Contribution to a conference proceedings/Contribution to a book

FZJ-2026-01057

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Parallel Reinforcement Learning and Gaussian Process Regression for Improved Physics-Based Nasal Surgery Planning

Rüttgers, M. (Corresponding author) ; Hübenthal, F. ; Tsubokura, M. ; Lintermann, A.FZJ*

2025
Springer Nature Switzerland Cham
ISBN: 978-3-031-85702-7 (print), 978-3-031-85703-4 (electronic)

This record in other databases:  

Please use a persistent id in citations: doi:10.1007/978-3-031-85703-4_6  doi:10.34734/FZJ-2026-01057

Abstract: Septoplasty and turbinectomy are among the most frequentbut also most debated interventions in the field of rhinology. A previouslydeveloped tool enhances surgery planning by physical aspects of respi-ration, i.e., for the first time a reinforcement learning (RL) algorithm iscombined with large-scale computational fluid dynamics (CFD) simula-tions to plan anti-obstructive surgery. In the current study, an improve-ment of the tool’s predictive capabilities is investigated for the afore-mentioned types of surgeries by considering two approaches: (i) trainingof parallel environments is executed on multiple ranks and the agentsof each environment share their experience in a pre-defined interval and(ii) for some of the state-reward combinations the CFD solver is replacedby a Gaussian process regression (GPR) model for an improved compu-tational efficiency. It is found that employing a parallel RL algorithmimproves the reliability of the surgery planning tool in finding the globaloptimum. However, parallel training leads to a larger number of state-reward combinations that need to be computed by the CFD solver. Thisoverhead is compensated by replacing some of the computations withthe GPR algorithm, i.e., around 6% of the computations can be savedwithout significantly degrading the predictions’ accuracy. Nevertheless,increasing the number of state-reward combinations predicted by theGPR algorithm only works to a certain extent, since this also leads tolarger errors.

---

Contributing Institute(s):

1. Jülich Supercomputing Center (JSC)

Research Program(s):

1. 5111 - Domain-Specific Simulation & Data Life Cycle Labs (SDLs) and Research Groups (POF4-511) (POF4-511)
2. SDLFSE - SDL Fluids & Solids Engineering (SDLFSE) (SDLFSE)
3. HANAMI - Hpc AlliaNce for Applications and supercoMputing Innovation: the Europe - Japan collaboration (101136269) (101136269)

---

Database coverage:
; ; Nationallizenz ; SCOPUS

---