Graph Neural Networks and Deep Reinforcement Learning in Job Scheduling

• diploma thesis download: https://dspace.cuni.cz/handle/20.500.11956/190558
• diploma thesis assignment: https://www.cs.cas.cz/~martin/diplomka68.html

Repository structure

root/
├── benchmarks/        # Benchmarks used for testing the models
│   ├── jssp/         # Benchmarks for Job-Shop Scheduling (JSSP)
│   └── fjsp/         # Benchmarks for Flexible Job-Shop Scheduling (FJSP)
├── data/             # Data produced by testing the models on the benchmarks
│   ├── scripts/      # Scripts used for analysing the data (very messy, I do not advise reusing it)
│   └── experiment_{dynamic|static}_{jssp|fjsp}_<model>.csv      # Experimental data from the testing of <model> on dynamic|static version of jssp|fjsp 
├── models/           # Tested models
│   ├── baseline/     # Priority Dispatching Rules (PDRs) used as a baselines for the experimental comparison
│       ├── jssp/     # PDRs for JSSP used from the model IEEE-ICCE-RL-JSP 
│       └── jssp/     # PDRs for FJSP used from the model End-to-end-DRL-for-FJSP
│   └── <model>/      # Each folder is named after the model in the section ## Compared models below
├── thesis/           # Diploma thesis in LaTeX
└── literature.md     # Relevant literature.md

Run the experiments yourself

To run the experiments yourself, follow this steps:

• 1. Go to /models/<model>/repo
• 2. Run the script experiment_{dynamic|static}_{jssp|fjsp}_<model>.py
  • it's possible that if the file is not found, it is in some subdirectory of the repo directory
  • to setup dependencies, I used poetry or regular pip requirements.txt for some models
    • unfortunately, some models do not have dependencies, because it was really hard to set them up and then I forgot to save them after the experiments :(

Compared Models

Wheatley

• SHOWCASE
• trained 15x15 Wheatley model, showcase for small benchmark instances
• dynamic JSSP implemented by adding "virtual machines" at the start of the jobs
• STATIC JSSP DATA
• DYNAMIC JSSP DATA

fjsp-drl

• SHOWCASE
• model runs for all available benchmarks
• no dynamic version
• STATIC JSSP DATA
• STATIC FJSP DATA

L2D

• SHOWCASE
• model runs for all available benchmarks
• dynamic version runs by implementing custom algorithm
• STATIC JSSP DATA
• DYNAMIC JSSP DATA

IEEE-ICCE-RL-JSP

• SHOWCASE
• I trained 5 models running on all available benchmarks
• working dynamic model implemented by using already available code, just needed to pass the starting times to the underlying mechanism already in the code by the author
• STATIC JSSP DATA
• DYNAMIC JSSP DATA

End-to-end-DRL-for-FJSP

• SHOWCASE
• model runs for all available benchmarks
• no dynamic version
• STATIC JSSP DATA
• STATIC FJSP DATA

Baselines

• JSSP BASELINE DATA IEEE-ICCE-RL-JSP
• FJSP BASELiNE DATA (code taken from End-to-end-DRL-for-FJSP)

Benchmarks

JSSP

Description of JSSP instances is available at http://jobshop.jjvh.nl/explanation.php

• Adams, Balas and Zawacks

• Demirkol, Mehta, and Uzsoy

• Fisher and Thompson

• Lawrence

• Applegate and Cook

• Storer, Wu and Vaccari

• Taillard

• Yamada and Nakano

FJSP

Description of FJSP instances is available at DataSetExplanation.txt

• Behnke and Geiger

• Instances for the general FJSSP by Brandimarte

• MPM-JSSP-instances by Hurink et al.

• Multiprocessor-JSSP/FMS-instances by Dauzère-Pérès and Paulli

• FJSSP-instances by Chambers and Barnes

• Instances for the general FJSSP with total flexibility by Kacem et al.

• FJSSP-instances for Mathematical Programming by Fattahi et al.