Predictive Analytics and Optimization

Concerning “Predictive Analytics and Optimization”, we are conducting research together with our scientific partners (e.g. JKU-FLLL) on the following topics, among others:

Transfer Learning

Research for transfer learning is directed to maximally leverage the available data before model deployment (offline learning) and the continuous improvement of these models after deployment (online learning). Regarding offline transfer learning we developed together with our scientific partner JKU-FLLL a new Deep Learning based Transfer Learning approach called Central Moment Discrepancy (CMD). It currently provides the state-of-the-art on several well-known benchmark problems for knowledge transfer between two scenarios.

Online Transfer Learning

In online transfer learning case the deployed models (offline models) are improved over time and new scenarios are added with little human interaction and minimum computational efforts. Recent advances in deep learning allow for the learning of additional tasks and scenarios without forgetting old ones, and without labor and time consuming retraining of the entire model. To utilize such methods in real world application, several application-related research problems are addressed.

Optimization and Control of Complex Tasks

Regarding the optimization and control of complex tasks, either repeated or time-dependent, we utilize and advance methodologies within the context of approximate dynamic programming and actor-critic learning. Such methodologies constitute alternatives of so-called reinforcement learning, and consist of two main parts: an approximation of a performance metric that captures the performance of a process subject to a set of optimization/control parameters, and a policy of selecting these parameters for either optimization or control.

Selected publications are summarized in the strategic project dasRES.