Structured Convolutional Kernel Networks for Airline Crew Scheduling

Authors: Yassine Yaakoubi, Francois Soumis, Simon Lacoste-Julien

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we report the results of experiments using Struct-CKN. First, we sanity-check Struct-CKN on the standard OCR dataset in Section 6.1, showing that it’s comparable to the state of the art. Then, in Section 6.2, we use the proposed predictor on the flight-connection dataset to warmstart Commercial-GENCOL-DCA.
Researcher Affiliation Academia 1GERAD, Polytechnique Montréal, Canada 2Mila, Canada 3Department of Computer Science and Operations Research, Université de Montréal, Canada Canada CIFAR AI Chair.
Pseudocode Yes Algorithm 1 Training the Struct-CKN Model
Open Source Code Yes 1The code is available at the following link: https:// github.com/Yaakoubi/Struct-CKN
Open Datasets Yes The training set in the flight-connection dataset consists of six monthly crew pairing solutions (50,000 flights per month) and the test set is a benchmark that airlines use to decide on the commercial solver to use.
Dataset Splits No The paper mentions training and test sets for both the flight-connection dataset and the OCR dataset, and for OCR it mentions 'pre-specified folds; one fold is considered the test set, while the rest as the training set'. However, it does not explicitly state a separate validation set split or how it is formed (e.g., specific percentages or sample counts for validation).
Hardware Specification Yes We use Pytorch (Paszke et al., 2019) to declare said model and perform operations on a 40-core machine with 384 GB of memory, and use K80 (12 GB) GPUs.
Software Dependencies No We use Pytorch (Paszke et al., 2019) to declare said model... The CRF model is implemented using Py Struct (Müller & Behnke, 2014), while the SDCA optimizer is implemented using SDCA4CRF. Scalers are implemented using Scikit-learn (Pedregosa et al., 2011). This text lists software packages and provides citations, but it does not specify explicit version numbers for any of them (e.g., PyTorch 1.9, Scikit-learn 0.24).
Experiment Setup No The paper mentions some aspects of the experimental setup such as using 'small batches of image maps as inputs to the structured predictor (e.g., 128 image maps)' and using the 'Newton Raphson algorithm' for line search. However, it does not provide concrete hyperparameter values such as specific learning rates, a defined number of epochs for training, or other detailed optimizer settings.