EUVIMEDEUVIMED entdecken
Crossref

Conflict-driven Adaptive Scaling for Safe Residual Reinforcement Learning

Khrystyna Mykich, Mykhailo Kopylets

Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì · 22.06.2026 · 19 · 444-457

DOI: 10.23939/sisn2026.19.444 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì, online erschienen am 17.05.2026. Verfasst von Khrystyna Mykich und 1 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Residual reinforcement learning combines a pre-trained base policy with residual corrections, but most existing approaches rely on a fixed scaling factor K, which creates a fundamental dilemma. A high K may destabilize an otherwise safe base policy. A low K, in contrast, prevents the residual policy from correcting errors effectively. This paper introduces CARS, or Conflict-Aware Residual Scaling, an adaptive methodology that dynamically adjusts K according to action conflict and safety constraint violations. It also presents Conflict-Aware Lab, a purpose-built information system for safe reinforcement learning research. CARS includes two components. The Kco∣nf component uses cosine similarity to detect disagreement between the base and residual policies. When conflict is observed, it increases residual influence. The Krisk component leverages Lagrangian multipliers to respond to safety violations. Conflict-Aware Lab integrates PyBullet-based UAV simulation, a two-layer safety architecture that combines Lagrangian constraints with CARS scaling, and experiment management tools. The system stores run parameters and evaluation outcomes, which supports reproducibility. Experiments on four drone navigation environments in PyBullet showed that CARS consistently outperformed fixed K baselines. The proposed approach improved route completion success and reduced collision frequency. The adaptive mechanism repeatedly selected scaling values above the baseline level. This confirmed its sensitivity to contextual signals and its ability to react flexibly to changing evaluation conditions. The results indicate that dynamic policy blending driven by conflict and safety signals improves both task performance and operational safety in reinforcement learning. Conflict-Aware Lab, in turn, provides reproducible infrastructure for this line of research

Zitieren

APA

Mykich, K., & Kopylets, M. (2026). Conflict-driven Adaptive Scaling for Safe Residual Reinforcement Learning. Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì, 19, 444-457. https://doi.org/10.23939/sisn2026.19.444

Vancouver

Mykich K, Kopylets M. Conflict-driven Adaptive Scaling for Safe Residual Reinforcement Learning. Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì. 2026;19:444-457. doi:10.23939/sisn2026.19.444

BibTeX
@article{Mykich2026,
  author = {Mykich, Khrystyna and Kopylets, Mykhailo},
  title = {Conflict-driven Adaptive Scaling for Safe Residual Reinforcement Learning},
  journal = {Vìsnik Nacìonalʹnogo unìversitetu "Lʹvìvsʹka polìtehnìka". Serìâ Ìnformacìjnì sistemi ta merežì},
  year = {2026},
  volume = {19},
  pages = {444--457},
  doi = {10.23939/sisn2026.19.444}
}
Kostenlos starten →