("cplex_persistent")
self._opt.set_instance(
self._model, skip_trivial_constraints=skip_trivial_constraints)
def test_skip_trivial_and_call_count_for_fixed_con_is_one(self):
self.setup(skip_trivial_constraints=True)
self._model.x.fix(1)
self.assertTrue(self._opt._skip_trivial_constraints)
self.assertTrue(self._model.c2.body.is_fixed())
with unittest.mock.patch(
"pyomo.solvers.plugins.solvers.cplex_direct.is_fixed",
wraps=is_fixed) as mock_is_fixed:
self.assertEqual(mock_is_fixed.call_count, 0)
self._opt.add_constraint(self._model.c2)
self.assertEqual(mock_is_fixed.call_count, 1)
def test_skip_trivial_and_call_count_for_unfixed_con_is_two(self):
self.setup(skip_trivial_constraints=True)
self.assertTrue(self._opt._skip_trivial_constraints)
self.assertFalse(self._model.c2.body.is_fixed())
with unittest.mock.patch(
"pyomo.solvers.plugins.solvers.cplex_direct.is_fixed",
wraps=is_fixed) as mock_is_fixed:
self.assertEqual(mock_is_fixed.call_count, 0)
self._opt.add_constraint(self._model.c2)
self.assertEqual(mock_is_fixed.call_count, 2)
def test_skip_trivial_and_call_count_for_unfixed_equality_con_is_three(
self):
self.setup(skip_trivial_constraints=True)
self._model.c2 = Constraint(expr=self._model.x == 1)
self.assertTrue(self._opt._skip_trivial_constraints)
self.assertFalse(self._model.c2.body.is_fixed())
with unittest.mock.patch(
"pyomo.solvers.plugins.solvers.cplex_direct.is_fixed",
wraps=is_fixed) as mock_is_fixed:
self.assertEqual(mock_is_fixed.call_count, 0)
self._opt.add_constraint(self._model.c2)
self.assertEqual(mock_is_fixed.call_count, 3)
def test_dont_skip_trivial_and_call_count_for_fixed_con_is_one(self):
self.setup(skip_trivial_constraints=False)
self._model.x.fix(1)
self.assertFalse(self._opt._skip_trivial_constraints)
self.assertTrue(self._model.c2.body.is_fixed())
with unittest.mock.patch(
"pyomo.solvers.plugins.solvers.cplex_direct.is_fixed",
wraps=is_fixed) as mock_is_fixed:
self.assertEqual(mock_is_fixed.call_count, 0)
self._opt.add_constraint(self._model.c2)
self.assertEqual(mock_is_fixed.call_count, 1)
def test_dont_skip_trivial_and_call_count_for_unfixed_con_is_one(self):
self.setup(skip_trivial_constraints=False)
self.assertFalse(self._opt._skip_trivial_constraints)
self.assertFalse(self._model.c2.body.is_fixed())
with unittest.mock.patch(
"pyomo.solvers.plugins.solvers.cplex_direct.is_fixed",
wraps=is_fixed) as mock_is_fixed:
self.assertEqual(mock_is_fixed.call_count, 0)
self._opt.add_constraint(self._model.c2)
self.assertEqual(mock_is_fixed.call_count, 1)
class TeamOrienteeringIlp:
def __init__(self,
num_teams,
vertex_reward,
edge_cost,
max_edge_cost,
max_vertices,
type_coverage=None,
min_type_coverage=None,
min_avg_type_conservation=None,
lazy_subtour_elimination=False,
solver='gurobi_persistent'):
self.logger = logging.getLogger(self.__class__.__name__)
if num_teams < 1:
raise ValueError('at least one team needed')
self._model = None
self._result = None
self._solver = None
self._team_max_vertices_constraints = []
self._team_max_edge_cost_constraints = []
self._lazy_subtour_elimination = lazy_subtour_elimination
self._solver_type = solver
self._vertex_reward = vertex_reward
self._num_teams = num_teams
self._max_edge_cost = max_edge_cost
self._max_vertices = max_vertices
self._type_coverage = type_coverage
self._min_type_coverage = min_type_coverage
self._min_avg_type_conservation = min_avg_type_conservation
if isinstance(edge_cost, dict):
# in this case, the edge costs is a dictionary (u, v) -> cost
self.logger.debug('Using sparse mode to build model')
self._is_graph_sparse = True
else:
# in this case, we have a matrix (array of arrays)
self.logger.debug('Using dense mode to build model')
self._is_graph_sparse = False
self._edge_cost = edge_cost
def build_model(self):
self.logger.info('Building model...')
self._model = aml.ConcreteModel()
# model parameters
self.logger.debug('Adding graph objects...')
self._model.TeamCount = aml.Param(initialize=self._num_teams,
mutable=True)
self._model.MaxEdgeCost = aml.Param(initialize=self._max_edge_cost,
mutable=True)
self._model.MaxVertexCount = aml.Param(initialize=self._max_vertices,
mutable=True)
self._model.Teams = aml.RangeSet(0, self._num_teams - 1)
self._model.Nodes = aml.RangeSet(0, len(self._vertex_reward) - 1)
self._model.r = aml.Param(
self._model.Nodes,
initialize=lambda model, n: self._vertex_reward[n])
if self._is_graph_sparse:
nodes_in, nodes_out = defaultdict(list), defaultdict(list)
for u, v in self._edge_cost:
nodes_in[v].append(u)
nodes_out[u].append(v)
self._model.Edges = aml.Set(initialize=self._edge_cost.keys())
self._model.NodesIn = aml.Set(
self._model.Nodes,
initialize=lambda model, node: nodes_in[node])
self._model.NodesOut = aml.Set(
self._model.Nodes,
initialize=lambda model, node: nodes_out[node])
self._model.d = aml.Param(
self._model.Edges,
initialize=lambda model, u, v: self._edge_cost[(u, v)])
else:
self._model.Edges = aml.Set(initialize=self._model.Nodes *
self._model.Nodes,
filter=lambda model, u, v: u != v)
self._model.d = aml.Param(
self._model.Edges,
initialize=lambda model, u, v: self._edge_cost[u][v])
# indicator variables for nodes and arcs
self._model.x = aml.Var(self._model.Edges * self._model.Teams,
domain=aml.Binary,
initialize=0)
self._model.y = aml.Var(self._model.Nodes * self._model.Teams,
domain=aml.Binary,
initialize=0)
# objective of the model: maximize reward collected from visited nodes
self._model.Objective = aml.Objective(
rule=lambda model: sum(model.y[n, t] * model.r[n]
for n in model.Nodes for t in model.Teams),
sense=aml.maximize)
# every team must leave and come back
self.logger.debug('Adding leave and return constraints...')
self._model.TeamsLeave = aml.Constraint(rule=lambda model: sum(
model.x[(u, v), t] for u, v in model.Edges for t in model.Teams
if u == 0) == model.TeamCount)
self._model.TeamsReturn = aml.Constraint(rule=lambda model: sum(
model.x[(u, v), t] for u, v in model.Edges for t in model.Teams
if v == 0) == model.TeamCount)
# every vertex must be visited at most once
self.logger.debug('Adding visit count constraint...')
self._model.VertexVisit = aml.Constraint(
(n for n in self._model.Nodes if n != 0),
rule=lambda model, node: sum(model.y[node, t]
for t in model.Teams) <= 1)
# incoming connnections = outgoing connections = node selected
# i.e. no sources or sinks (implies path is connected)
# enforces consistency between x and y (i.e. node touched by arcs if and only if it is selected)
# and at most one path passes from the node
self.logger.debug(
'Adding consistency and connectedness constraints...')
if self._is_graph_sparse:
def in_rule(model, node, team):
return sum(model.x[(v, node), team]
for v in model.NodesIn[node]) == model.y[node, team]
def out_rule(model, node, team):
return sum(model.x[(node, v), team]
for v in model.NodesOut[node]) == model.y[node,
team]
else:
def in_rule(model, node, team):
return sum(model.x[(node, v), team] for v in model.Nodes
if v != node) == model.y[node, team]
def out_rule(model, node, team):
return sum(model.x[(v, node), team] for v in model.Nodes
if v != node) == model.y[node, team]
self._model.Incoming = aml.Constraint(
((n, t) for n in self._model.Nodes for t in self._model.Teams),
rule=in_rule)
self._model.Outgoing = aml.Constraint(
((n, t) for n in self._model.Nodes for t in self._model.Teams),
rule=out_rule)
# subtour elimination, if required
if not self._lazy_subtour_elimination:
self.logger.debug('Adding subtour elimination constraints...')
self._model.u = aml.Var(self._model.Nodes * self._model.Teams,
bounds=(1.0, len(self._model.Nodes) - 1))
self._model.SubTour = aml.Constraint(
((u, v, t) for u, v in self._model.Edges
for t in self._model.Teams if u != 0 and v != 0),
rule=lambda model, u, v, t:
(model.u[u, t] - model.u[v, t] + 1 <= (len(model.Nodes) - 1) *
(1 - model.x[(u, v), t])))
if self._type_coverage and (self._min_type_coverage
or self._min_avg_type_conservation):
assert (not self._min_type_coverage
or not self._min_avg_type_conservation
or len(self._min_avg_type_conservation) == len(
self._min_type_coverage))
self.logger.debug('Adding coverage information...')
# type_coverage is a binary tensor s.t. C_ijk = 1 iff vertex j covers option k of type i
# min_type_coverage is a vector s.t. c_i is the minimum options of type i that the vaccine must cover
self._model.Types = aml.RangeSet(0, len(self._type_coverage) - 1)
self._model.Options = aml.RangeSet(
0,
len(self._type_coverage[0][0]) - 1)
self._model.TypeCoverage = aml.Param(
self._model.Types * self._model.Nodes * self._model.Options,
initialize=lambda model, t, n, o: self._type_coverage[t][n][o])
# indicator variable 1 iff at least one team visits at least one vertex of option i of type j
self._model.OptionCovered = aml.Var(self._model.Types *
self._model.Options,
domain=aml.Binary,
initialize=0)
self._model.OptionCoveredConstraint = aml.Constraint(
self._model.Types * self._model.Options,
rule=lambda model, typ, option: sum(
model.y[n, t] * model.TypeCoverage[typ, n, option]
for n in model.Nodes
for t in model.Teams) >= model.OptionCovered[typ, option])
self.logger.debug(
'Enforcing minimum coverage and/or conservation with %d types and %d options per type',
len(self._type_coverage), len(self._type_coverage[0][0]))
if self._min_type_coverage:
self.logger.debug('Adding minimum coverage for each type...')
# sum of above indicator variables must be at least the minimum option coverage for each type
self._model.MinOptionCoverage = aml.Param(
self._model.Types,
initialize=lambda model, typ: self._min_type_coverage[typ])
self._model.MinOptionCoverageConstraint = aml.Constraint(
self._model.Types,
rule=lambda model, typ:
(model.MinOptionCoverage[typ],
sum(model.OptionCovered[typ, o]
for o in model.Options), None))
if self._min_avg_type_conservation:
# every vertex must cover a minimum number of different options
self.logger.debug('Adding average vertex conservation...')
self._model.MinOptionConservation = aml.Param(
self._model.Types,
initialize=lambda model, typ: self.
_min_avg_type_conservation[typ])
self._model.MinOptionConservationConstraint = aml.Constraint(
self._model.Types,
rule=lambda model, typ: sum(model.y[n, t] * (
sum(model.TypeCoverage[typ, n, o] for o in model.
Options) - model.MinOptionConservation[typ])
for t in model.Teams
for n in model.Nodes) >= 0)
else:
self.logger.info('No coverage enforced')
self._solver = SolverFactory(self._solver_type)
self._solver.set_instance(self._model)
self.update_max_vertices(self._max_vertices)
self.update_max_edge_cost(self._max_edge_cost)
self.logger.info('Model build!')
return self
def update_max_vertices(self, max_vertices):
def get_constraint(team):
if max_vertices > 0:
return pmo.constraint(
expr=sum(self._model.y[n, team] for n in self._model.Nodes
if n != 0) <= self._model.MaxVertexCount)
else:
return None
self._max_vertices = max_vertices
self._model.MaxVertexCount.set_value(max_vertices)
self._team_max_vertices_constraints = self._update_constraint_for_all_teams(
self._team_max_vertices_constraints, get_constraint,
'MaxVerticesForTeam%d')
if max_vertices < 0:
self.logger.info('No maximum vertex count enforced.')
else:
self.logger.info('Maximum vertex count for each tour is %d',
self._max_vertices)
def update_max_edge_cost(self, max_edge_cost):
def get_constraint(team):
if max_edge_cost > 0:
return pmo.constraint(expr=sum(
self._model.x[(u, v), team] * self._model.d[u, v]
for u, v in self._model.Edges) <= self._model.MaxEdgeCost)
else:
return None
self._max_edge_cost = max_edge_cost
self._model.MaxEdgeCost.set_value(max_edge_cost)
self._team_max_edge_cost_constraints = self._update_constraint_for_all_teams(
self._team_max_edge_cost_constraints, get_constraint,
'MaxEdgeCostForTeam%d')
if max_edge_cost > 0:
self.logger.info('Maximum edge cost for each tour is %f',
self._max_edge_cost)
else:
self.logger.info('No maximum edge cost enforced.')
def _update_constraint_for_all_teams(self, current_constraints,
constraint_fn, name_fmt):
for name in current_constraints:
constr = getattr(self._model, name)
try:
self._solver.remove_constraint(constr)
except KeyError:
# happens after model is built, but not solved. not sure why
pass
setattr(self._model, name, None)
del constr
new_constraints = []
for team in range(self._num_teams):
name = name_fmt % team
constr = constraint_fn(team)
if constr is None:
continue
setattr(self._model, name, constr)
new_constraints.append(name)
self._solver.add_constraint(constr)
return new_constraints
def solve(self, options=None):
# if logging is configured, gurobipy will print messages there *and* on stdout
# so we silence its logger and redirect all stdout to our own logger
logging.getLogger('gurobipy.gurobipy').disabled = True
class LoggingStdOut:
def __init__(self):
self.logger = logging.getLogger('stdout')
def write(self, message):
self.logger.debug(message.rstrip())
def flush(self, *args, **kwargs):
pass
sys.stdout = LoggingStdOut()
try:
return self._solve(options)
except Exception:
# restore stdout so that handlers can print normally
# https://docs.python.org/3/library/sys.html#sys.__stdout__
sys.stdout = sys.__stdout__
raise
finally:
sys.stdout = sys.__stdout__
def _solve(self, options=None):
''' solves the model optimally
'''
self.logger.info('Solving started')
if self._model is None:
raise RuntimeError('must call build_model before solve')
self._subtour_constraints = 0
while True:
res = self._solver.solve(options=options or {},
tee=1,
save_results=False,
report_timing=True)
if res.solver.termination_condition != TerminationCondition.optimal:
raise RuntimeError(
'Could not solve problem - %s . Please check your settings'
% res.Solution.status)
self._solver.load_vars()
team_tours_dict = self._extract_solution_from_model()
self.logger.debug('Solution contains the following tours: %s',
team_tours_dict)
valid_solution = True
all_tours_edges = []
for i, tour in enumerate(team_tours_dict):
this_tour_edges = self._extract_tours_from_arcs(tour)
self.logger.debug(
'Solution for team %d contains the following tour(s):', i)
for tt in this_tour_edges:
self.logger.debug(' %s', tt)
if len(this_tour_edges) > 1 or not any(
u == 0 for u, v in this_tour_edges[0]):
assert self._lazy_subtour_elimination, 'subtour elimination failed'
valid_solution = False
self._eliminate_subtours_dfj(this_tour_edges)
self.logger.debug(
'Subtour elimination constraints updated (%d inserted so far)'
% (self._subtour_constraints))
else:
all_tours_edges.append(this_tour_edges[0])
if valid_solution:
break
res.write(num=1)
self.logger.info('Solved successfully')
self._result = all_tours_edges
return self._result
def _eliminate_subtours_dfj(self, tours):
''' adds DFJ subtour elimination constraints
'''
for tour in tours:
tour_nodes = set(i for i, _ in tour)
for team in range(self._num_teams):
self._subtour_constraints += 1
name = 'Subtour_%d' % self._subtour_constraints
constraint = pmo.constraint(body=sum(
self._model.x[i, j, team] for i in tour_nodes
for j in tour_nodes
if i != j and (i, j) in self._model.Edges),
ub=len(tour_nodes) - 1)
setattr(self._model, name, constraint)
self._solver.add_constraint(getattr(self._model, name))
def _extract_solution_from_model(self):
''' returns a list of dictionaries i -> list of j containing the tours found by the model
'''
tours = []
for t in range(self._num_teams):
vertices = [
n for n in self._model.Nodes
if 0.98 <= self._model.y[n, t].value
]
self.logger.debug('Team %d selected nodes %s', t, vertices)
edges = {}
for i, j in self._model.Edges:
if 0.98 <= self._model.x[i, j, t].value <= 1.5:
edges[i] = j
tours.append(edges)
self.logger.debug('Team %d selected edges %s', t, edges)
assert set(vertices) == set(edges)
return tours
@staticmethod
def _extract_tours_from_arcs(arcs):
''' given a dictionary of arcs, returns a list of tours, where every tour is a list of arcs
'''
assert set(arcs.keys()) == set(
arcs.values()), 'arcs do not form a set of tours: %r' % arcs
not_assigned, assigned = set(arcs.keys()), set()
tours = []
while not_assigned:
tour, cursor = [], not_assigned.pop()
while not tour or cursor not in assigned:
tour.append((cursor, arcs[cursor]))
assigned.add(cursor)
not_assigned.discard(cursor)
cursor = arcs[cursor]
tours.append(tour)
return tours
def explore_edge_cost_vertex_reward_tradeoff(self, steps):
# introduce variables for vertex reward and edge cost
self._model.VertexReward = aml.Var()
self._model.AssignVertexReward = pmo.constraint(expr=sum(
self._model.y[n, t] * self._model.r[n] for n in self._model.Nodes
for t in self._model.Teams) == self._model.VertexReward)
self._model.EdgeCost = aml.Var()
self._model.AssignEdgeCost = aml.Constraint(expr=sum(
self._model.x[u, v, t] * self._model.d[u, v]
for (u, v) in self._model.Edges
for t in self._model.Teams) == self._model.EdgeCost)
self._solver.add_var(self._model.VertexReward)
self._solver.add_constraint(self._model.AssignVertexReward)
self._solver.add_var(self._model.EdgeCost)
self._solver.add_constraint(self._model.AssignEdgeCost)
# step 1: obtain maximum vertex reward
self.logger.info('Obtaining maximum vertex reward...')
del self._model.Objective
self._model.Objective = aml.Objective(expr=self._model.VertexReward,
sense=aml.maximize)
self._solver.set_objective(self._model.Objective)
self.solve()
max_reward = aml.value(self._model.VertexReward)
self.logger.info('Maximum reward is %f with cost %f', max_reward,
aml.value(self._model.EdgeCost))
# step 2: obtain minimum edge cost
self.logger.info('Obtaining minumum edge cost...')
del self._model.Objective
self._model.Objective = aml.Objective(expr=self._model.EdgeCost,
sense=aml.minimize)
self._solver.set_objective(self._model.Objective)
self.solve()
max_cost = aml.value(self._model.EdgeCost)
self.logger.info('Minimum cost is %f with reward %f', max_cost,
aml.value(self._model.VertexReward))
# step 3: obtain minumum edge cost, conditioned on maximum vertex reward
self.logger.info(
'Obtaining minimum cost conditioned on maximum reward...')
self._model.ForcedReward = aml.Param(initialize=max_reward)
self._model.MaxVertexReward = pmo.constraint(
expr=self._model.VertexReward == self._model.ForcedReward)
self._solver.add_constraint(self._model.MaxVertexReward)
self.solve()
max_cost_max_reward = aml.value(self._model.EdgeCost)
self.logger.info('Cost is %f with reward %f', max_cost_max_reward,
aml.value(self._model.VertexReward))
# step 4: iterate between these two values
self._solver.remove_constraint(self._model.MaxVertexReward)
del self._model.MaxVertexReward
del self._model.ForcedReward
self._model.EdgeCostSlackValue = aml.Param(initialize=0.0,
mutable=True)
self._model.EdgeCostSlack = aml.Var(within=aml.NonNegativeReals)
self._model.Epsilon = aml.Param(initialize=1e-4)
del self._model.Objective
self._model.Objective = aml.Objective(
rule=lambda model: model.VertexReward + model.Epsilon * model.
EdgeCostSlack,
sense=aml.maximize)
self._solver.add_var(self._model.EdgeCostSlack)
self._solver.set_objective(self._model.Objective)
for i in range(steps):
value = max_cost_max_reward + i * (
max_cost - max_cost_max_reward) / float(steps - 1)
self.logger.info('======')
self.logger.info('Iteration %d - Cost bound is %f', i + 1, value)
self._model.EdgeCostSlackValue.set_value(value)
self._model.EdgeCostConstr = pmo.constraint(
expr=self._model.EdgeCost +
self._model.EdgeCostSlack == self._model.EdgeCostSlackValue)
self._solver.add_constraint(self._model.EdgeCostConstr)
vaccine = self.solve()
reward, cost = aml.value(self._model.VertexReward), aml.value(
self._model.EdgeCost)
self.logger.info('Obtained reward %f with cost %f', reward, cost)
yield vaccine, reward, cost
self._solver.remove_constraint(self._model.EdgeCostConstr)
del self._model.EdgeCostConstr
