This repository contains the code that corresponds to the CPAIOR 2024 submission Learning From Scenarios for Repairable Stochastic Scheduling by Kim van den Houten et al.
Besides the installation of the requirements.txt, it is needed to install the IBM CPLEX Optimization Studio (full edition, available via an academic licence) in order to be able to run all experiments in this repository. You may need to make the CPLEX optimizer available to the scripts by creating a symbolic link in the virtualenv, for example (adjust paths for your system):
ln -s /opt/ibm/ILOG/CPLEX_Studio2211/cpoptimizer/bin/x86-64_linux/cpoptimizer /path/to/Learning-From-Scenarios-for-Repairable-Stochastic-Scheduling/venv/bin/
Note that all scripts in this repository require the working directory to be /path/to/Learning-From-Scenarios-for-Repairable-Stochastic-Scheduling, and for this directory to be in PYTHONPATH.
The experiments shown in the papers use the following instances:
-
PSPlib j30 instances: j301_1 - j303_10 (30 instances).
-
PSPlib j90 instances: j901_1 - j903_10 (30 instances).
-
Small industry instances: 5_1_factory 1, 5_2_factory_1, 5_3_factory 1, 5_4_factory_1, 5_5_factory_1, 10_1_factory 1, 10_2_factory_1, 10_3_factory 1, 10_4_factory_1, 10_5_factory_1, 20_1_factory 1, 20_2_factory_1, 20_3_factory 1, 20_4_factory_1, 20_5_factory_1.
-
Large industry instances: 40_1_factory 1, 40_2_factory_1, 40_3_factory 1, 40_4_factory_1, 40_5_factory_1, 60_1_factory 1, 60_2_factory_1, 60_3_factory 1, 60_4_factory_1, 60_5_factory_1.
Decision-focused learning: Pytorch ADAM optimizer with default settings, batch size is 10, random shuffling is true. Deterministic model: IBM CP Optimizer with default settings. Stochastic model: IBM CP Optimer with default settings.
Settings per instance set:
PSPlib j30 instances: time budget = 30 min, the number of scenarios:
- deterministic, 100 scenarios
- stochastic, 50 scenarios
- decision-focused, 50 scenarios
PSPlib j90 instances: time budget = 60 min, the number of scenarios:
- deterministic, 100 scenarios
- stochastic, 50 scenarios
- decision-focused, 50 scenarios
Small industry instances: time budget: 5_i instances 600 sec, 10_i instances 1200 sec, 20_i instances 2400 sec, the number of scenarios:
- deterministic, 100 scenarios
- stochastic, 10 scenarios
- decision-focused, 25 scenarios
Large industry instances: time budget: 40_i instances 3600 sec, 60_i instances 3600 sec, the number of scenarios:
- deterministic, 100 scenarios
- stochastic, 10 scenarios
- decision-focused, 25 scenarios
The experiments files can be found in the experiment_scripts folder where there exists a separate folder for the industry experiments and the PSPlib experiments. Note that before running one should ensure that the settings are given as desired. Besides the scripts that run the different algorithms (run_deterministic.py, run_stochastic.py, run_SFGE.py), there are also scripts to visualize the results (plot_normalized.py to get an overview per instance, and plot_summarized.py to get the aggregated boxplots). These plots can also without running be found in the plots folder.
The results are provided in the results folder, in which a separate folder exists for each instance set, comprising two folders for the two different penalty settings (small and large). To analyze the significance of the obtained results, one could run experiment_scripts/paper_significance.py.