Vehicle Routing Problem

Strategic Research Report

Combinatorial Optimization
State of the Art

From "XL" scale metaheuristics to Agentic AI designers. A comprehensive analysis of the vehicle routing landscape in 2024–2025.

Paradigm Shift: Hybridization

The boundary between "Learning" and "Searching" has eroded. The SOTA is no longer pure OR or pure AI, but a symbiotic fusion.

HGS (Genetic Search)

Undisputed champion for quality. Uses "Giant Tour" representation and "Catastrophe" diversity management.

Agentic AI

LLMs don't solve the route; they write the solver code. Enables "Zero-Shot" custom constraints.

Scaling to "XL" (10k Nodes)

Optimality Gap (%)

While Exact methods break and NCO struggles at scale, Hybrid Genetic Search (PyVRP) maintains <1% gap even at 10,000 nodes.

The Algorithmic Core

The mechanics powering the 2025 leaderboard.

🧬

Hybrid Genetic Search

The King
  • Giant Tour: Ignores capacity initially to traverse search space.
  • Split Algorithm: Uses shortest-path on DAG to segment tour into valid routes.
  • PyVRP: The SOTA implementation bridging C++ speed with Python ease.
🧠

Neural Optimization

The Frontier
  • Diffusion Models: Generates entire solution structure simultaneously (non-autoregressive).
  • Generalization: Uses "Multi-Expert" modules to handle varying constraint tightness.
📐

Exact Methods (BCP)

The Proof
  • Neural Pricing: Uses ML to accelerate column generation.
  • RouteOpt: Modular solver proving optimality for 360-400 node instances.
1. Interpreter Agent

Reads natural language: "Drivers need a 15m break every 3h."

2. Coder Agent

solver.Add(drive_time <= 180)

3. Refiner Agent

Tests synthetic cases. Debugs errors. Optimizes params.

The Agentic Shift

In 2025, we stop asking AI to "solve" the problem and start asking it to "write the solver." Frameworks like AFL can automatically generate valid code for 91.67% of custom VRP variants, democratizing access to complex optimization.

Industrial Driver

Electric Vehicle Routing

Physics dictates the algorithm. Solvers must now handle non-linear charging curves and battery degradation.

Non-Linear Charging Piecewise Linear approx.
Last-Mile Dynamic Q-Learning (RL)

Practitioner's Guide

Offline Planning

Quality First

  • Goal: Save 1-2% fleet cost
  • Time: Minutes/Hours
  • Tool: PyVRP

Real-Time Dispatch

Speed First

  • Goal: Instant Assignment
  • Time: Milliseconds
  • Tool: VROOM (C++)

Enterprise Scale

Size First

  • Goal: 30,000+ Nodes
  • Tech: List Variables
  • Tool: Hexaly

Complex Logic

Flexibility First

  • Goal: Weird Constraints
  • Tech: CP / Agentic AI
  • Tool: OR-Tools / AFL

Synthesized from "The State of the Art in Combinatorial Optimization" (2025)