Avtomatizirana sinteza in analiza znanstvenih modelov

Nazaj na seznam za leto 2019

---

Oznaka in naziv projekta

N2-0128 Avtomatizirana sinteza in analiza znanstvenih modelov
N2-0128 Automating the Synthesis and Analysis of Scientific Models

Logotipi ARRS in drugih sofinancerjev

Projektni partnerji

• Institut "Jožef Stefan", Odsek za tehnologije znanja

Projektna skupina

• prof. dr. Sašo Džeroski, vodja projekta

• prof. dr. Ljupčo Todorovski

• prof. dr. Zoran Levnajić

• prof. dr. Michelangelo Ceci

• doc. dr. Panče Panov

• dr. Dragi Kocev

• dr. Jovan Tanevski

• dr. Martin Breskvar

• dr. Matej Petković

• Jure Brence

Vsebinski opis projekta

Slovensko:

Projekt SESAME se bo spopadel z zelo ambiciozno nalogo avtomatizacije znanstvenega modeliranja in omogočil, tako ljudem kot strojem, lažje prispevati k zakladnici znastvenih spoznanj. V obdobju revolucije podatkovnih znanosti in umetne inteligence, ko so podatki temelj za uporabo strojnega učenja v mnogih znanstvenih podvigih, se bo SESAME osredotočil na sintezo in analizo znanstvenih modelov. Glavna značilnost teh modelov je, da predstavljajo globoko razumevanje preučevanih sistemov in pojavov, so ljudem razumljivi in uporabljeni za razlago dotičnih tem. Po drugi strani trenutno najpopularnejši pristopi v strojnem učenju gradijo modele, ki ne morejo predstaviti razlogov za svoje odločitve. Tako znanstvenikom ne omogočajo razlage ali poglobljenega razumevanja modeliranih pojavov, kar je običajno poglavitni interes pri modeliranju v znanosti.

SESAME bo neposreden odgovor na potrebe znanstvenikov. Na področju dinamičnih sistemov bo integriral modeliranje na podlagi znanja, ki gradi obrazložitvene modele, ter modeliranje na podlagi empiričnih podatkov. Razvil bo metode strojnega učenja za sintezo modelov, ki bodo modele gradile iz podatkov in obstoječega znanja. Poleg tega bo razvil metode za analizo modelov, ki bodo znale razložiti obnašanje modelov pod različnimi pogoji. S tem bo pohitril ročno gradnjo in analizo modelov, in tako olajšal pot do novih znanstvenih odkritij.

Za dosego teh visokih ciljev bo SESAME predlagal formalizme za predstavitev znanstvenih modelov dinamičnih sistemov, njihovega obnašanja in domenskega znanja. Formalizmi bodo večnivojski in bodo dopuščali tako človeško razumevanje kot natančne simulacije sistemov. Podpirali bodo označevanje, shranjevanje in priklic modelov, njihovih komponent, obnašanj in tudi kriterijev kakovosti. S tem bo olajšana večkratna uporaba modelov ter ponovljivost procesov modeliranja. Moč pristopov strojnega učenja, ki bodo temeljili na teh reprezentacijah, bo predstavljena na študijah primerov iz zahtevnih področij uporabe.

Angleško:

SESAME will tackle the audacious task of automating scientific modeling, facilitating humans and machines in contributing to the pool of scientific knowledge. Set against the backdrop of the on-going revolution in data science and artificial intelligence, where machine learning uses data as the foundation for a wide range of scientific endeavours, SESAME will focus on the synthesis and analysis of scientific models. Such models embody deep knowledge about studied systems and phenomena, can be understood by humans, and used for explanation. In contrast, currently dominant machine learning approaches produce models that cannot explain their thinking, and leave the needs of scientists, whose very enterprise is founded on explanation, fundamentally unmet.

SESAME will directly respond to these unmet needs. Focusing on dynamical systems, it will integrate knowledge-driven modelling that builds explanatory models, and empirical data-driven modelling. It will develop machine learning approaches for model synthesis that will learn models from both data and existing knowledge, as well as for model analysis, seeking to explain model behaviour under different conditions. By doing so, it will alleviate the major bottleneck in science caused by the manual construction and analysis of models.

To achieve its grand goals, SESAME will propose representations for scientific models of dynamical systems, their behaviours, and domain specific modeling knowledge. Model representations will be multi-layered and allow for both human understanding, and precise simulation of system behaviour. They will support annotating, storing and retrieving models, model components, behaviours, as well as model quality criteria, facilitating their reuse and the reproducibility of modeling efforts. The power of the machine learning approaches, based on these representations, will be demonstrated through case studies in challenging application domains.

Osnovni podatki sofinanciranja so dostopni na spletni strani SICRIS.

Delovni sklopi projekta

• DS1: Predstavitev osnovnih gradnikov pri modeliranju dinamičnih sistemov (predznanje, modeli, prostori modelov, podatki oz. obnašanja in mere kvalitete modelov)
• DS2: Shranjevanje podatkov in predznanja pri modeliranju dinamičnih sistemov
• DS3: Strojno učenje za gradnjo modelov
• DS4: Strojno učenje za analizo modelov
• DS5: Vrednotenje razvitih pristopov in študije primerov uporabe le-teh

Bibliografske reference

• BRENCE, J., TODOROVSKI, L., DŽEROSKI, S. (2021). Probabilistic grammars for equation discovery. Knowledge-based systems.

• VERDEL, N., TANEVSKI, J., DŽEROSKI, S., MAJARON, B. (2020). Predictive model for quantitative analysis of human skin using photothermal radiometry and diffuse reflectance spectroscopy. Biomedical optics express.

• SIMIDJIEVSKI, N., TODOROVSKI, L., KOCIJAN, J., DŽEROSKI, S. (2020). Equation discovery for nonlinear system identification. IEEE access.

• MIGNONE, P., PIO, G., DŽEROSKI, S., CECI, M. (2020). Multi-task learning for the simultaneous reconstruction of the human and mouse gene regulatory networks. Scientific reports.

• BRENCE, J., TANEVSKI, J., ADAMS, J., MALINA, E., DŽEROSKI, S. (2020). Learning surrogates of a radiative transfer model for the sentinel 5P satellite. In Proc. 23rd International Conference on Discovery Science. LNCS 12323: 217-230.

• SZLUPOWICZ, M. A., BRENCE, J., ADAMS, J., MALINA, E., DŽEROSKI, S. (2020). Machine learning of surrogate models with an application to sentinel 5P. In Proc. 23rd International Multiconference Information Society. Volume A:104-107.

• KOSTOVSKA, A., TOLOVSKI, I., SIMIDJIEVSKI, N., TODOROVSKI, L., DŽEROSKI, S., PANOV, P. (2020). Process-based modelling of dynamical systems : an ontology of core entities (IJS delovno poročilo, 13150).

• KOSTOVSKA, A., TOLOVSKI, I., DŽEROSKI, S., PANOV, P. (2020). A system for semantic annotation and querying of datasets, models and experiments in the domain of process-based modelling of dynamical systems (IJS delovno poročilo, 13151).

• RADINJA, M., ŠKERJANEC, M., DŽEROSKI, S., TODOROVSKI, L., ATANASOVA, N. (2020). Automatization of urban drainage modelling by hybridising domain knowledge and equation discovery. In Book of abstracts iEMSs Conference.

• LUKŠIČ, Ž., TANEVSKI, J., DŽEROSKI, S., TODOROVSKI, L. (2019). Meta-model framework for surrogate-basedparameter estimation in dynamical systems. IEEE access.

• VERDEL, N., TANEVSKI, J., DŽEROSKI, S., MAJARON, B. (2019). Hybrid technique for characterization of human skin by combining machine learning and inverse Monte Carlo approach. In. Proc. European Conferences on Biomedical Optics, SPIE: 110751K-1-110751K-8.

• TOLOVSKI, I., KOSTOVSKA, A., SIMIDJIEVSKI, N., TODOROVSKI, L., DŽEROSKI, S., PANOV, P. (2019). Towards reusable process-based models of dynamical systems : a case study in the domain of aquatic ecosystems. In Proc. 42nd International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO).

• VERDEL, N., TANEVSKI, J., DŽEROSKI, S., MAJARON, B. (2019). A machine-learning model for quantitative characterization of human skin using photothermal radiometry and diffuse reflectance spectroscopy. In Proc. Photonics in Dermatology and Plastic Surgery, SPIE: 1085107-1-1085107-10.

• VERDEL, N., TANEVSKI, J., DŽEROSKI, S., MAJARON, B. (2019). Hybrid technique for characterization of human skin by combining machine learning and inverse Monte Carlo approach. Zbornik povzetkov 16. Slovenskega srečanja o uporabi fizike.

• KRALJ NOVAK, P., ŠMUC, T., DŽEROSKI, S. (2019). 22nd International Conference on Discovery Science. Proceedings LNCS 11828.

Magistrska dela

• TOLOVSKI, I. (2019). Designing, populating and querying repositories of semantically annotated computational experiments : master thesis = Načrtovanje, polnjenje in preiskovanje zbirk semantično označenih računalniških poskusov: magistrsko delo.

• KOSTOVSKA, A. (2019). Designing, populating and querying repositories of semantically annotated datasets : master thesis.

Odprto dostopna programska oprema

• ProGED

• Probmot

---

Nazaj na seznam projektov po letih