How do I assign a different arc cost function for each vehicle type? (Python/RoutingModel)

[Oliver Wilkie]

I'm trying to use the RoutingModel and I’m struggling to correctly set different cost function callbacks for each vehicle type.

This is a snippet from my function

# set costs
for courier in couriers:
 distance_function = distance_matrix.distance
 model.SetArcCostEvaluatorOfVehicle(distance_function, courier.vehicle_num)

When I run it, it produces a solution but the solution is incorrect. The end_node for the first vehicle isn’t a depot and this exception is thrown

  File "/Users/oliverw/caviar/projects/caviar-ml-service/env/lib/python2.7/site-packages/ortools/constraint_solver/pywrapcp.py", line 3220, in SolveWithParameters
    return _pywrapcp.RoutingModel_SolveWithParameters(self, parameters, assignment)
TypeError: in method 'RoutingModel_SolveWithParameters', argument 1 of type 'operations_research::RoutingModel *'

When I replace

# set costs
for courier in couriers:
 distance_function = distance_matrix.distance
 model.SetArcCostEvaluatorOfVehicle(distance_function, courier.vehicle_num)

with

model.SetArcCostEvaluatorOfAllVehicles(distance_function)

a valid solution is returned with no exceptions.

I’ve looked in the python examples but can’t find a use of SetArcCostEvaluatorOfVehicle in the python examples.
I saw in the Java examples that SetVehicleCost was another option but that didn’t help me either.

Do I need to set any specific parameters on the routing model in order for the SetArcCostEvaluatorOfVehicle to be used correctly?

Thanks in advance for any advice you guys have!

Oli

​

[Oliver Wilkie]

Using a C debugger I’ve found the point where this is going wrong.

The method in question is RoutingModel::GetFingerprintOfEvaluator. This method tries to build a fingerprint for each evaluator based on the distances they return for each node pairing. The error occurs when evaluator->Run() is called.

      .....
      if (from != to && !IsStart(to) && (!from_start || !IsEnd(to))) {
        row[row_size] = evaluator->Run(IndexToNode(from), IndexToNode(to));
        ++row_size;
      }

     ....

The stack trace is

        (lldb) thread backtrace

* thread #1: tid = 0x2e80a, 0x000000010000f0a3 Python`PyObject_Call + 28, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=EXC_I386_GPFLT)

  * frame #0: 0x000000010000f0a3 Python`PyObject_Call + 28

    frame #1: 0x000000010008ed8b Python`PyEval_CallObjectWithKeywords + 93

    frame #2: 0x000000010434becb _pywrapcp.so`PyCallback2NodeIndexNodeIndex(_object*, IntType<operations_research::_RoutingModel_NodeIndex_tag_, int>, IntType<operations_research::_RoutingModel_NodeIndex_tag_, int>) + 43

    frame #3: 0x000000010475839f libortools.dylib`operations_research::RoutingModel::GetFingerprintOfEvaluator(ResultCallback2<long long, IntType<operations_research::_RoutingModel_NodeIndex_tag_, int>, IntType<operations_research::_RoutingModel_NodeIndex_tag_, int> >*) const + 287

    frame #4: 0x0000000104758dff libortools.dylib`operations_research::RoutingModel::ComputeCostClasses() + 2207

    frame #5: 0x000000010475c063 libortools.dylib`operations_research::RoutingModel::CloseModel() + 147

    frame #6: 0x000000010475fdc1 libortools.dylib`operations_research::RoutingModel::Solve(operations_research::Assignment const*) + 33

When I use SetArcCostEvaluatorOfAllVehicles instead of SetArcCostEvaluatorOfVehicle it never calls RoutingModel::GetFingerprintOfEvaluator and so the problem doesn’t happen.

I have no idea if this strange behaviour is

a) me not using it correctly
b) a bug in the source code
c) a bug resulting from building incorrectly

Would really be grateful for any tips!

​

[Oliver Wilkie]

Figured it out! Its to do with scoping in the for loop.

Because distance_function is inside the forloop (and not used again in pythonland) it must get deallocated by python’s garbage collector.

I used this instead and it worked

vehicle_call_backs = [lambda i, j: matrix.Distance(i, j) for vehicle in range(vehicles)]
    for vehicle in range(vehicles):
      routing.SetArcCostEvaluatorOfVehicle(vehicle_call_backs[vehicle], vehicle)

​

[Michael Powell]

If you did it via callback methods like I did in C# then that's quite likely the case. It is necessary to maintain references to those unmanaged instances, or at minimum mark them not for GC, between callbacks.

[Benoît Boucher]

Thank you so much Oliver! I spent the whole day trying to figure it out!
Store the distance function outside of the forloop containing the evaluator that uses it... Genius!