FI MU student attends a prestigious AI conference in Canada
His interest in machine learning and probability systems led him to science during his undergraduate studies. He is part of a team that has created a new algorithm for optimizing decision-making processes and recently presented this result personally at the prestigious 38th Annual AAAI Conference on Artificial Intelligence in Vancouver. In the interview, Vojtěch Kůr gave us details of his research, his experiences in Canada and the importance of scientific work for his future career.
How did you get into research?
I started working on my bachelor thesis under the supervision of Assoc. Vojtěch Řehak from the Department of Programming Theory FI MU. He invited me to collaborate further in the area of the decision-making processes. First I tested an algorithm written by my colleague David Klaska and played with different inputs to compare the results and behaviour of the algorithm. My role was threefold - I implemented the calculations and algorithm into the wider infrastructure, looked for the best model setup, and ran and evaluated the experiments.
Could you explain in simple terms the main problem your team is addressing?
The core of our work is an optimization problem, which is to find a strategy with a desired stationary distribution, i.e., that has the right ratio of visits to states on average. Think of it as a simple system of a factory, where we would like to spend on average 90% of the time on production and 10% of the time on maintenance. We would then look for a schedule that satisfies this. The problem is that even if the factory meets its long-term goal, there may be times when it is shut down too often or not enough in a short period. This is the problem of local instability.
What are the main findings of your research?
We have been working with modelling in the context of what are called Markov Decision Processes (MDPs), which are flow charts that help in deciding what actions to take in different situations to achieve the best overall outcome.
We have formally defined the notion of local stability with respect to steady-state frequencies. We then showed that there is no reasonably fast algorithm for finding the best strategy. So we proposed our own to find such strategies.
How reliable is your algorithm?
Our algorithm is fast, although it does not always find the best solution. But we have shown that it can often find very good solutions, even for large systems. An algorithm that would solve every problem accurately would be too slow. So instead we use easily computable values to help us estimate the stability of the system, and then optimize these values using mathematical methods, specifically gradient descent using automatic differentiation.
How can we use your results in real life?
By streamlining decision-making strategies, we can make systems more reliable and predictable. This can be applied in many areas such as manufacturing, logistics, and even technology where consistent performance is critical.
Specifically, let's imagine that a piece of equipment is expected to fail on average no more than once a month. But if, after a long period of error-free operation, it fails twice in two weeks, the user will not be satisfied. Although the average failure frequency is fine, this may in practice lead to periods of more frequent failures, which we do not want. Therefore, it is appropriate to look for optimization tools.
You presented the results at the prestigious AAAI-24 conference in Vancouver. What was the experience like for you?
I went to the conference alone from the whole university. I spent a week there - the first three days consisted of smaller workshops, and then the next four days the conference itself. It also included the so-called Poster Sessions. There were posters of various papers displayed in the large hall, and there was always at least one person, usually one of the authors. Participants could then browse among them and discuss the articles that caught their attention based on the posters. We also had a place in one such block. So there I introduced the interested participants to the details of our research and answered questions. One master's student from Aachen was so enthusiastic about our research that we discussed various possibilities for improvements and modifications together. We were very pleased with this interaction.
Photo: Poster Session at the AAAI-24 conference in Vancouver
In order to qualify for a student scholarship, I had to volunteer one day at the conference. So I handed out giveaways like pens and notepads. But during this time I was able to meet other students. Finally, I even participated in a quiz competition organized at AAAI-24 in a team whose core consisted of 4 students from King's College, but we also had students from Lyon, Athens and the USA. By attending seminars and lectures I had the opportunity to learn about the latest research in AI, of course a lot of space was given to the now very popular large language models and their capabilities and limitations. I am very grateful for these opportunities.
How did working on this project influence your future direction?
Working on the project influenced me a lot. It has been interesting to see how the research comes together, how different colleagues contribute to the work, how they complement each other and also how they resolve disagreements. It was really a team effort that was successful in the end. I also picked up a lot of practical knowledge and skills, such as working with the PyTorch library for automatic differentiation. I also got to improve my knowledge of probability theory and MDP theory, which I didn't even know until then.
What advice would you give to computer science students who want to pursue research?
I don't know how competent I am to give advice, but I think it's good to ask around the various labs and research groups in the department, and show an active interest in participating in research. I would say that people are rather scarce and therefore any extra hand and mind is useful. Also, students can sign up for lab seminars. For example, Formela (the lab I'm in) offers different seminars each semester that discuss topics that are either not otherwise taught, or are discussed in a lot of depth in the seminar. In this way, for example, I was able to get a good grasp of the basics of probability theory, which is crucial for this research.
What do you plan to study next?
First of all, I plan to finish my Bachelor's degree and eventually my Master's degree. Then I will see, fortunately that is still a long way to go. During my studies, I would like to continue to participate in some research in the field of artificial intelligence.
Anything else you'd like to mention?
I would like to thank my colleagues. It was mainly thanks to them that the paper whose conclusions I presented at the conference was written. Thank you.
Thank you for the interview.
Prof. RNDr. Antonín Kučera, Ph.D. the Head of the Laboratory of Formal Methods, Logic and Algorithms (Formela) adds:
"The results that Vojta reported at AAAI 2024 were finalized during the hot summer of 2023, when most of the students were enjoying their holidays and colleagues were taking well-deserved vacations. Instead, Vojta was busy with his scientific work. It was amazing to see how skillful and efficient he was in solving the problems we faced in the final stages of the research. Although the chances of success were not great (the percentage of accepted papers at the AAAI conference is below 20% and the competition is fierce), his commitment was enormous. I sincerely hope that other bright students of our faculty will be inspired by this example. They are welcome; and I think not only in our lab..."
The full paper can be found in the conference proceedings: https://doi.org/10.1609/aaai.v38i18.29993.
Author: Marta Vrlová, Office of External Relations and Partnerships, FI MU
Photo: Vojtěch Kůr
- Attachments
