Routing Simulator (Environment Model)• Dynamic customer requests• Travel time & energy model• Time progression & event updates

State Representation + Knowledge Guidance (SAP)• Encodes customers/depots/EVs• Masks infeasible actions (battery/time)

Multi-Head Attention (MHA) Encoder• Contextual dependencies• Multi-head attention produces embeddings

MARL — DDQN + PER (CLDE)• Double DQN for Q-values• Prioritized Experience Replay• Centralized learning / decentralized execution

Double-Adaptive VNS (DA-VNS)• Adaptive shaking (diversify)• Adaptive VND (intensify)• Operator weight updates

Outputs / Results• Optimized routes• Distance, service rate, time-window metrics• Human-interpretable insights

Explainable AI (XAI) Layer• Attention visualization (MHA)• Feature importance (SHAP/LIME)• Neighborhood contribution (DA-VNS)

Fig. 2. General architecture of the proposed framework integrating MHA-enhanced MARL, Double-Adaptive VNS (DA-VNS), and an Explainable AI layer for DMDEVRPTW.

nnd