Our CEO Jürgen Schmidt said that I would be the best person to write a blog on Causal AI. If a non-technical person writes it, it might read more clearly, he said. Well, I feel flattered…
Causal AI is currently being discussed intensively at conferences, new developments are emerging all the time, and it is one of the more recent buzzwords around the use of AI. Unlike AI itself, whose roots go back to the 1950s, the newer school of thought behind Causal AI is in fact a relatively young technology. The concept comes from Judea Pearl, who introduced it in 2018 in his “Book of Why”.
What distinguishes correlation from causality?
Events that occur at the same time often suggest that one causes the other. When there is a high degree of statistical agreement, we speak of correlation. However, correlation is not the same as cause.
We know many examples of random correlation that are easy to explain. The higher the temperature, the more traffic jams there are in the city. So heat causes traffic jams. Of course, we know that roadworks are set up more frequently during the summer holidays. Summer roadworks cause traffic jams, not the heat. Studies that only look for statistical correlations therefore often lead to false conclusions.
It becomes trickier when there is a causal relationship, but it is not so easy to see what is cause and what is effect. My son is a firefighter and once jokingly pointed out to me that the amount of damage is higher the more firefighters take part in an operation. So it would be better if fewer emergency responders came; then the damage would be smaller. In this case, too, it is quickly clear that more emergency responders (effect) are called to a major incident (cause) than to a traffic accident with sheet-metal damage.
Other causal relationships are harder to clarify. For example, the dependency between supply and demand. Is more being invested by competitors in the development of new products in a certain category because demand is rising? Or are they bought more readily because there is a broader offering? A classic chicken-and-egg question that is difficult to answer with conventional machine learning methods. A causal model can capture cause-and-effect relationships from large volumes of data and understand connections.
Simulation
The possible applications, however, go far beyond simply explaining „What is the reason for this?” The effects of interventions can be simulated and expensive wrong decisions avoided using Causal AI. What impact does closing a location, a significant price increase, or the market entry of a competitor have? – The failure of an important customer or supplier? In this way, stress tests, for example, can also be implemented and the resilience of companies assessed. The old human dream of being able to find out the „What if?” in a digital simulation has become reality.
How do causal AI models work?
As early as the 1930s, methods from econometrics and the social sciences were described that deal with causality. However, causal models must not be based solely on data; they must be designed in such a way that they take into account the creators’ understanding of the existing causal mechanisms.
The core is understanding and modelling cause-and-effect relationships. For illustration, a so-called directed acyclic graph (DAG) is often used. Pearl significantly advanced what are known as Bayesian networks for representing causal relationships using graphs. The arrows are drawn when there is a causal relationship between two variables, represented as nodes, even though the observed statistical correlation does not depend on the direction of the arrow. Bayesian networks are powerful and can determine the true origin of an effect even in complex problems, and propose targeted optimizations.
Judea Pearl describes three levels of knowledge in the “Book of Why”:
As mentioned at the outset, research around Causal AI has gained real momentum in recent years. Some research findings still seem rather abstract at first, while others are already suitable for implementation in real projects.
Where is Causal AI used in practice?
Many industries can benefit from the use of these technologies. From insurance (better understanding of risk factors), to energy supply (preventing power outages), to supply chain management: wherever complex systems are optimized, causal models help identify the decisive levers and act efficiently.
We would be glad to exchange views if you are considering using Causal AI to support the optimization of business and manufacturing processes or strategic decisions. People have to make the decisions; algorithms can provide the foundations for them.
And: Don’t worry, our developers and data scientists will take care of implementing the project, not me.