Mid-term project scheduling in shipyards is a large-scale project scheduling problem where hundreds to thousands of activities form complex network structures. Resource leveling problem is an optimization problem that minimizes the variability of reso...
Mid-term project scheduling in shipyards is a large-scale project scheduling problem where hundreds to thousands of activities form complex network structures. Resource leveling problem is an optimization problem that minimizes the variability of resource usage during the project period to maximize resource utilization efficiency. Uneven resource usage causes resource overload during specific periods, leading to additional costs and project delays. Resource leveling problem is classified as NP-hard, and as project size increases, the solution space grows exponentially, making optimal solutions practically impossible for large-scale projects. Existing studies attempted to solve this problem using Meta heuristic algorithms and exact algorithms. However, Meta heuristic algorithms suffer from degraded solution quality and local optimal trapping as problem complexity increases, while exact algorithms face scalability issues with non-linear increases in computation time as problem size grows, making it difficult to obtain solutions within reasonable timeframes in large-scale projects like actual shipyards.
This paper proposes a two-stage approach combining an enhanced Kahn's Algorithm and attention-based deep reinforcement learning to effectively solve the resource leveling problem in large-scale projects. Project scheduling decision-making is divided into activity sequencing, which determines execution sequence of activities, and activity scheduling, which determines the timing of each activity. The first stage uses an enhanced Kahn's algorithm that considers priority based on total resource usage to determine the execution sequence satisfying precedence relationships between activities. While the original Kahn's algorithm only considers topological sorting, this research derives more efficient sequences by reflecting resource usage patterns. The second stage applies deep reinforcement learning with an Attention mechanism to determine the optimal start time for each activity. The attention mechanism integrates activity characteristics and resource status at the scheduling moment through state representation, simultaneously considering possible start time candidates, and the agent selects optimal actions based on learned policies that align with resource leveling objectives. This research proposes a methodology based on the REINFORCE algorithm while utilizing existing heuristics as a baseline to gradually improve policies, reducing instability in early learning stages and enabling robust performance even in large-scale problems.
Performance verification of the proposed model was conducted through experimental evaluation using the PSPLIB benchmark dataset and actual shipyard mid-term scheduling data. Experimental results demonstrate that the proposed methodology consistently outperforms existing methods including exact algorithms, meta heuristics, and heuristics in reducing resource usage variance for large-scale projects ranging from 299 to 5,800 activities. Notably, the algorithm implements robust scheduling without constraint violations even in highly complex real world problems considering both precedence and time range constraints, while maintaining practical computation times. Ablation experiments validate the contribution of the attention mechanism and enhanced Kahn's algorithm, confirming that the combination of these two components plays a critical role in performance improvement.
This paper presents the first attention-based deep reinforcement learning approach for the resource leveling problem. By establishing a learning environment that satisfies precedence relationships while minimizing solution space constraints, the global optimum reachability is improved. The algorithm demonstrates high generalization performance, overcoming computation time limitations in various large-scale projects and providing solutions within practical timeframes. Through this achievement, this research provides a practical solution applicable to production planning optimization in large-scale manufacturing industries including shipyards and demonstrates the feasibility of applying deep reinforcement learning in industrial practice.