Comparing Encoder Architectures for the Discrete Berth Allocation Problem

Abstract

This study looks at whether we can replace the Graph Neural Network (GNN)
encoder in a reinforcement learning framework with a direct attention mechanism to solve the Dynamic Discrete Berth Allocation Problem (DDBAP): the challenge of assigning ships to specific docking spots over time. We tested two alternative designs, a Pure Attention Transformer and a topology-aware Edge-Transformer, against a baseline Graph Attention Network (GAT) using 27 simulated shipping scenarios. On average, the baseline GNN performed relatively better overall, scoring an average cost of 646.58 compared to 650.72 for the Edge-Transformer and 658.63 for the Pure Transformer. However, the best choice depends entirely on the size of the problem. The GNN works best in low-traffic situations with plenty of available berths because it ignores irrelevant background information. In contrast, when the port gets highly congested, such as 120 ships competing for just 5 berths, the global attention mechanism performs much better because it can anticipate long-term queue delays. Finally, while the attention-based models take significantly longer to train over 30,000 samples, they both process decisions in less than a second during live testing. This makes them highly practical for real-time maritime scheduling.