Livoa LogoLivoa
Online Network
Target Network
Prioritized Experience Replay
Loss Function
Knowledge-Integration Module
argmaxa Q(s, a; θ)
Agent 1
Agent 2
Agent n-1
Agent n
Environment
Cooperation Module
Simulator
Initial Solution
Time-Aware Variable Neighborhood Search
Shaking
Local Search
Final Solution

Routing Simulator (Environment Model)

• Dynamic requests

• Travel/energy/time updates

State Representation + SAP

• Encode customers/depots/EVs

• Mask infeasible actions

Multi-Head Attention (MHA) Encoder

• Contextual dependencies

• Attention-based embedding

MARL — DDQN + PER (CLDE)

• Centralized learning

• Decentralized execution

Double-Adaptive VNS (DA-VNS)

• Adaptive shaking

• Adaptive VND refinement

Outputs / Results

• Optimized routes

• Distance / service rate

• Time-window feasibility

• Explanations

Explainable AI (XAI) Layer

• Attention viz (MHA)

• SHAP/LIME feature importance

• Operator contribution (DA-VNS)

nn

by hh

0
0 uses