Flexible job-shop co-scheduling of machines and automated guided vehicles (AGVs) faces a quality–cost dilemma: global metaheuristic optimization produces high-quality schedules but is computationally expensive, whereas rule-based dispatching is ...

http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
https://www.riss.kr/link?id=T17548778
서울 : 동국대학교 일반대학원, 2026
학위논문(석사) -- 동국대학교 일반대학원 , 컴퓨터·AI학과 컴퓨터공학전공 , 2026. 8
2026
한국어
기계-AGV 동시 스케줄링 ; 유연잡샹 스케줄링 ; 선택적 시뮬레이션 기반 최적화 ; 이종 그래프 어텐션 네트워크 ; 강화학습 ; 메타 휴리스틱 ; NSGA-II ; 디지털 트윈
서울
65 p. ; 26 cm
지도교수: 석문기
I804:11020-000000094415
0
상세조회0
다운로드다국어 초록 (Multilingual Abstract)
Flexible job-shop co-scheduling of machines and automated guided vehicles (AGVs) faces a quality–cost dilemma: global metaheuristic optimization produces high-quality schedules but is computationally expensive, whereas rule-based dispatching is ...
Flexible job-shop co-scheduling of machines and automated guided vehicles (AGVs) faces a quality–cost dilemma: global metaheuristic optimization produces high-quality schedules but is computationally expensive, whereas rule-based dispatching is cheap but myopic. Under a limited optimization budget for each scheduling decision, committing to a fixed optimization level overspends on easy states and underspends on bottlenecks. This thesis reframes the optimization scope—how much budget to spend and which operations to optimize—as a learned, budget-constrained decision variable.
The proposed framework pairs a heterogeneous graph attention (HGAT) encoder that summarizes the shop-floor state with a two-head reinforcement-learning controller that sets an optimization budget and operation-type importance weights, thereby selecting a core operation set. A bounded NSGA-II searches only over this core set and returns a typed decision artifact—urgency shifts and machine/AGV preferences—that a rule-based decoder applies through a unified priority interface, which provably reduces to pure dispatching when the core set is empty. The controller is trained in two tiers: behavioral-cloning distillation from an NSGA-II teacher, followed by proximal policy optimization (PPO) fine-tuning against an in-loop full-scope reference.
Across three benchmark families (Kumar, Deroussi, and MK10, spanning 4 to 15 machines), under matched budgets the selective controllers surpass full-scope optimization when the budget is tight, and the margin widens as shop size grows—up to about 20% on MK10, with roughly 58 times fewer evaluations to reach a target solution. PPO closely tracks an ε-minimal core-set oracle, whereas behavioral cloning over-provisions as the shop scales, so PPO attains comparable quality preservation with a more parsimonious scope. The framework also preserves a Pareto front for operator-preference tuning that scalar optimizers discard. These results indicate that treating optimization scope as a learned, budget-constrained decision is an effective way to navigate the quality–cost trade-off in flexible job-shop co-scheduling.
목차 (Table of Contents)