Risk-Aware Route Optimization Using Minimum Spanning Tree Initialization and a Transferable Reinforcement Learning Policy

Efficient route planning is a central challenge in applications such as autonomous area monitoring, infrastructure inspection, and logistics, where an agent must visit a set of locations while accounting not only for travel distance but also for spatially varying traversal risk. Classical distance-only optimizers minimize path length but are blind to such nongeometric costs. We propose a hybrid framework that integrates minimum spanning tree (MST)–based initialization with a transferable reinforcement-learning (RL) refinement policy for risk-aware routing. The MST provides a structured, low-cost initial tour; the RL agent then refines it under a multicriteria cost that combines distance and risk exposure. The key design choice is that the policy’s action-value function is linear in instance-independent, size-normalized features of a candidate edge exchange. As a result, a single 10-parameter weight vector is trained once and deployed on unseen instances and unseen problem sizes without any per-instance retraining. Across instances of 15–100 nodes (30 independent instances per size; 12 at n = 100), the proposed MST–RL method reduces multicriteria cost by 16%–27% over the MST baseline (p < 10⁻³ at every size) and decisively outperforms a randomly initialized RL policy (up to 44% at n = 100, p < 10⁻³), confirming the value of MST initialization. It attains cost competitive with, though not superior to, exhaustive local search and ant colony optimization while running 24–126 × faster per instance at n = 100, with inference time that does not grow with problem size. A reward-weight sweep demonstrates controllable distance–risk trade-offs. These properties make the method attractive for on-board, real-time deployment on resource-constrained hardware.