COMET-Project Deepred (2019-2022)
Deep Learning based Predictive Analytics and Optimization
The recent success of deep learning approaches in the fields of image and speech recognition are driving the expectations of machine learning in other (industrial) fields of application as well, e.g., prediction of process and/or product parameters, development of virtual sensors or adaptive/self-learning assistance systems. In this project, such industrial applications will be improved on the basis of the existing know-how in the area of deep learning, representation learning and transfer learning. One focus is on the support of many similar processes with the aim of finding the right solution without expensive, comprehensive and process-specific data collection.
The optimal control of complex systems (production plants, smart grids, optimization algorithms, etc.) will be done more and more as model-based and predictive. The combination of prediction models based on machine learning and their integration into optimized and adaptive control algorithms at all levels of the automation pyramid represents a future-oriented approach. In this project, methods for the industrially suitable implementation of this approach are evaluated, whereby the uncertainty of predictions and estimates or the proof of stability plays an essential role.
- New methods for multi-task learning: Regarding the transfer-learning approaches, we aim to extend our state-of-the-art transfer learning (CMD) to multiple scenarios. This, together with research on scenario similarity, will allows us to transfer knowledge from more than one source scenario and to weight scenarios by their usefulness for the target task.
- Novel methods for online transfer learning: The online learning case requires methodologies from multiple research directions (learning without forgetting, increasing capacity to networks, trust in data, handling of real-world data...).
- Reinforcement learning based optimization framework: Regarding the optimization and control of complex tasks, we target the development of a reinforcement learning framework appropriate for optimization and control of complex tasks (i.e., tasks that may not easily be captured by model-based approaches).
This project is subsidized in the frame of COMET – Competence Centers for Excellent Technologies by BMVIT, BMDW, State of Upper Austria and its scientific partners. The COMET program is handled by FFG.