Dynamic Hybrid Temporal Attention With Attention on Attention for the Capacity Vehicle Routing Problem

With the increasing utilization of deep reinforcement learning (DRL) in recent studies to solve vehicle routing problem, it has been observed that these models mainly adopt an encoder and decoder structure. However, the time and space features embedded in encoder and decoder are not fully considered. To address these limitations, we propose a dynamic hybrid temporal attention (DHTA) model that incorporates an attention on attention (AOA) mechanism to tackle the capacity vehicle routing problem. A noted feature of DHTA is its utilization of a multidimensional attention mechanism based on AOA within encoder, along with a temporal pointer decoder that dynamically focus on the decision-making information. Experimental results demonstrate that our model exhibits superior performance compared to state-of-the-art DRL-based models and several traditional heuristic algorithms.