Chronos-QRL: Outperforming Google OR-Tools, Attention Model, and POMO on the International Dynamic VRPTW Benchmark via 28-Qubit Floquet-MBL Superposition in 4GB VRAM
Year: 2026
Type: preprint
Abstract: The Dynamic Vehicle Routing Problem with Time Windows (Dynamic VRPTW) constitutes the most demanding international standard for evaluating autonomous dispatch algorithms for massive fleets under combinatorial uncertainty O(2^N). Classical local search metaheuristics (Google OR-Tools GLS and HGS-VRP) suffer an exponential growth in computation time when re-optimizing routes live, while state-of-the-art autoregressive deep models (Attention Model – AM, ICLR 2019; and POMO, NeurIPS 2020) are confined to the TC0 circuit class, collapsing under correlated congestion cascades of order k ≥ 6. In this work, we present Chronos-QRL, a Quantum Reinforcement Learning agent that maps the topological and temporal state of the Solomon RC101 graph (100 clients + central depot, Q = 200, Tmax = 240) onto an N = 28 qubit Floquet-MBL Time Crystal (268,435,456 states in 3.221 GB of VRAM) coupled with a Quantum Natural Policy Gradient optimizer (Fubini-Study Metric QNPG + SPSA) with zero Backpropagation. Audited over 16,000 dynamic epochs at a speed of 5,940,668 decisions/second (2.693 ms on a 4GB VRAM RTX 3050 Ti GPU), Chronos-QRL unsupervisedly discovers the critical 6-node subgraph, raises the bottleneck detection ROC-AUC from 0.5121 (AM/POMO collapse) to 0.9985 (+48.64 pts), and reduces the dynamic optimality gap relative to the Best Known Solutions (BKS) from +39.41% (Solomon I1), +23.01% (OR-Tools), and +20.55% (POMO) down to just +0.83% (95.6% BKS hit rate).
Language: English
Subfield: Industrial and Manufacturing Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222958461
Cites:
Cited by:
Related to: