Bi-Weekly Talk: Tarik Viehmann: Combining Knowledge-based Control and Data-Driven Models in Industrial Domains

Mittwoch, 22.02.2023, 10.30 Uhr

Ort: RWTH Aachen University, Informatikzentrum - Ahornstr. 55, Erweiterungsgebäude E3, Raum 9u10

Vortragender: Tarik Viehmann

 

Abstract:

Knowledge-based systems enable domain experts to shape the program flow by their experience and preferences. However, the design of proper domain models is typically difficult and driven by manual labor. Additionally, some important dynamics of a domain can be either unknown or too complex to model.
We will take a look at applications from robotics and manufacturing, where knowledge-driven systems are well-suited provide top-level control, but where some decisions cannot easily be made by domain experts.

Firstly, when forging metal pieces to optimize workpiece properties for use in safety critical applications, assistance systems using Case-based Reasoning (CBR) allow forging experts to gear the design of forging strategies according to their preferences. However, CBR systems require users to define similarity measures for all relevant process parameters, some of which may lack a proper model to estimate the impact on the desired target parameters. Data-driven methods for clustering, classification and similarity synthesis are promising approaches to derive the local similarities for these parameters.

Secondly, serving customer requests in smart warehouse and factory settings is a common task in autonomous robotics. The RoboCup Logistics League replicates common challenges, such as uncertain online arrival of customer requests, bottlenecks through limited production capacities or unforeseen delays and failures during operation. Existing goal reasoning approaches leverage the knowledge of domain experts to make dynamic decisions on the shopfloor to control and coordinate multiple robots. However, optimizing priorities of individual tasks in face of the randomized and partially unknown demand is difficult. One way to overcome the issue is to automatically learn the priorities over a larger set of simulated or real games.