本研究探討在柔性作業車間排程問題（Flexible Job-shop Scheduling Problem， FJSP）下，應用深度強化學習（Deep Reinforcement Learning， DRL）優化 total tardiness 的方法。不同於過去多數研究聚焦於 makespan，total tardiness 的目標函數具有有界於零、且在數值特性上與 makespan 不同等特點，使得為 makespan 設計的狀態特徵與訓練配置未必能直接適用。 本研究以 Wang 等人（2024）提出的雙重注意力網路（Dual Attention Network， DAN）為基礎，將其目標由 makespan 改為 total tardiness，並系統地檢驗四個層面的設計。在特徵層面，於原框架的手工特徵之外引入 Normalized Operation Due Date、Slack per Operation 與 Estimated Tardiness 等交期特徵，以及新設計的前瞻特徵 Projected Slack Ratio，並檢驗這些特徵對 total tardiness 的效果;同時將隨機閘門(stochastic gates)嵌入式特徵選擇方法改編至強化學習目標，檢驗完整三十個手工特徵中可移除的冗餘程度。在架構層面，檢驗殘差連接對此淺層注意力網路的作用及其機制。在測試策略層面，檢驗取樣次數作為測試階段計算預算對 total tardiness 的影響。在訓練層面，以標準的訓練診斷量檢驗訓練的穩定性，並據以檢驗是否需要額外的學習率排程或梯度裁剪。實驗在兩種問題規模（10×5 與 15×10）上進行，並將所提出的排程器與優先派工法則( priority dispatching rules )及數學規劃時限上界進行比較。;This study investigates the use of Deep Reinforcement Learning (DRL) to optimize total tardiness in the Flexible Job-shop Scheduling Problem (FJSP). Unlike most prior work that minimizes makespan， the total-tardiness objective is bounded below at zero and differs in numerical character from makespan， so the state features and training configuration designed for makespan do not necessarily transfer directly. Building on the Dual Attention Network (DAN) of Wang et al. (2024)， this study changes its objective from makespan to total tardiness and systematically examines four aspects of the design. At the feature level， it adds due-date features — normalized operation due date， slack per operation， and estimated tardiness — together with a newly designed forward-looking feature， the projected slack ratio， to the original handcrafted features and examines their effect on total tardiness; it also examines， through an embedded stochastic-gate selection method adapted to the reinforcement-learning objective， the extent to which the full set of thirty handcrafted features can be reduced. At the architectural level， it examines the role of residual connections in this shallow attention network and the mechanism behind it. At the test-strategy level， it examines the effect of the number of sampling iterations， viewed as a test-time compute budget， on total tardiness. At the training level， it uses standard training diagnostics to examine training stability and， based on these， whether additional learning-rate scheduling or gradient clipping is needed. Experiments are conducted on two problem sizes (10×5 and 15×10)， and the proposed scheduler is compared against priority dispatching rules and a time-limited mathematical-programming upper bound.
