Direction-aware deep policy learning for efficient capacitated arc routing

Feng Xue

doi:10.1016/j.engappai.2026.114695

Direction-aware deep policy learning for efficient capacitated arc routing

2026-04-15

0

PRE

AI

DOI：10.1016/j.engappai.2026.114695

Original

Original request for help

Save

Abstract

En 中文

• First direction-aware single-stage neural solver (DaAM) that directly decides arc and direction under capacity for CARP. • Supervised-then-RL scheme (pre-train on expert labels, PPO with self-critical baseline) stabilizes learning and speeds convergence. • Plug-and-play dynamic-programming path optimizer improves depot-return decisions and further lowers tour cost. • New Beijing-CARP benchmark with 20k training and 10k test instances per scale enables fair evaluation of learning-based CARP. • Near meta-heuristic quality with much faster inference: ≤7.28% gap to MAENS on medium scales; strong large-instance generalization.

Keywords:

Direction-aware

Single-stage neural solver

Capacitated Arc Routing

Reinforcement Learning

Dynamic Programming

Direction-aware deep policy learning for efficient capacitated arc routing

Abstract

Journal

Organization