Saeid Khoobdel is a Master’s student in the research group of Prof. Dr. Claus Atzenbeck at the Institute for Information Systems (iisys). There he is working on artificial intelligence (AI) for the development of knowledge bases. These make it possible to bring man and machine together. Out of personal interest, he took part in a Math Olympiad as part of an international team. The aim was not to solve mathematical problems, but to train an AI to solve problems. The team delivered impressive results.
Mr. Khoobdel, what kind of event was it exactly that you took part in?
“It was an international competition on the Kaggle platform, an online community for data scientists and AI experts. Kaggle regularly organizes competitions where teams from all over the world compete against each other to develop innovative solutions to real-world problems. The competition we took part in is called the “AI Mathematical Olympiad” and focuses on solving mathematical problems using artificial intelligence.”
What is the particular challenge of this competition?
“The challenge is not just for AI to solve mathematical problems – but to do so at a level comparable to the best human participants in the International Mathematical Olympiad. Mathematical problem solving is seen as a key milestone in the development of AI because it often requires sophisticated logical structures and high abstraction skills that are particularly difficult for machines to grasp.”
How exactly can we imagine the course of the competition?
“In the competition, 110 novel mathematical problems were presented, covering a wide range of difficulty levels – from simple arithmetic to sophisticated algebraic and geometric reasoning. As the AI models in training draw on large datasets from the internet, the challenge was to ensure that the models had not encountered the test questions in advance, which could distort the assessment of their real-world problem-solving abilities. Therefore, the tasks were specially created by an international team to ensure a fair and transparent assessment framework.”
Is it just about math or other topics too?
“Interestingly, such competitions are not just about mathematics in the traditional sense. Mathematics teaches AI models to understand logic and abstraction, which are important for many other complex problems, such as in engineering and finance. An AI model often has to abstractly “imagine” which solution makes sense over many steps. This poses a particular challenge for a machine, as it is usually based on concrete rules and has difficulty looking beyond direct data. Such competitions therefore also promote the development of models that better understand logical and abstract patterns and are therefore able to grasp complex relationships in different areas.”
What exactly motivated you to take part?
“My main motivation for taking part was the desire to learn, in line with Confucius’ motto: “I hear and forget. I see and remember. I do and understand.” I am convinced that we learn the most when we face challenges. Taking part in such demanding competitions, which reflect current challenges in the industry, encourages us to look for innovative solutions and deepens our own understanding of the field.
Another important motivation was the opportunity to work in a team and achieve more together. By working in a team, you can often break the “mental loop” by getting caught up in going it alone after a while. With such cutting-edge challenges, it is crucial to try out new methods and read lots of scientific articles to stay up to date. However, this requires a lot of time, which is often not available. In a team, however, we can pool our knowledge and time effectively, which gives us the opportunity to learn from each other and make faster progress together. One of the great advantages of being human is that by working together we can find solutions that would be difficult to achieve alone.”
What else appealed to you?
“At the end of such competitions, forums are often created in which participants can share their solutions and learn from each other’s strategies. It’s inspiring to see the different approaches that have been taken and the ideas that you can adopt for your own work. A good solution can even attract the attention of companies and research institutions that are interested in working together. These opportunities to make new contacts and possibly find other exciting projects also motivate me a lot.”
You were part of an international team. How did the group come about?
“Our team came about through shared interests and previous projects we had worked on together. Liam and I had met on an open source project and realized that we both shared a strong passion for deep learning and mathematics. We are both members of an international Discord community that brings together tech enthusiasts from all over the world. In this group, we support each other with programming problems and exchange ideas – but our conversations are not just about technology. Sometimes we also play chess together, talk about books we’ve read, movies we’ve seen or discuss everyday topics.
One day, Liam wrote to the group and asked if anyone was interested in taking part in the AI Mathematical Olympiad. Noah and I signed up immediately, and so our first team of three was formed. Shortly afterwards, we were joined by Rho and Lara, who were active in another group from Vancouver, so our team eventually consisted of five people from three different countries. Liam and Noah live in Sweden, Rho and Lara in Canada, and myself in Germany. This international composition brought a variety of perspectives and ideas to the team, which made our collaboration particularly enriching and exciting.”
What problem did you solve as a team?
Our team was faced with the challenge of solving mathematical problems that require a deep understanding of logic, abstraction and mathematical thinking. The tasks in the AI Mathematical Olympiad on Kaggle are designed to push the AI’s capabilities to their limits. It’s not just about pure calculations, but about the models recognizing complex relationships and patterns and understanding problems in a similar way to human thinkers.”
What was your approach?
Our main idea was to use an ensemble approach, which we called “The Archipelago of Intelligence” – a name inspired by the book The Gulag Archipelago by Russian Nobel Prize winner Alexander Solzhenitsyn. Without going into too much technical detail, this means that we trained several models with different architectures and parameters, with each model taking on a specific task in the solution process. Together, these models formed a pipeline in which they worked together to arrive at the optimal solution. This communication between the models allowed us not only to achieve higher accuracy than the standard models at the time, but also to understand where potential sources of error lie. This is important because the accuracy of a model that solves mathematical problems cannot be calculated simply by comparing the model response with the correct solution. It can happen that a model’s answer is close to the correct solution, but it has made fundamental errors in the solution path. Another model, on the other hand, may be further away from the correct answer but have only made a small error in an intermediate step, such as a simple multiplication. Such insights are crucial in order to improve the models in a targeted manner and achieve reliable results.
This problem is also evident in large language models, which often struggle to solve even simple tasks correctly when they require a certain level of understanding and logic. Additionally, it was a challenge to train our models with limited resources. Unlike large companies, we did not have a large number of GPUs available to train our models over several days. Therefore, our goal was to achieve high accuracy with a limited amount of data and computing power. Overall, this approach allows us to understand the mental path that a model follows and to make targeted improvements if the model makes errors or tends to be distorted in certain parts.”
How much time was required? Was the time to solve the tasks limited?
The competition on Kaggle, the AI Mathematical Olympiad, took place over a period of around three months. However, as our team was formed later, we only had around six weeks to develop and submit our project. Despite this limited time, each team member already had experience and ideas from similar projects, which enabled us to quickly find a common approach.
Our first task was to hold meetings in which each member presented their approach. This allowed us to decide together which approach was the most promising, as we did not have enough time to implement all the ideas. This decision was followed by detailed planning and task allocation to ensure that everyone knew exactly what needed to be done. As we were all studying and partly working part-time, our available time for the project was extremely limited and we had to complete most of the tasks in our free time.
This strict organization and common understanding of time constraints helped us to successfully complete the project within the given time frame. However, the work on optimizing our solution approach continues to this day. Our goal is to further improve the model so that we can publish it in the future and others can also benefit from our results.”
Was this more of a hobby activity or did it also benefit your work at iisys?
“Taking part in this competition was not just a personal project, but a valuable addition to my work in the Visual Analytics team at iisys, where I work in the field of deep learning. The experience I gained during this and other Kaggle competitions directly supports me in my professional work, especially in recognizing patterns and correlations in data. The competition furthered my understanding of machine learning and mathematical modeling and allowed me to test algorithms and strategies that are helpful in my projects – an enrichment that strengthens my professional work at iisys.”
