本研究期望建構一套兼具交期感知能力與多目標學習能力之雙目標深度強化學習 的彈性零工式排程問題架構，作為同時改善makespan與total tardiness 之有效方法。;This study is based on the Dual Attention Network (DAN) and extends the original single-objective reinforcement learning framework for the Flexible Job-Shop Scheduling Problem (FJSP) into a bi-objective reinforcement learning scheduling problem that simultaneously minimizes makespan and total tardiness. A disjunctive graph is adopted to describe the precedence relations among operations， machine competition relations， and tardiness-related information.
In terms of problem formulation， this study adopts a Multi-Objective Markov Decision Process (MOMDP)， in which a reward vector is used to simultaneously represent the two objectives of makespan and total tardiness. In addition， due date-related features， including due date， estimated tardiness， and critical ratio are incorporated into the state representation to enhance the model’s ability to perceive due date risk. In terms of model design， this study further adopts the concept of multi-level attention pooling to collect graph-structured information from each layer， thereby improving the representation of global features. Moreover， to alleviate the training instability caused by reward scale discrepancies and gradient imbalance in bi-objective learning， this study adopts reward scaling and gradient normalization to dynamically adjust the weights between the two objectives. In addition， the learning rate is adjusted during training to further improve training stability and model performance.
This study aims to construct a bi-objective deep reinforcement learning framework for the flexible job-shop scheduling problem that integrates both due date awareness and multi-objective learning capability， and to provide an effective approach for simultaneously improving makespan and total tardiness.
