Is there a way to get the optimal objective value of MILP solved by SCIP as information_function?

[thupchnsky]

Hi,

I know that there are many things in ecole.reward can be added into information function when defining the environment. For example, I can define a SCIP solver as default_env = ecole.environment.Configuring(observation_function=None, information_function={"nb_nodes": ecole.reward.NNodes(), "time": ecole.reward.SolvingTime()}, scip_params=scip_parameters). So I can get the number of nodes explored and the solving time. However, is there a reward function that returns the objective value of the original optimization problem? I try to search it in doc but could not find it.

[AntoinePrv]

Hello @thupchnsky

There is no such things in Ecole for now, but you can define it for yourself using PySCIPOpt:

class BestObjVal:
    def before_reset(self, model):  # Edit added missing function
        pass

    def extract(self, model, done):
        m = model.as_pyscipopt()
        return m.getSolObjVal(m.getBestSol())

And then:

default_env = ecole.environment.Configuring(
    observation_function=None,
    information_function={
        "nb_nodes": ecole.reward.NNodes(),
        "time": ecole.reward.SolvingTime(),
        "best_obj": BestObjVal(),
    },
    scip_params=scip_parameters
)

[thupchnsky]

Thanks for your reply! It seems that the custom class must have a reset method. Otherwise it gives me errors. However, when I add a dummy reset function to BestObjVal (I guess nothing needs to be reset here?..), it will cause fatal error to my jupyter kernel and force it to restart automatically. So I am wondering what should be the correct reset method here? Thanks!

[AntoinePrv]

It seems that the custom class must have a reset method. Otherwise it gives me errors.

You're right, my bad. An empty before_reset is needed.

When I add a dummy reset function to BestObjVal (I guess nothing needs to be reset here?..), it will cause fatal error

It seems the problem is coming PySCIPOpt.

In Configuring, in reset, nothing has been done so SCIP is in Stage SCIP_STAGE_PREOBLEM.
In this case, SCIPgetBestSol may return a NULL pointer because no solution exists, then in getSolObjVal pyscipopt does not check that NULL pointer and crashes.

Fortunaltely, getNSols does not do that, therefore, we add a check:

def extract(self, model, done):
    m = model.as_pyscipopt()
    if m.getNSols() <= 0:
        return ...  # decide what to do
    else:
        return m.getSolObjVal(m.getBestSol())

[thupchnsky]

It works! Thanks for the reply!

[AntoinePrv]

For reference, I opened an issue scipopt/PySCIPOpt#493