EURO 2025 Leeds
Abstract Submission

2828. SupPy – a GPU-accelerated Python toolbox for superiorization of feasibility-seeking projection algorithms

Invited abstract in session TC-43: GenAI and Learning, stream Software for Optimization.

Tuesday, 12:30-14:00
Room: Newlyn GR.07

Authors (first author is the speaker)

1. Tobias Becher
Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ)
2. Yair Censor
Department of Mathematics, University of Haifa
3. Niklas Wahl
German Cancer Research Center (DKFZ)

Abstract

Superiorization is an approach to constrained optimization problems that perturbs an underlying feasibility-seeking projection algorithm to steer it toward a solution with reduced objective function value. Contrary to constrained optimization, the superiorization method does not target the constrained optimum, but aims to find a satisfactory solution of improved objective function value compared to the solution of the bare feasibility-seeking algorithm.
SupPy has three main modules: The first module implements flexible feasibility-seeking problems and projection algorithms. An object-oriented structure enables easy definition and solving of small geometric examples in various ways (sequentially, simultaneously, block-iterative and string-averaging), subgradient projections, split problems, as well as more complex implementations for various linear algorithms (Kaczmarz and Landweber methods, ARM, ART3+, etc).
The second module enables the definition of perturbation strategies, which can be combined with a feasibility algorithm via a third module which manages their combination in order to perform superiorization.
SupPy was tested on small geometric problems, as well as on real world applications like radiotherapy planning (RP) and image reconstruction (IR). All algorithms can be run both on the CPU and GPU, with the GPU allowing fast computation for problems heavily relying on large matrix multiplications (runtime of a few seconds for RP and IR with appropriate methods).

Keywords

Status: accepted

Back to the list of papers