Interactive Machine Learning with Evolving Fuzzy Systems

The major goal of this project is to develop a new methodological framework for overcoming the current limitations of on-line machine learning (ML) systems in industrial installations, social media platforms, health-care systems, web mining tools, predictive maintenance frameworks etc. Currently, ML systems are mostly oriented more on a precise on-line processing functionality where continuously arriving data streams are processed and high-qualitative models are learnt from them for various purposes such as decision support, forecasts of states, classifications, quality control etc. Indeed, outputs of these learning processes and/or internal model building stages may be shown to the user, but currently this is basically restricted within a passive supervision frontend, at most allowing some rudimentary feedback by human users (cold interaction). However, current systems do not foresee an advanced interaction and communication methodology, where the human is stimulated and would be thus willing and able to bring in her/his knowledge about the process. It is expected that, within an advanced interactive system, both, humans and machines, benefit from each other, achieving knowledge gains for humans as well as performance boosts for the ML models likewise. This request induces several basic research questions and challenges on methodological level --- to be tackled in this project within 4 work-packages (and associated goals). Especially, new methodologies need to be developed for achieving a better understandability of the models, for keeping human feedback on an economic level and for a proper automated integration of various forms of human input into the adaptive evolving models --- this includes an appropriate handling of human uncertainty and vagueness and will lead to a new hybrid stream learning paradigm, which is expected to have significant impact for the interactive ML and the evolving modelling community. A prototypical software tool with an appropriate GUI frontend is planned to be established, which will be used for evaluation and test purposes on several real-world applications (Goal 5, WP 5). Upon successful establishment in this project, it is thus expected to be appealing for several industrial partners of the proposers department and will have thus also an impact on further application-based projects. It will be also made accessible to the research community. The responsibility will be in the hands of a renowned international researcher, namely Dr. Edwin Lughofer, who co-authored several publications (around 50 journal papers) in the fields of evolving (fuzzy) models, interpretability as well as active learning, and also has long-term connections to and collaborations with international research facilities working on the same and/or similar research topics. A Post-Doc and a PhD student are required to be employed in the department for the realization of all project goals.