半導體封裝廠之 Wire Bonding 工作站具有多台平行機台、工件動態到達、交期與批量差異，以及 recipe 切換所產生之換線換型時間等特性，使工件排序與機台指派形成高度複雜之動態排程問題。本研究以降低逾期為主要目標，建立單一站點、多台平行機台之事件驅動式排程環境，並採用近端策略最佳化（Proximal Policy Optimization，PPO）學習派工策略。Agent 於每一決策時點同時選擇工件與機台，並透過動作遮罩排除尚未到達、已完成或機台不可用之非法動作；reward 則納入準時完成批量、逾期、在製品及換線換型時間等因素，以兼顧交期與作業效率。 為驗證方法成效，本研究以基因演算法（Genetic Algorithm，GA）與模擬退火法（Simulated Annealing，SA）作為比較方法，並設定兩種規模，搭配寬鬆及緊迫交期，共形成四組實驗情境。實驗結果顯示，PPO 在四組情境中皆取得最低之批量加權總逾期時間，並於其中三組情境取得最低之總逾期時間。相較於 GA，PPO 之批量加權總逾期時間降低 10.57% 至 42.82%；相較於 SA，則降低 21.25% 至 50.33%。此外，PPO 整體亦具有較高之準交率，以及較低之逾期批量、換線換型時間、換線換型次數與最大完工時間。研究結果顯示，PPO 能在不同問題規模與交期壓力下有效學習工件－機台配對策略，並兼顧逾期控制、批量交付與工作站運作效率，可作為半導體封裝廠智慧派工與動態排程之參考。;The Wire Bonding work center in a semiconductor packaging plant is characterized by multiple parallel machines， dynamic job arrivals， differences in due dates and batch quantities， and setup times caused by recipe changes. These characteristics make job sequencing and machine assignment a highly complex dynamic scheduling problem. With tardiness reduction as the primary objective， this study develops an event-driven scheduling environment for a single work center with multiple parallel machines and applies Proximal Policy Optimization (PPO) to learn a dispatching policy. At each decision point， the agent simultaneously selects a job and a machine. An action mask is employed to eliminate infeasible actions involving jobs that have not yet arrived， jobs that have already been completed， or machines that are unavailable. The reward function incorporates the quantity completed on time， tardiness， work-in-process， and setup time to balance due-date performance and operational efficiency. To evaluate the effectiveness of the proposed method， Genetic Algorithm (GA) and Simulated Annealing (SA) are adopted as comparison methods. Two problem scales are considered under both loose and tight due-date settings， resulting in four experimental scenarios. The experimental results show that PPO achieves the lowest quantity-weighted total tardiness in all four scenarios and the lowest total tardiness in three of them. Compared with GA， PPO reduces quantity-weighted total tardiness by 10.57% to 42.82%; compared with SA， the reduction ranges from 21.25% to 50.33%. In addition， PPO generally achieves a higher on-time delivery rate， as well as lower tardy quantity， setup time， number of setups， and makespan. These findings indicate that PPO can effectively learn job–machine pairing policies under different problem scales and due-date pressures while simultaneously improving tardiness control， batch delivery performance， and work-center operational efficiency. Therefore， the proposed approach can serve as a reference for intelligent dispatching and dynamic scheduling in semiconductor packaging plants.
