Constrained Reinforcement Learning for the Dynamic Inventory Routing Problem under Stochastic Supply and Demand

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hasturk, Umur, Schrotenboer, Albert H., Roodbergen, Kees Jan, Ursavas, Evrim
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915700288782336
author Hasturk, Umur
Schrotenboer, Albert H.
Roodbergen, Kees Jan
Ursavas, Evrim
author_facet Hasturk, Umur
Schrotenboer, Albert H.
Roodbergen, Kees Jan
Ursavas, Evrim
contents Green hydrogen has multiple use cases and is produced from renewable energy, such as solar or wind energy. It can be stored in large quantities, decoupling renewable energy generation from its use, and is therefore considered essential for achieving a climate-neutral economy. The intermittency of renewable energy generation and the stochastic nature of demand are, however, challenging factors for the dynamic planning of hydrogen storage and transportation. This holds particularly in the early-adoption phase when hydrogen distribution occurs through vehicle-based networks. We therefore address the Dynamic Inventory Routing Problem (DIRP) under stochastic supply and demand with direct deliveries for the vehicle-based distribution of hydrogen. To solve this problem, we propose a Constrained Reinforcement Learning (CRL) framework that integrates constraints into the learning process and incorporates parameterized post-decision state value predictions. Additionally, we introduce Lookahead-based CRL (LCRL), which improves decision-making over a multi-period horizon to enhance short-term planning while maintaining the value predictions. Our computational experiments demonstrate the efficacy of CRL and LCRL across diverse instances. Our learning methods provide near-optimal solutions on small scale instances that are solved via value iteration. Furthermore, both methods outperform typical deep learning approaches such as Proximal Policy Optimization, as well as classical inventory heuristics, such as (s,S)-policy-based and Power-of-Two-based heuristics. Furthermore, LCRL achieves a 10% improvement over CRL on average, albeit with higher computational requirements. Analyses of optimal replenishment policies reveal that accounting for stochastic supply and demand influences these policies, showing the importance of our addition to the DIRP.
format Preprint
id arxiv_https___arxiv_org_abs_2503_05276
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Constrained Reinforcement Learning for the Dynamic Inventory Routing Problem under Stochastic Supply and Demand
Hasturk, Umur
Schrotenboer, Albert H.
Roodbergen, Kees Jan
Ursavas, Evrim
Optimization and Control
Green hydrogen has multiple use cases and is produced from renewable energy, such as solar or wind energy. It can be stored in large quantities, decoupling renewable energy generation from its use, and is therefore considered essential for achieving a climate-neutral economy. The intermittency of renewable energy generation and the stochastic nature of demand are, however, challenging factors for the dynamic planning of hydrogen storage and transportation. This holds particularly in the early-adoption phase when hydrogen distribution occurs through vehicle-based networks. We therefore address the Dynamic Inventory Routing Problem (DIRP) under stochastic supply and demand with direct deliveries for the vehicle-based distribution of hydrogen. To solve this problem, we propose a Constrained Reinforcement Learning (CRL) framework that integrates constraints into the learning process and incorporates parameterized post-decision state value predictions. Additionally, we introduce Lookahead-based CRL (LCRL), which improves decision-making over a multi-period horizon to enhance short-term planning while maintaining the value predictions. Our computational experiments demonstrate the efficacy of CRL and LCRL across diverse instances. Our learning methods provide near-optimal solutions on small scale instances that are solved via value iteration. Furthermore, both methods outperform typical deep learning approaches such as Proximal Policy Optimization, as well as classical inventory heuristics, such as (s,S)-policy-based and Power-of-Two-based heuristics. Furthermore, LCRL achieves a 10% improvement over CRL on average, albeit with higher computational requirements. Analyses of optimal replenishment policies reveal that accounting for stochastic supply and demand influences these policies, showing the importance of our addition to the DIRP.
title Constrained Reinforcement Learning for the Dynamic Inventory Routing Problem under Stochastic Supply and Demand
topic Optimization and Control
url https://arxiv.org/abs/2503.05276