Hybrid Deep Neural Network Scheduler for Job-Shop Problem Based on Convolution Two-Dimensional Transformation.

Area of Science:

• Operations Research
• Artificial Intelligence
• Machine Learning

Background:

• Job-shop scheduling problems (JSSPs) are complex combinatorial optimization challenges.
• Existing methods often struggle with scalability and efficiency for large-scale JSSPs.
• Deep learning offers potential for enhanced scheduling solutions.

Purpose of the Study:

• To propose a novel hybrid deep neural network scheduler (HDNNS) for solving JSSPs.
• To improve the efficiency and performance of job-shop scheduling.
• To develop a scheduler with strong generalization capabilities.

Main Methods:

• A hybrid deep neural network scheduler (HDNNS) framework is introduced.
• Job-shop scheduling problems are decomposed into classification-based subproblems.
• Convolutional two-dimensional transformation (CTDT) is employed to regularize scheduling data for deep learning.

Main Results:

• HDNNS achieved a 9% improvement in the MAKESPAN index compared to HNN and a 4% improvement over ANN on the ZLP dataset.
• The HDNNS method exhibited significantly shorter training times than the DEEPRM method for identical neural network structures.
• Experimental results demonstrated excellent generalization performance, enabling the scheduler to handle large-scale JSSPs with limited training data.

Conclusions:

• HDNNS is an effective and efficient approach for solving job-shop scheduling problems.
• The proposed method offers superior performance and faster training compared to existing techniques.
• HDNNS shows promising generalization capabilities for real-world, large-scale scheduling applications.