maybe a bug in ecole.observation.Pseudocosts?

[ghost]

when I try to run the example as follows:

class ExploreThenStrongBranch:
    def __init__(self, expert_probability):
        self.expert_probability = expert_probability
        self.pseudocosts_function = ecole.observation.Pseudocosts()
        self.strong_branching_function = ecole.observation.StrongBranchingScores()

    def before_reset(self, model):
        self.pseudocosts_function.before_reset(model)
        self.strong_branching_function.before_reset(model)

    def extract(self, model, done):
        probabilities = [1-self.expert_probability, self.expert_probability]
        expert_chosen = bool(np.random.choice(np.arange(2), p=probabilities))
        if expert_chosen:
            return (self.strong_branching_function.extract(model,done), True)
        else:
            return (self.pseudocosts_function.extract(model,done), False)
        
        
scip_parameters = {'separating/maxrounds': 0, 'presolving/maxrestarts': 0,
                   'limits/time': 100, 'timing/clocktype': 2}
observation_function = { "scores": ExploreThenStrongBranch(expert_probability=0),
                         "node_observation": ecole.observation.NodeBipartite() }
env = ecole.environment.Branching(observation_function=observation_function,
                                  scip_params=scip_parameters, pseudo_candidates=True)


env.seed(0)
instance_path = 'data/instances/setcover/train_500r_1000c_0.05d/instance_0000.lp'
instance = ecole.scip.Model.from_file(instance_path)
observation, action_set, _, done, _ = env.reset(instance)

I find something hard to understand:
(1)
input observation['scores'][0].shape
output 994
not equal to num of variables 1000, why?
(2)
input observation['scores'][0][action_set]
output

array([       nan,        nan,        nan,        nan,        nan,
              nan,        nan,        nan,        nan,        nan,
              nan,        nan,        nan,        nan,        nan,
              nan, 0.18499214, 0.05232925, 0.11698202,        nan,
       0.18961997,        nan, 0.10188215, 0.17122907, 0.18852313,
              nan, 0.17757535, 0.19141571, 0.20885814, 0.20852026,
       0.11995715, 0.11698202, 0.15915762, 0.20743212, 0.06429137,
       0.1105206 , 0.17639468, 0.20891465, 0.05871806,        nan,
       0.20902973, 0.16134182, 0.04683438, 0.10060939, 0.04683438,
       0.03530609, 0.19141571, 0.17757535, 0.14443708, 0.0414829 ,
       0.09042382, 0.18490181, 0.13457705, 0.18852313, 0.10276868,
       0.13367503, 0.1713421 , 0.21513443, 0.09491988, 0.15915762,
       0.15915762, 0.27444468,        nan, 0.18961997, 0.13457705,
       0.19141571, 0.03853465, 0.02273904, 0.13457705, 0.09491988,
              nan, 0.18499214, 0.11995715, 0.13547365, 0.20891465,
       0.14707062, 0.18756717, 0.14443708, 0.00268779, 0.14707062,
       0.18852313, 0.22890247, 0.14707062, 0.73026008, 0.14707062,
       0.10276868, 0.14443708, 0.14707062, 0.11698202, 0.04683438,
       0.17519231, 0.13457705, 0.09491988, 0.18756717])

so when I run observation['scores'][0][action_set].argmax() it returns 0
the variable it chose to branch is wrong?

(3) when I set expert_probability=1, the StrongBranching does not have nan in observation['scores'][0][action_set]

so I'm confused if there's a bug in ecole.observation.Pseudocosts?

[ghost]

here's my instance_0000.lp used in code (rename the file by removing the suffix .csv )
instance_0000.lp.csv

[AntoinePrv]

Hello @CXCellTrack,

1. I can reproduce that observation['scores'][0].shape == 994. Furthermore, we also have
   instance.as_pyscipopt().getNVars() == 1000 and env.model.as_pyscipopt().getNVars() == 994.
   Somwhere in the preprocessing SCIP must realize that some variables can be eliminated (or even are unused).
   Ecole get the data from SCIP, but it is true that it can sometimes be surprising.

2. Yes, we fill in NaN when the observation can not be computed for a variable. Here you have ecole.environment.Branching(..., pseudo_candidates=True), so I assume that some Pseudocosts can not be computed for pseudo candidates.

3. That must be because Pseudocosts cannot be computed for pseudo candidates but StrongBranchingScores can.

Hope this helps.

[ghost]

thanks for your reply,
when I use ecole.environment.Branching(..., pseudo_candidates=False) the NaN disappeared, though I'm not very clear about what does pseudo_candidates mean, it works :)

[AntoinePrv]

pseudo_candiadates=True means the Branching environment will return a larger set of branching variables for the action set, extracted with SCIPgetPseudoBranchCands.
If I'm correct these are all non fixed variables, even if their value in the solution of the relaxation is integer.

With pseudo_candidates=False, the Branching environment will compute the action set with SCIPgetBranchCands that only include variables whose solution value to the relaxation is fractional.

[ghost]

thanks very much, I think I've got it!