Manufacturing Dispatching With Transfer Learning for Job-Due Cost Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The manufacturing industry faces challenges in finding an efficient dispatching rule for processing jobs due to the complexity of sequencing hundreds of jobs across multiple machines and processing stages, with existing reinforcement learning methods failing to minimize past-due and inventory costs effectively, especially in dynamic environments and across different factory configurations.
Innovation Solution
A deep reinforcement learning and transfer learning-based system is developed to formulate manufacturing dispatching processes, incorporating state representations of machine and job queue states, slack time information, and new reward functions to minimize lateness and tardiness, while enabling the transfer of optimal dispatching policies between factories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional reinforcement learning methods are used for job scheduling, then the system can learn dispatching policies, but it cannot effectively minimize past-due and inventory costs
Solution Approach 1:
The patent transforms the discrete job scheduling problem into a continuous control problem by introducing slack time as a continuous parameter. The deep deterministic policy gradient algorithm operates in this continuous space to directly minimize past-due and inventory costs, rather than learning discrete dispatching rules. This parameter transformation enables effective cost minimization while maintaining manageable system complexity through gradient-based optimization.
Solution Approach 2:
The patent replaces traditional reinforcement learning mechanisms (discrete policy learning, trial-and-error exploration) with a deep deterministic policy gradient approach that uses gradient descent optimization. This substitution allows the system to directly optimize for cost minimization by following gradient directions in the continuous slack time space, achieving better cost minimization effectiveness without the sample inefficiency and convergence issues of traditional RL methods.
2Reliability
If deep reinforcement learning is used to minimize job-due related costs, then cost minimization improves, but the system becomes more complex
Solution Approach 1:
The patent introduces slack time as a continuous parameter that transforms the discrete scheduling decision space into a continuous control space. This parameter change enables the use of deep deterministic policy gradient methods, which optimize cost minimization through gradient descent in the continuous space. The approach achieves superior cost minimization performance while keeping the model architecture relatively simple compared to discrete RL methods, as it directly learns continuous control policies without requiring complex discrete action spaces.
3Manufacturing precision
If optimal dispatching policies are customized for each factory, then local optimization improves, but the system lacks adaptability across different factories
Solution Approach 1:
The patent develops a universal deep deterministic policy gradient framework that can be deployed across different factories and manufacturing configurations. The model learns generalizable policies in the continuous slack time space that adapt to different factory settings, job characteristics, and cost structures. This universal approach maintains local optimization performance by adapting to specific factory conditions while providing cross-factory versatility through transfer learning capabilities, eliminating the need for complete policy retraining in new environments.
Solution Approach 2:
The patent creates a dynamic policy that continuously adapts to changing factory conditions, job characteristics, and cost parameters through the continuous slack time representation. The deep deterministic policy gradient algorithm enables real-time adaptation to different factory configurations by learning flexible continuous control policies rather than fixed discrete rules. This dynamic approach maintains optimal local performance while providing adaptability across different factories through the same underlying framework.
Data Source
AI summary
Example implementations described herein are directed to a system for manufacturing dispatching using reinforcement learning and transfer learning. The systems and methods described herein can be deployed in factories for manufacturing dispatching for reducing job-due related costs. In particular, example implementations described herein can be used to reduce massive data collection and reduce model training time, which can eventually improve dispatching efficiency and reduce factory cost.