Shipyard production environments are characterized by high uncertainty, including work delays, equipment failures, and emergency operations. In the execution planning phase, the feasibility of plans is directly affected by on-site uncertainties, makin...
Shipyard production environments are characterized by high uncertainty, including work delays, equipment failures, and emergency operations. In the execution planning phase, the feasibility of plans is directly affected by on-site uncertainties, making dynamic planning essential to reflect real-time production conditions. However, most existing research on execution plan optimization has been limited to static optimization based on deterministic data. While some studies have employed simulation, fuzzy theory, and rescheduling techniques to address uncertainty, they face limitations in supporting real-time decision-making in dynamic production environments due to issues such as poor integration between simulation and optimization processes, inadequate responses to unpredictable disruptions, and computational burdens from repetitive plan revisions.
To address these challenges, this study proposes a multi-agent reinforcement learning-based dynamic scheduling framework that supports real-time decision-making based on production monitoring data. To effectively account for the complexity and distributed nature of decision-making in shipyard production processes, three methodologies are proposed. First, for state encoding, an Edge-Feature-Augmented Heterogeneous Graph (EFA-HG)-based approach is proposed to effectively capture complex interdependencies and constraints among various production entities in shipyard operations. Second, for inter-agent communication, a Relation-Aware Message (RAM)-based communication scheme is introduced to enable agents to exchange high-dimensional relational information directly applicable to decision-making. Third, for multi-agent learning strategy, a Decompose-Integrate-Adapt DTDE (DIA-DTDE) learning strategy is proposed to maintain the scalability of conventional Decentralized Training with Decentralized Execution (DTDE) while mitigating environment non-stationarity to ensure learning stability.
The effectiveness of the proposed methodologies is validated on the most generalized machine scheduling problem types: unrelated parallel machines and flexible job shop. Specifically, case studies are conducted on four scheduling problems: quay wall allocation for post-stage outfitting operations, crane scheduling in steel stockyards, flexible job shop scheduling with crane transportation, and grand assembly process scheduling. The results demonstrate that the dynamic scheduling algorithm based on the proposed framework consistently outperforms priority dispatching rules and genetic programming methods, with performance gaps widening as problem scale and complexity increase. The contributions of the three proposed methodologies are verified as follows. The EFA-HG-based state encoding achieves an average performance improvement of 28.1% over conventional graph-based encoding, while RAM-based communication shows an average improvement of 11.1% compared to independent decision-making without information exchange. Furthermore, the DIA-DTDE learning strategy achieves an average performance improvement of 6.3% over the conventional DTDE strategy.
Based on these findings, this study provides the following theoretical contributions to the field of deep reinforcement learning-based dynamic scheduling. First, it presents an improved approach for effectively modeling the structural characteristics of complex production environments such as shipyards through EFA-HG-based state encoding. Second, it proposes a heterogeneous multi-agent learning framework that enables stable learning of cooperative policies for multiple decision-making elements through RAM-based communication and DIA-DTDE learning strategy. Additionally, this study offers the following practical contributions to shipyard production management. First, it enables immediate decision-making that reflects real-time site conditions when disruptive events occur in shipyard production environments, minimizing production disruptions and ensuring operational stability. Second, it supports cooperative decision-making among diverse departments performing different planning tasks, enabling productivity improvement from a system-wide perspective rather than individual departmental optimization.