In the paper, the authors highlight the mathematical principles behind the payroll engine VIREN. They explain how artificial intelligence optimizes the tool. The paper was awarded at the leading IAAI conference organized by the (Association for the Advancement of Artificial Intelligence).
Wannes Meert: With VIREN, Teal Partners has built a tool that converts knowledge into executable computer code. VIREN calculates the potential impact of changes in companies' wage policies. The way VIREN calculates aligns with our AI research domain. VIREN translates knowledge, in this case, tax and social-judicial regulations, into executable computer code that allows users to perform business calculations. You can also apply these formulas in other knowledge domains. We worked out an additional translation to another computer code to realize reverse calculations based on how Teal Partners had already translated knowledge into efficient code for calculations from data to results. Theoretically-mathematically, this is potentially an unsolvable task, but it worked out fine in this case because we were able to turn an abstract-mathematical problem into a concrete case.
Wannes Meert: Suppose you add one and one. Calculated in the mathematical, logical direction, you come up with two. But if you start from a formula with unknown factors and your result is ten, you have to calculate in reverse. Constraint programming allows you to calculate in two directions. This is a lot more difficult and less efficient, even for the computer, but today's technology makes it possible. In the VIREN case, it was even easier than expected.
Wannes Meert: The unknown factors in our case are a person's age and seniority, whether he has a cell phone or company car, and so on. Numerous factors can influence a tax result. Those factors you are going to enter as constraints, literally translated limitations. Suppose you want to know how many days you have to work at least to keep a net wage of 1000 euros. That 1000 euros, that's your outcome. You want to know the impact of the choice of a cell phone or a company car. This is not easy to calculate because you have to play with your variables, with all the elements influencing the tax rate. All the choice options together give you an exponential number of possibilities. And, since corona, everyone understands what an exponential curve is (laughs). You can see how it quickly becomes unrealistic to do those calculations by heart to try all the possibilities. By constraint programming, you reduce the playing field enormously. The software knows: I'm not going to look at certain options anymore because I know they're never going to be better than the ones I already calculated. That is automatic reasoning, and the computer can help with that.
Wannes Meert: Most people think of AI exclusively in the subdomain of machine learning. The computer learns to recognize pictures of dogs or traffic lights after seeing many images of them. But AI is much broader. In constraint programming, we don't write down the computer code. We write down the goals of the computer program. We don't explicitly program but describe the environment, such as the tax code, and an objective, for example, "maximize profit". The computer itself finds the solution by automatic reasoning, which is precisely a core element of AI. In VIREN, we also use machine learning. Since VIREN often performs similar tasks, we can learn from previous calculations and perform new calculations faster. So VIREN uses a combination of reasoning and learning.
Wannes Meert: This year, the conference took place online. Therefore there was less interaction, but nevertheless, we received many positive reactions. Especially from sectors that rely heavily on legislation, such as insurers and mutualities, there was interest in building similar tools. It triggers companies to try out this form of AI in their domain.
Wannes Meert: The VIREN tool is already being used in the field, but our parallel calculation method allowed us to add extra functionality, especially optimizations. Our collaboration also underlines the innovative nature of the VIREN solution. For us, it is fascinating to be able to test our research against the real world. Teal Partners employs strong programmers. We have worked together intensively, brainstormed a lot, and experimented. This is a complex subject matter, and they managed to absorb the information in no time. They also realized the conversion of our prototype to the VIREN environment in less than a week. That is extremely fast.
Wannes Meert: In the past year, our research group has worked with about twenty companies. My job is to stimulate the link between the university and companies and detect how our technology can be helpful. Fundamental research is fascinating, but you also want to contribute something to the world as a scientist. Realizing something that is also useful outside the university, building things that work. After all, we remain a sub-part of the engineering faculty (laughs). We always work in the long term. What we come up with today doesn't go into production tomorrow. For example, we had been developing the technology we used in the project with Teal Partners for years. It was the first time we did a project in the context of legislation and compliance, though.
Wannes Meert: The concept of turning rules into constraints and improving through machine learning has a wide range of applications. For example, in the construction sector. The software suggests designs of technical drawings based on the specs provided by a customer. Suppose he wants to introduce a new component: what are the feasible options or the influence on the price? A building materials company can enter constraints such as pressure or temperature to make the proper selection. In the logistics industry, constraint programming is already more familiar. The factors could be the number of packages, the available trucks, the available drivers, the routes. You ask the computer to give you an optimal distribution, and it gives you the numbers and routes.
Wannes Meert: The VIREN project is finished, but if we find a new research project, we'd love to. If we find a topic relevant for both of us, in terms of research and in practice, we'll start right away.
The paper that came about thanks to the collaboration between Teal Partners and the AI department at KU Leuven describes the use of AI to calculate the potential impact of changes in wage policy. Besides Wannes Meert, Sebastijan Dumancic from KU Leuven and Teal Partners Stijn Goethals, Tim Stuyckens, Jelle Huygen and Koen Denies collaborated on the paper.