VERICOMP 2.0: Comparing and Recommending Verified IVP Solvers in a Flexible Way (Vortrag)

Aus Kompetenzportal
Wechseln zu: Navigation, Suche
VERICOMP 2.0: Comparing and Recommending Verified IVP Solvers in a Flexible Way (Vortrag)
Autor Ekaterina Auer, Lorenz Gillner, Wolfram Luther, Andreas Rauh
Tagung/Veranstaltung SCAN 2020
Ort Szeged, Ungarn (online)
vom 13. September 2021 bis 15. September 2021
Review editorial
Titel Proceedings
Autor Proceedings
Verlag
Erscheinungsjahr 2021

Methods with result verification, for example, interval analysis, have been applied in engineering since the 1970s at the latest. They help not only to prove automatically that computer-obtained results are correct, but also to represent bounded uncertainty and propagate it through system models in an easy-to-understand, deterministic manner. After over half a century of research, many up-to-date libraries are available (and still emerging), implementing the concepts for a variety of programming languages and computer algebra systems such as C++, Python, Matlab, Julia. Regardless of the merits of such methods and their general accessibility, they are rarely used outside of the university context or cooperations.

Aside from the necessity to learn new material, an impediment on the way of a larger application of verified techniques in industrial engineering practice is the lack of information which of the available tools to choose for a given task. Sometimes, making an inappropriate choice for a program or approach can lead to too conservative results discouraging the use of the whole branch of methods. To improve the situation at least in one area, we have been working on a web-based platform VERICOMP for promoting, comparing and recommending verified initial value problem software for ordinary differential equations (IVPS) for over a decade. Almost as a by-product, VERICOMP offers developers of new IVPS a possibility to compare their solvers with the established ones. Here, VERICOMP can be of use for facilitating such projects as ARCH-COMP (cps-vo.org/group/ARCH/FriendlyCompetition), a competition on verifying continuous and hybrid systems.

Differential equations are fundamental in many applied areas of science as a mathematical model for dynamic systems or processes, with techniques for comparing traditional, non-verified software available since the seventies (archimede.dm.uniba.it/~testset/testsetivpsolvers/). Some of the challenging tasks on the way to highlight advantages of various tools are to develop a standard set of problems, a set of fair criteria and testing conditions, a means to easily incorporate new software into the comparison, and, last but not least, a means to present and visualize the gathered information. The goal is to allow a user to obtain knowledge easily and to grasp it immediately. Moreover, the maintenance of these sets and means should be as flexible as possible facilitating, for example, entry of new (specialized or standard) problems or replacement of criteria if the need arises.

In this talk, we focus on the current version of VERICOMP available at vericomp.fiw.hs-wismar.de that enhances the previous one especially from the point of view of flexibility. Additionally, we demonstrate how ideas from the area of recommender systems can be employed to produce an automatic suggestion about the right tool to use for a given application. Moreover, we discuss in general what kind of metadata, data, quality criteria, metrics, and visualization are required to be able to compare and recommend IVPS. Finally, we give an outlook on the possibility of easier management and comparability of different solvers inside VERICOMP based on containerization. This task requires standardisation of data flow and IVPS interfaces (which are quite non-uniform at the moment).