A Multiagent Transformer-Based Algorithm for Multitask Dynamic Scheduling With Constrained Machines.

Area of Science:

• Operations Research
• Artificial Intelligence
• Manufacturing Systems

Background:

• Modern manufacturing demands simultaneous task handling in dynamic environments using shared, constrained machines.
• Existing multitask scheduling algorithms primarily focus on knowledge transfer, neglecting collaborative efforts for shared resources.

Purpose of the Study:

• To propose a novel multiagent transformer (MAT)-based algorithm for multitask dynamic scheduling with constrained machines.
• To address the limitations of current algorithms by incorporating collaborative decision-making for shared resources.

Main Methods:

• Formulated the multitask scheduling problem as a sequential multiagent decision-making process.
• Developed a joint policy network for adaptive heuristic selection, leveraging common and task-specific knowledge.
• Designed a comprehensive reward function to guide collaborative decision-making across tasks.

Main Results:

• The proposed MAT algorithm demonstrated superior performance against 14 state-of-the-art competitors across 270 instances.
• Ablation studies confirmed the effectiveness of reward mechanisms and the joint policy network in enhancing decision-making.
• The algorithm successfully enabled collaborative machine sharing for addressing multiple tasks.

Conclusions:

• The MAT-based algorithm effectively solves multitask dynamic scheduling problems with constrained machines through collaborative efforts.
• The joint policy network and comprehensive reward function are crucial for informed, collaborative decision-making.
• This research advances scheduling algorithms by integrating multiagent collaboration for resource-constrained environments.