What is "Machine Learning"?
Machine learning is the branch of artificial intelligence that deals with algorithms that aim at enabling machines to learn autonomously from data. These algorithms are mostly statistical based prediction algorithms and they are not new. For instance, the beginning of artificial neural networks dates back to the 1950s. The reason why artificial intelligence, particularly machine learning, is so hyped at the moment is that progresses in computer technology rendered it possible to acquire and process huge data sets in real time. This is a necessary requirement for statistical machine learning. I made the experience that many people associate AI and machine learning with human-like aware robots that we know from science-fiction movies such as Terminator or Star Trek. So, let me make this crystal clear, we are by no means close to develop such robots. Machine learning refers to statistical computer programs that are able to solve specific tasks in an intelligent manner.
Why do you want to use machine learning algorithms for plausibility checks?
First of all, plausibility checks belong to the typical task's clerk. Clerks often receive long lists and need, e.g., to check whether invoices can be paid out or need to be corrected. Usually, such decisions are repetitive and are based only on a very sparse set of criteria. These are optimal conditions for statistical learning algorithms.
Which plausibility checks do you specifically want to automatize?
Well, our first project deals with plausibility checks for meter readings. Let me give you some background information: energy companies measure your energy consumption by a technical device – a meter. These measurements can be inaccurate because of technical or human errors. Some energy companies even ask customers to submit the meter reading on their own. Possibly, you know that from experience. In order to detect false measurements early, there is a detection system integrated in the software systems of energy companies. This detection system marks meter readings as potentially implausible. Then, a clerk has to check if these potentially implausible readings are likely to be really false or not. If they are likely to be false the customers is contacted, otherwise an invoice is sent to the customer. These checks are expensive. Let me give you some numbers: on average, 4% of the meter readings are classified as potentially implausible by the early detection system. A mid-size energy company has around 100'000 meter readings per month. Hence, 48'000 implausible meter readings per year. A clerks needs 2 to 3 minutes to check a single reading. Roughly, this corresponds to costs of 1 to 2 € per check. Eventually, this results in 50'000 € expenses per year.
Were you able to automatize these plausibility checks?
It is still too early to come to a final conclusion. Our algorithms do these checks since 4 or 5 months. However, the results so far are overall very good. We have accuracy rates around 90%. That means our algorithms classified 90% of the meter readings correctly. This is better as the accuracy rate of
What are the future projects in this area?
Oh god, there are so many! For instance, we would like to automatize the entire process chain starting with potentially implausible meter readings. The next step is to automatize plausibility checks for invoices. Apart from that, we got an algorithm that automatically recognizes incoming mails and distribute them to the correct department. Another project is so-called chat bots. These are algorithms that automatically handle customer requests. In the best case, these chat bots are even able to chat with customers in a way that the customer does not recognize that he or she texts with a robot.
Kauffeldt, T.F., and Kazanc, H. (2019): Artificial Intelligence and Machine Learning in the Energy Industry in Realization of utility 4.0 (ed. by O. Doleski), Wiesbaden: Springer, pp. 449-463