Skip to content

Latest commit

 

History

History
34 lines (25 loc) · 1.79 KB

README.md

File metadata and controls

34 lines (25 loc) · 1.79 KB
Raw

Job Shop Scheduling Problem with Reinforcement Learning

This is the implemention of JSSP with RL. The framework used for RL is actor critic and the dataset comes from Tianchi competition.

Problem formulation

see https://tianchi.aliyun.com/competition/entrance/531831/information for more details

Action design

The agent action takes from {0, 1} where the action array has the same size as the number of experts.

Reward design

The reward takes the following formula:

r = 1 - M/N

where M is the amount of left jobs and N is the total job amount.

State design

The state is the concat of job status and job distribution time

How to use

  • the data file contains the dataset for the JSSP
  • the RL_brain.py is the reinforcement learning algorithm for the problem
  • the trainer.py is the training function for the problem
  • the job_env.py is the environment for the JSSP
  • the utils.py contains useful functions that are used many times
  • the run.py is the main file of the project

One could run the following code to train the agent:

python run.py --lr=0.01 --gamma=0.9 --seed=2020 --num-steps=100 --max-episode-length=1e5

Reference on better JSSP solution

A paper named A Reinforcement Learning Environment For Job-Shop Scheduling is published in arXiv on 4/9, 2021. One could refer to this paper for action design, state design and most importantly, reward function design. The paper also provides a JSSP standard environment and a baseline solution for the environment, which might be of great help for those who are interested in JSSP. The paper could be found here and the corresponding code could be found here