© iStockphoto
We at Fraunhofer ITWM are mostly mathematicians and do research in different fields. Two major fields and methods are »Artificial Intelligence« (AI) and »Machine Learning« (ML). Our goal is to improve and understand these methods and to open up new fields of application. This also includes the application in business processes.
For many, the words »Artificial Intelligence« (AI) as well as »Machine Learning« (ML) are not really tangible. In this context, the word »black box« often comes up with the thought: »We put our data into the AI methods and at the end we get a very good prediction/evaluation/analysis of our data.«
Unfortunately, this is not quite the case in practice. Sure we can analyze data faster using AI methods. However, we will not find out information that is not already contained in the data itself. Therefore, the use of AI always stands or falls with the quality of the data.
We also need the expertise of the industry for every analysis and interpretation or for setting up the models. Without this, we would perhaps also prove correlations that do not even exist.
Consequently, the basis of our AI methods is the best possible quality of the data (we will explain exactly what this means in one of the next blog posts). Likewise, understanding this data is essential. Without these two prerequisites, a good design of AI methods is difficult.
When we talk about artificial intelligence, this buzzword is probably understood differently by everyone, yet everyone talks about it. With the term algorithm, this is different again. Some shy away from it, to others it is the epitome of AI.
If we consult Wikipedia, we get the following definition: »An algorithm is a unique set of actions for solving a problem or class of problems.«
Artificial Intelligence is basically a stringing together of different action rules with the goal of achieving the best description of the data. What exactly »the best description of the data« means in this context, then depends on the description of the problem.
Problems that we will work on and address here come from the areas of »conspicuity detection«, »time series« and »explainability«. In these individual areas we will look at this topic in some detail.
In recent weeks and months, there has been repeated talk of boosting Germany's digitization. It is precisely this boost within the administration and business processes that generates a pool of data that must be regularly analyzed and assessed by experts.
For example, the introduction of e-bills has made it possible for the first time to use algorithms to check the data itself and to have relevant data checked by experts. Through more and more digital communication, it is also possible to learn the current assignment of messages to clerks, as well as the detection of unnecessary e-mail traffic by directly sending the message to the people who really need it. Without this data being available digitally, this is only possible to a limited extent or with enormous effort.
If this data is now available digitally, this is the best prerequisite for machine learning (ML). Simple statistical analyses based on small amounts of data are available in almost every company. With these digitally acquired, enormous amounts of data, machine learning processes are now also possible in many companies. To what extent the data can be used for the relevant problems with the already existing methods is another question.
This not only offers the opportunity to put the algorithms into practice, but also to develop and research new procedures and analysis methods for questions of this kind.
If the computer can already find everything and provide better forecasts and evaluations than the expert, what is their job anyway? Will they all be replaced now? We read such questions again and again. To answer them, it is essential to know that people program the algorithms. These search for an optimal solution, whereby optimal is determined again by the programmers. Predictable events can already be taken into account. New, unpredictable changes are difficult to incorporate. AI therefore makes the work of the administrator easier, but the administrator is never completely replaced, in order to intercept the unpredictable.
If changes take place in the data, the algorithm must also learn these. This also requires the information of the people, who make the correct assessment through their own experience.
As we can see, there is a change in work, but not a replacement.
In our project EP-KI (decision support for business processes with the help of new AI methods) we develop new algorithms which are relevant for the application. For this purpose, we have targeted the following application fields:
We will work on these with the following research fields:
We have marked topics that exactly fit into these research fields accordingly. Thus, it will be possible to find quick answers for specific questions. We are looking forward to your feedback!
Fraunhofer Institute for Industrial Mathematics ITWM