Modern factories, heating systems and greenhouses are increasingly being controlled by artificial intelligence. However, often no one understands exactly why the software makes certain decisions. This is problematic. Researchers at Chemnitz University of Technology have developed a solution.
Their method has just been accepted at one of the most important scientific conferences in the world: the International Conference on Machine Learning, or ICML for short. Out of almost 24,000 submissions, only 6,300 made it into the program. Chemnitz University of Technology is taking part with its idea for control technology.
The solution from Chemnitz
The team led by Streif has therefore developed a process that subsequently explains the decisions made by the control software. It uses both physical knowledge about the respective process as well as data from the operation of the plant and information about how individual decisions are related. The researchers are now providing reliable explanations for decisions that previously often seemed like a "black box". As a result, people can better understand the software's suggestions and are more likely to trust them.
In tests on a greenhouse, a residential heating system and an industrial process, the method improved the comprehensibility of AI decisions by 53 percent compared to other approaches. The project is funded by the European Social Fund Plus program with more than 2.13 million euros.
The acceptance at the ICML, says Streif, "underlines the international visibility of research on trustworthy and explainable AI for engineering applications". Seven young researchers from four faculties at Chemnitz University of Technology are working in the MORE-KIBA group. Their goal: machines should not only be smart, but also be able to explain what they are doing.
The preprint version of the publication is available here.
