Evolutionary job scheduling with optimized population by deep reinforcement learning

Abstract

The sorting operation of the production line in a heavy industrial scenario has the double complexity of the problem and the data. To improve production efficiency, the operation needs to be optimized. Aimed at this problem, this article designs a data representation method and an evolutionary job scheduling algorithm with an optimized population by deep reinforcement learning (DRL). Moreover, a real industrial dataset is contributed. The representation method represents the job data by referring to the bag-of-words model. The evolutionary algorithm uses DRL to initialize the genetic algorithm (GA)'s population and further evolves the population through the GA to obtain the final scheduling result. The experimental results indicate that the evolutionary algorithm has achieved the largest decrease in the average times for frame clearing on the real and simulated validation datasets, which are 12.54$\mathrm{\%}$ and 11.43$\mathrm{\%}$, respectively. It is of great significance for subsequent scheduling of the full-scenario digital twin.

Keywords:

• Genetic algorithm
• reinforcement learning
• deep learning
• job scheduling
• combinatorial optimization

Acknowledgments

The authors are grateful to the engineers of the Hunan Speedbot Robot Co. Ltd for their assistance in this study.

Disclosure statement

No potential conflict of interest was reported by the authors.

Additional information

Funding

This work was supported in part by the National Key Research and Development Program of China [grant number 2018AAA0102200]; in part by the National Natural Science Foundations of China (NSFC) [grant numbers 61572507, 61532003 and 61622212]; and in part by the National University of Defense Technology pre-research project [grant number ZK21-41].