VA
V. Anica-Popa
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
1 records found
1
Global-Local Attention vs Graph Neural Networks in the Reinforcement Learning Approach for the Dynamic Berth Allocation Problem
Bridging the Optimality Gap in Dynamic Berth Allocation Problem via Global-Local Attention
The Dynamic Berth Allocation Problem (DBAP) is an optimization problem in maritime logistics that seeks to minimize vessel delays and improve terminal efficiency through effective berth scheduling. This paper investigates how replacing the Graph Neural Network (GNN) encoder in a reinforcement learning approach to the DBAP with a Global-Local Attention mechanism affects the optimality gap. This new encoder architecture is designed to capture both local vessel-berth interactions and global scheduling information. The performance of the proposed architecture is evaluated against the baseline GNN across 2,700 benchmark instances spanning various terminal sizes and traffic congestion levels. Empirical results demonstrate that the Global-Local Attention variant yields a consistent improvement in schedule quality, reducing normalized operational costs by an average of 2.8%. While this modification provides notable optimization gains, particularly in lower congestion scenarios, it introduces an 81.4% increase in trained parameters, presenting a distinct trade-off between scheduling optimality and computational efficiency.
...
The Dynamic Berth Allocation Problem (DBAP) is an optimization problem in maritime logistics that seeks to minimize vessel delays and improve terminal efficiency through effective berth scheduling. This paper investigates how replacing the Graph Neural Network (GNN) encoder in a reinforcement learning approach to the DBAP with a Global-Local Attention mechanism affects the optimality gap. This new encoder architecture is designed to capture both local vessel-berth interactions and global scheduling information. The performance of the proposed architecture is evaluated against the baseline GNN across 2,700 benchmark instances spanning various terminal sizes and traffic congestion levels. Empirical results demonstrate that the Global-Local Attention variant yields a consistent improvement in schedule quality, reducing normalized operational costs by an average of 2.8%. While this modification provides notable optimization gains, particularly in lower congestion scenarios, it introduces an 81.4% increase in trained parameters, presenting a distinct trade-off between scheduling optimality and computational efficiency.