LLM-Assisted Offline Reward Design for Reinforcement Learning-Based Irrigation Scheduling
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171024 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Qiufen Wang und 2 weiteren Autor:innen.
Abstract (Originalsprache)
This study develops an LLM-assisted offline workflow for reward design in reinforcement learning-based irrigation scheduling under a seasonal water allocation. Researchers provide agronomic knowledge and feedback on candidate-policy behavior, and the language model translates these inputs into executable rewards. Eight candidates are compared over four generations. The selected reward, aridity–soil curriculum, and FAO-56 exploration mask are held fixed during final policy training and transfer evaluation. Across 24 transfer scenarios comprising two withheld sites, three soil textures and four seasons, proximal policy optimization achieves a mean net benefit of 1943 USD ha⁻¹, 2.8 per cent above the strictest tuned rule and 9.1 per cent above the untuned FAO-56 rule. The paired mean gain is 52.5 USD ha⁻¹, with site means of +127.9 at Lubbock and −23.0 at Urbana. Grouping shared-weather outcomes into eight site–season blocks gives a two-sided sign-flip p-value of 0.117; joint resampling of sites, seasons, and training seeds gives an exploratory percentile range of −43.3 to 159.6 USD ha⁻¹. The combined curriculum and exploration mask increase convergence frequency from 5/10 to 10/10 runs in the evaluated ablation. The results connect offline reward design with stage-sensitive water allocation and show how its scheduling value varies between semi-arid and humid conditions.
Zitieren
APA
Wang, Q., Wang, T., & Wang, Z. (2026). LLM-Assisted Offline Reward Design for Reinforcement Learning-Based Irrigation Scheduling. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171024
Vancouver
Wang Q, Wang T, Wang Z. LLM-Assisted Offline Reward Design for Reinforcement Learning-Based Irrigation Scheduling. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171024
BibTeX
@article{Wang2026,
author = {Wang, Qiufen and Wang, Tianhao and Wang, Ziyue},
title = {LLM-Assisted Offline Reward Design for Reinforcement Learning-Based Irrigation Scheduling},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171024}
}
EUVIMED