Project
Title
Modellbildung aus Experimentaldaten: Maschinelles Lernen und Modellevaluierung unter Abhängigkeiten und Verteilungsverschiebungen (Modellbildung)
Acronym
Modellbildung
Start
23.05.2018
End
31.07.2020
Coordinating Institute
Leibniz-Institut für Agrartechnik und Bioökonomie e.V. (ATB)
Coordinator
Allocated to research program
Summary
The analysis of experimentally obtained data is one central approach to gain new knowledge in the natural sciences. Modern machine learning methods have the potential to derive complex predictive models from very large experimentally generated data collections, and to discover novel relationships between variables of interest. However, common methodological tools in machine learning are based on independence and distributional assumptions which can be violated by experimental data in several ways. For example, experimental data is often influenced by our choice of experimental protocols and the time and location at which measurements ae carried out, because factors that are characteristic for a specific region or time period will leave traces the data that is being collected. Experimental data therefore always provide a view of reality that is shaped by the way in which measurements are carried out.
The aim of the project is the development of machine learning methods that explicitly reflect the measurement process and the resulting properties of experimental data. To this end, we develop approaches to correct spatial and temporal distribution shifts in data; we study the formal properties of the resulting algorithms and the complexity of the optimization problems that need to be solved. Results of the project will be used for solving data analysis problems in precision crop production and precision livestock farming. In crop production data, challenges include spatial variation (over individual fields or entire production regions) and temporal variation (because of short-term weather patterns or longer-term changes in temperatures); in livestock production, individual differences between observed animals often complicate data analysis. Within the project we furthermore collaborate with research groups from geophysics and cognitive psychology.
Funding
Deutsche Forschungsgemeinschaft (DFG)
Grant agreement number
LA3270/1-1
Funding framework
DFG Nachwuchsgruppe im Emmy Noether-Programm