ds4dm / branch-search-trees Goto Github PK

Dear authors,

I know that the instances you used are from some standard library, but I am wondering if there is a way to generate some synthetic instances by ecole (the platform also from your group)? Is it possible? I have used ecole for some time. For the candidate node embedding I guess we can use the node_observation.column_features (maybe plus some post-processing transform to get what you designed in paper), but I could not find a way to find a cutoff value for a MILP problem in ecole. Is the cutoff value necessary here?

Thanks anyway for this interesting work and I like it very much!

Hi,
I was trying to use the model provided in trained-models.

I had a few doubts:
for instance:

for map18 for seeds 0 1 2 3 4
I am getting the following number of nodes:
653, 747, 739, 519, 681

However, the paper reports shifted geometric mean of ~575.

Could the results vary too much ?

Hello! We have been reproducing the results of your insightful paper entilted " Parameterizing Branch-and-Bound Search Trees to
Learn Branching Policies " and encountered some issues. First of all, we appreciate that you publish your code and data in the repo.
And we utilize the data from the repo and the hyperparameters from the model file trying to reproduce retrain the model. The val/test
acc of NT and val acc of TG are similar to the results reported in the paper. But, the test acc of TG is 74.17%.
We are looking forward to discussing with you.
ps:

our model hyperparam:

•     Namespace(depth=5, dim_reduce_factor=2, dropout=0.0, eval_batchsize=500, hidden_size=64, infimum=8, lr=0.01, lr_decay_factor=0.1, lr_decay_schedule=[20, 30], momentum=0.9, norm='none', num_epochs=40, opt='adam', out_dir='trained-models/TreeGatePolicy_hidden64_depth5_lr0.01', policy_type='TreeGatePolicy', seed=0, test_h5_path='data/test.h5', top_k=[2, 3, 5, 10], train_batchsize=32, train_h5_path='data/train.h5', use_gpu=True, val_h5_path='data/val.h5', weight_decay=1e-05)
  

model hyperparam from repo:

•     Namespace(depth=5, dim_reduce_factor=2, dropout=0.0, eval_batchsize=500, hidden_size=64, infimum=8, lr=0.01, lr_decay_factor=0.1, lr_decay_schedule=[20, 30], momentum=0.9, norm='none', num_epochs=40, opt='adam', out_dir='trained-models/TreeGatePolicy_hidden64_depth5_lr0.01', policy_type='TreeGatePolicy', seed=0, test_h5_path='data/test.h5', top_k=[2, 3, 5, 10], train_batchsize=32, train_h5_path='data/train.h5', use_gpu=True, val_h5_path='data/val.h5', weight_decay=1e-05)
  

Best,
Zhi

Dear professor,

Sorry for bothering.

When I run the collect.py file. I got this error,

AttributeError: 'pyscipopt.scip.Model' object has no attribute 'getNNodesLeft'

This is my 'paths' setting

paths = {
'MYSYSTEM': {
'out_dir': 'new_dataset',
'instances_dir': 'branch-search-trees-dataset/train_instances/',
'cutoff_dict': 'branch-search-trees-dataset/cutoff_dict.pkl',
},
}

This is my run command
python collect.py -n air05.mps.gz \
-s 0 \
-k 10 \
--setting sandbox \
--system MYSYSTEM \
-v

The version of PySCIPOpt is 1.4.9, which is the same as in requirements.txt and I checked the functions inside the model through the dir(model) method, and did not find functions such as getNNodesLeft, getCandsState, getNodeState, and getMIPState.

Could you help me? How do you think about this error? Thank you very much, professor.

Sincerely.

Great job! Very inspiring.

However, I think there is a lack of comparison of solving time between different methods in the same instances.
Although the paper reports the number of nodes explored by learned and SCIP policies, the number of nodes and solving time are not exactly equivalent. Neural network policies maybe run faster than default SCIP policies with the same number of explored nodes.

A lot of people pay attention to ML4CO because of its potential to speed up solving MILP problems. It would be better and more intuitive that there is a solving time comparison, just like Exact Combinatorial Optimization with Graph Convolutional Neural Networks, Gasse et al.

Hi,
I ran the evaluate_scip.py with 'nu25-pr12.mps.gz' instance( and a few others such as mine-166-5.mps.gz), it shows status ( scip-getStatus() as infeasible).

It becomes feasible if I comment this "m.setObjlimit(cutoff_value)".

Can you please help with this.

Hello! I have a question. In the file brancher.py, maybe the code self.branchexec_count += 1 in line 56 should be placed after line 59? Otherwise, there is bias during the test phase because the same code in class SCIPCollectBrancher in line 161 is placed after line 154.

Hi,
I know your customized version of pyscipopt must be installed to run the code. I have a question. Your customized version of pyscipopt is based on version 1.4.9. However, Pyscipopt has been updated to version 4.0.0, involving many added, removed or fixed functionality. If I want to use some new added function which exists in higher version of pyscipopt such as model.redirectOutput(), what should I do?
Thank you in advance! Looking forward to your reply!

While the dataset we constructed is collected in an inherently sequential manner (it is a rollout of the SCIP default branching scheme), we randomly shuffle all the data and train as if all the datapoints are independent and identically distributed (iid). The motivating question is how accurate is this iid assumption?

There are two sides of the argument:
(1) Pro-iid camp: The hand-crafted features are themselves temporal and we are effectively in a Markovian setting.

(2) Anti-iid camp: While the features are themselves temporal, they are not temporal enough. There's still lots of temporal information we are failing to capture, which ultimately hinders generalization performance.

Currently our GatingNet is a feedforward multi-layer perceptron (MLP). The feature request is to modify the GatingNet to be a recurrent neural network. Instead of breaking sequentiality of our collected data, we preserve sequentiality and train in episodes or collected SCIP rollouts. This will require coding up new data-loaders and padding the sequences to allow for batch-wise training of episodes. Backpropagation through time (BPTT) may also need to be employed as for some of the more difficult instances, the rollouts can easily get into the thousand time-step range.

We measure if using a recurrent GatingNet improves generalization performance, both the (tangential but "standard" ML loss and accuracy metrics) and the MILP solver evaluation metrics.