Official codebase for our paper "NICE: Robust Scheduling through Reinforcement Learning-Guided Integer Programming" accepted at AAAI 2022
Project Page | AAAI 2022 Paper
Integer programs provide a powerful abstraction for representing a wide range of real-world scheduling problems. Despite their ability to model general scheduling problems, solving large-scale integer programs (IP) remains a computational challenge in practice. The incorporation of more complex objectives such as robustness to disruptions further exacerbates the computational challenge. We present NICE (Neural network IP Coefficient Extraction), a novel technique that combines reinforcement learning and integer programming to tackle the problem of robust scheduling. More specifically, NICE uses reinforcement learning to approximately represent complex objectives in an integer programming formulation. We use NICE to determine assignments of pilots to a flight crew schedule so as to reduce the impact of disruptions. We compare NICE with (1) a baseline integer programming formulation that produces a feasible crew schedule, and (2) a robust integer programming formulation that explicitly tries to minimize the impact of disruptions. Our experiments show that NICE produces schedules that are more robust to disruptions than the baseline formulation, with computation times that are lower than those of the robust integer program.
Experiemental code for NICE logits scheduling approach. Disruption experiments are run with:
python3 test_disrupt.py -l feasibility -i 100 -p model_wts/s0_d1_m0.0_b1.0.ckpt -d flights_delay -n 50 3 -r 2 -c 2 > constrained_experiment/delayed_50_buffer_exp.txt
-l specifies the type of baseline linear program you are comparing the logits approach to. "feasibility" in this case refers to the baseline IP.
-i specifies the number of iterations it should be run for.
-p is the saved neural network to use for the NICE and RL scheduling approach.
-d specifies the disruption type. "flights_delay" indicates that flights will be delayed
-n specifies the parameters for the disruption type. The first value ("50") is the percentage of flights that get delayed. The second ("3") is the maximum number of days any one flight will be delayed.
-r inidicates the "n" value for the weight extraction method
-c indicates the density of flights for the given schedule.
Note that we changed the name of the NICE scheduler while developing the code. In the codebase, it is referred to as "CARLO", "LOGITS", and "NICE". They are all synonymous.
For the model training (all of the experiments in the training_scripts folder), we used the Python package versions found in training_requirements.txt with Python 3.7.3. All of the other experiments used the package versions in testing_requirements.txt with Python 3.8.2.
training_scripts/run_exps.sh: script to train all of the NICE neural networks.time_experiments:
run_exp_{X}.sh runs the experiments where the buffer IP failed to create a single schedulebufferX.txt shows the results of those experiments (no output is produced because the experiments never finished running)selection_experiments:
generate_commands.py produces the run_exps.sh fileparse.py extracts all of the data out of the files produced by the experiment.parsed.txt contains the parsed datas{A}_d{B}_m{C}_b{D}_n{E}.txt contains all of the experimental data. The number of disruptions produced is listed at the end. Note that half of these experiments used the move-up reward function, which we use in the technical appendix of the paper. A indicates the seed value. B indicates the scheduling density of the experiment. C and D indicate the reward function used (C = 1.0 and D = 0.0 indicates the move-up reward function. C = 0.0 and D = 1.0 indicates the buffer reward function.) E indicates the n value used for weight extraction in the NICE approach. run_exps.sh runs all of the experiments where we ran each network/n value combo for 20 iterations to pick the best scheduling method. This is the experiment run in the Scheduling Parameter Selection section of the paper.robustness_experiments:
run_exps.sh runs the experiments that produced the data listed in the baseline scheduling scenario. It also has the experiments used for the move-up scheduling in the technical appendix of the code. delayed_{X}_{Y}_exp[_v{Z}].txt is the result of the experiment where X% of flights are delayed, and we use the NICE/RL scheduler using the network trained with the Y (buffer or moveup) reward function. If [_v{Z}] is included, that means, rather than the baseline IP, the NICE/RL schedulers were compared against the Z (Buffer or Moveup) IP scheduler. Note that, when Y or Z is moveup or Moveup, that experiment was included in the technical appendix of the paper.constrained_experiment:
run_exp.sh runs the experiments in the constrained scheduling environment (where the scheduling density is 2).delayed_{X}_buffer_exp.txt is the result of the experiment where X% of the flights are delayed. We did not compare against the buffer IP because, as the time experiments show, the buffer cannot produce a single schedule within 90 minutes of running in this environment.If you find this code useful, please consider citing our paper:
@inproceedings{DBLP:conf/aaai/KenworthyNCB22,
author = {Luke Kenworthy and
Siddharth Nayak and
Christopher Chin and
Hamsa Balakrishnan},
title = {{NICE:} Robust Scheduling through Reinforcement Learning-Guided Integer
Programming},
booktitle = {Thirty-Sixth {AAAI} Conference on Artificial Intelligence, {AAAI}
2022, Thirty-Fourth Conference on Innovative Applications of Artificial
Intelligence, {IAAI} 2022, The Twelveth Symposium on Educational Advances
in Artificial Intelligence, {EAAI} 2022 Virtual Event, February 22
- March 1, 2022},
pages = {9821--9829},
publisher = {{AAAI} Press},
year = {2022},
url = {https://ojs.aaai.org/index.php/AAAI/article/view/21218},
biburl = {https://dblp.org/rec/conf/aaai/KenworthyNCB22.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
} In case of any queries, please reach out to Luke or Siddharth.