In this episode I am talking with Stefan Edelkamp from University Bremen (UHB). UHB are responsible for work package 4 and the IT-related partner in the FLOURplus consoritum. Being a kind of exotic in the area of traditional bakery industry, Stefan give us insides into his work and their specific taks within the project.
Stefan Edelkamp is a Professor in the Institute for Artificial Intelligence at the Faculty of Computer Science and Mathematics of the University of Bremen. He earned his Ph.D. from Freiburg University and has lead a junior research group at Technical University of Dortmund. His scientific interests are centered around Algorithmic Intelligence and include such areas as Heuristic Search, Action Planning, Game Playing, Machine Learning, Multi-Agent Simulation, Model Checking, External-Memory Algorithms, Parallel and Distributed Computing, Algorithm Engineering, Computational Biology, Decision Diagrams, Priority Queues, Navigation Systems, Network Security, and Intrusion Detection. Stefan Edelkamp has organized successful Workshops (e.g., MOCHART-06, SPIN-07), Conferences (e.g., KI-11, ICAPS-11) as well as Dagstuhl-Seminars (e.g., on Directed Model Checking and on Graph Search Engineering) and won several Performance Awards at International Planning Competitions. Together with Stefan Schroedl he is a co-author of the book “Heuristic Search - Theory and Applications” that has been recently published by Morgan Kaufmann / Elsevier Science.
Before going into details of the use of artificial intelligence in the baking industry, Stefan is introducing is in the term of artificial intelligence itself and how much it is used already in our daily life.
The study of computers being capable of doing work that humans think would require intelligence. The study of not being lost in new situations and of being able to adapt gathered knowledge onto new cases. I have a sticker on my computer „I'm teaching this machine to think“.
Nowadays, washing machines have controls based on fuzzy logic calculations, spam filters help concentrate on the more important aspects in live, recommender systems at Amazons or NetFlix broaden the view on previously unseen products, pricing policies at Wallmart show integral patterns (beer & chips go together), smartphone speech recognition allows for easier interfacing with connected devices, "Did you mean" at Google search covers common typos, personalized news and ads are injected on many webpages, data mining and filters help surveillance for agencies like NSA, autonomous robots are entering every-day live in household, at work, on streets or in sports.
UHB’s main work in work package 4 is the development of process tool and management, data inclusion and optimization. They do the core inference computation in the tool, which is seen as the contribution of FLOURplus - the main competitive advantage that SME-AGs and SMEs will receive from the project. They are trying to making sense of the analytical data for flour, process and products.
The process tool is developed in collaboration with the association in an agile and rather rapid prototyping style. So the final structure of the program will be an iterative design. However, we have a picture on how it looks like. The core idea is correlation and prediction to make the baking process more reliable and efficient. Flour quality varies over time. The flour type differs, and other aspects, too. We need supervised and unsupervised learning, where the teachers are analytical or sensory responses. Maybe active learning (asking questions to the user) can help to overcome the knowledge acquisition bottleneck. With help from our expert project partners at the ttz and sensory grading from consumers we have support with data collection and classification. Because flour types are a given for bakers and product specifications remain constant our focus lies solely on process parameter optimisation.
The development and programming of such a process tool will be realized in different steps and it will adhere to the process specification developed with the partner of the project. The first step is to infer correlations based on the baking trials at TTZ and UCC.
There is a mix of new and established analytical methods that is applied to all process steps for different products of the flour, dough and of the baked product: white bread and rolls. Second agreed step is the offer of entering flour specification from the miller into the system, for us to collect data for them to receive some feedback on the ingredients and suggested baking parameters.
UHB will first work with libraries like Weka or SVMlib to get first correlation results. Clustering and classification methods are first choice and the arrangement of values like Principle Component Analysis (Singular Value Decomposition based on Spectral Analysis/Eigenvector Study). There are Statistik-Tools like SPSS but also Machine Learning Tools like Weka that offer (PLS and) PCA.
The real-live setting of industrial and SME backing is a challenge for current ML approaches, as the data is sparse and confidential. Machine learning may be cast as applied statistics, so there is much noise to be filtered. Especially when data from different measurements is integrated at various times it can be challenging to appropriately tune the process parameters.