Tea manufacturing: Blends of raw materials with variable qualities
Avoiding production errors is not the issue at the Martin Bauer Group; rather, it is how to manufacture products of consistent quality using raw materials with varying properties. The group produces herbal and fruit tea blends for supermarkets and drugstores. As the ingredients of the botanical raw materials vary, warehouse and production planning are very time-intensive operations at the Martin Bauer Group. We solved this pooling problem for tea blends with optimization software that takes into account stock levels, storage periods, laboratory analytics, intermediate and end products, and the various quality requirements of customers. With the solution jointly developed by researchers from our working group and FAU Erlangen-Nürnberg, the dispatchers are able to quickly run through various scenarios, which otherwise would have too many combinations to be solved by humans alone. This kind of problem is not limited to tea manufacturing either, but also arises in many other areas of the food industry and in industrial manufacturing.
OBER: Optimal inventory planning quantifies uncertainties of forecasts
“Out of stock” has been an oft-heard refrain in recent times. Wood, bathroom fittings, canned vegetables and toilet paper are not in stock when customers come in to buy them. At the same time, goods that are not in demand are taking up valuable space. Before now, businesses have usually relied on very simple forecasts based on average sales to date, even though these predictions are riddled with uncertainties. In the OBER research project, we combine forecasts specially designed for the wholesale sector with mathematical optimization, taking into account restrictions such as the best price, available storage space and financial resources. Moreover, the AI we developed quantifies the uncertainty of the prediction. It calculates the optimum strategic course of action even for goods that will not be ordered for a few months.
AutoML – automatic selection of the best model at any given moment
Finding the most suitable mathematical procedure for any given application is time-consuming. For a solution, we looked to AutoML (automated machine learning). We use an umbrella model that automatically analyzes the various algorithms and independently selects the most suitable model. With Online AutoML, moreover, it is possible to continuously review whether the model currently being used is still the best one. Because when production suddenly changes the recipe for gingerbread or if a different car model is to be manufactured, then another machine learning algorithm might be better suited to the new task. AutoML is therefore versatile and can be used in many domains, as its abstraction at the mathematical level works for many applications.