Menu Close

SibIASA team members are winners of the mega-grant competition

A team of Krasnoyarsk scientists have become the winners of the ninth mega-grant competition of the Ministry of Science and Higher Education of the Russian Federation. They will develop and study hybrid modelling and optimization methods in complex systems and create a laboratory at the Siberian Federal University (SFU) under the guidance of the world-class leading scientist Predrag Stanimirovic (Serbia).

Questions are answered by SibIASA employees, key members of the mega-laboratory team.

What are hybrid modelling and optimization methods?

To simplify, analytical work uses both rigorous mathematical methods and heuristic methods, meaning methods based on the intuition and experience of specialists, but not sufficiently strict from the point of view of mathematics. In fact, rigorous mathematical methods later become cluttered with all sorts of heuristics, which makes them more efficient, and heuristic methods eventually get a more rigorous mathematical justification than they had at the beginning. This is the path to symbiosis.

Imagine that you need to get from Krasnoyarsk to a hunter’s lodge lost in the taiga in a mountainous area. Suppose you have two options: a snowmobile and a helicopter. A snowmobile can get quite close to the lodge, so you will only have to ski through the taiga for five kilometres. That is, it is a fairly accurate tool, easy to use if the conditions for its use coincide with the real ones. But you will ride it for two days and ten hours. And success is highly dependent on the terrain, which is not very accurately known to you. And it is by no means a fact that you can even get there. The helicopter will fly there in six hours, but the nearest landing spot may be 50km from the lodge, and this is not known in advance. Flying a helicopter is difficult, and there’s no guarantee that you’ll be able to. Which option should you choose? A hybrid one! You take a snowmobile and skis with you on a helicopter and, of course, invite a pilot. In six hours, you are near the lodge, then you drive the snowmobile for a couple of hours and get on your skis. After nine hours you reach the goal. So, the task of optimization (minimizing your distance to the lodge) is solved using the (simplest) method of hybridizing an accurate and easy method with a powerful and reliable one. To everyone’s pleasure. Cooperation makes what was complex a lot easier. Especially when the cooperators are optimizers.

A more complex and less obvious example is to obtain an emergent effect, meaning new properties of the hybrid that were not present in its constituent algorithms while they were being applied independently.

Why is it necessary to move towards the automated solving of modelling and optimization problems in complex systems?

For the same reasons as any automation, namely to transfer formalized, routine, well-studied processes and procedures for making decisions and their execution to a machine in order to free up a person for creative work impossible for machines and computers. For example, driving was once a highly skilled profession that required serious training, special skill and work experience among other things. Not everyone could become a successful driver. For example, female drivers were very rare and deserved great respect. Then automated mechanisms appeared on cars, for example, power steering, automatic transmission, anti-lock braking systems, hill start assist, forward collision avoidance systems and parking assistance systems. Now almost everyone can drive a car and can do so relatively safely. So, in the field of intelligent data analysis systems and decision support, one of the most important and rapidly developing areas is so-called AutoML, or machine learning automation. Machine learning itself automates decision making, so here you need to understand that we are talking about automating the development of machine learning technologies. Our hybrid algorithms are self-learning while solving the problem, that is, it turns out that we are developing methods for automating the design of algorithms that will automatically generate technologies to automatically solve modelling and decision-making problems. So to speak, ‘I’m teaching teachers to teach.’ Teachers are needed. They are the most important people in our lives. But teachers who teach students to be teachers are also very much needed. We, in such terms, turn out to be those who teach university teachers to teach schoolteachers, who will then teach pupils.

However, we should not understand the automation of modelling and optimization method design as simply translating known mathematical operations into computer language, or in other words, programming. Unfortunately, when designing such methods, the developer must work with complex objects, large dimensions and huge amounts of data. Modern requirements for such design tasks greatly exceed the intellectual capabilities of a person and become more and more difficult every day, so our ultimate goal is to formulate to the machine (algorithms implemented on a computer) only a high-level description of the desired end result, and the machine will search for a way to achieve this result on its own, without or with minimal human involvement.

How will hybridization avoid the need to involve highly-qualified specialists in the field of modelling and optimization?

The need to involve such specialists is not going away. But they will have a tool, the use of which greatly facilitates their work and allows them to concentrate on the most intellectual and demanding tasks that only people can solve. For example, to answer the question of whether it is necessary to use intelligent computer systems for automated decision-making support in a given situation. Maybe, the so-called ‘developers of artificial intelligence systems’ will finally have the opportunity to think about whether these systems should be developed and implemented at all in a particular situation. Is it fair, is it humane, do these developers even understand what their system actually does and how it makes a decision? For example, the well-known case when a self-learning chatbot in the shortest possible time learned to communicate with people in such a way that they could not distinguish it from other people, a brilliant achievement of artificial intelligence developers! Well, at the same time the chatbot became a sexist, a racist and learned other nasty things. Or the Pentagon’s development of a tank identification system which recognized all American tanks as enemy ones only because they were illuminated by the sun, as in all the photographs presented for training. The system had learnt that if a tank is brightly lit by the sun, then it is an enemy tank. As it was in the training photos. Such cases have finally led developers to an idea that is self-evident from our point of view, namely that it is important from whom to learn and what to learn. For ‘artificial intelligences’ learning at an astronomical rate, this is much more important than even for humans. And yes, hybridization per se has nothing to do with it. It’s just one of the approaches that allows us to improve the quality and efficiency of automation modelling and optimization in complex systems.

How will cooperation with the school of a leading foreign scientist allow you to bring research in the field of machine learning and artificial intelligence to a new level?

Cooperation is always a good idea and a useful activity, especially cooperation with a world-class scientist and his colleagues whose knowledge and experience complement and expand the knowledge and experience of Krasnoyarsk scientists. Cooperation with the research group of Professor Stanimirovic will make it possible to develop methods for modelling and optimizing complex systems that effectively use both the advantages of heuristic methods for solving global optimization problems (applicability to a wide class of problems with a diverse objective function landscape, including in a complex high-dimensional space) and the ability of classical optimization methods to provide fast and guaranteed convergence for a wide class of problems. This will lead to the development of general approaches to hybridization, methods of self-configuring and self-tuning of such algorithms, and hyper-heuristic hybrid algorithms. This will require that representatives of the Krasnoyarsk Scientific School master the methods of work of the leading scientist, what is possible only if Krasnoyarsk researchers have direct access to developments (for example, to a program code) and other elements of the scientific tools of a leading mathematician. Great opportunities to improve approaches will be provided by the organization of direct scientific exchange at the level of ideas, rather than completed results, and this will be ensured by the face-to-face presence of a leading scientist in Krasnoyarsk. Thanks to such cooperation, the created laboratory can claim the role of the leading Russian research centre in the field of hybrid modelling and optimization methods and one of the leading laboratories in the world.

Will this raise the status of Krasnoyarsk researchers in world science?

Yes. That’s the short answer. In more detail, the last thing we thought about when forming the research area and applying for a mega-grant was this status. That isn’t the point. It’s high enough as it is. Among Krasnoyarsk scientists who are members of the laboratory team is a winner of the Anita Borg Prize from Google, world champions in evolutionary optimization, winners of the Russian-German competition of partner scientific and educational programmes (the project ‘Youth, science, artificial intelligence: Krasnoyarsk – Ulm: 30 years together’), and team members and (co-)heads of international grants with scientists from Austria, Germany, Slovenia, Finland and Japan. And among the competitors we beat within the mega-grant competition are Skolkovo, Innopolis and “Sirius”. So, everything is good with the status. But the integration of the intellect and experience of Krasnoyarsk scientists and research groups of the invited leading scientist, of course, will bring the team to a new level of fundamental research.

It is not only and not so much the status in world science that is important, but the benefits from the planned results, including for the Krasnoyarsk Territory. In answering to the possible question of what can be the practical use of fundamental theoretical research that is not aimed at direct practical application, I will give an extended version of the well-known phrase: ‘There is nothing more practical than a good theory, and above all its correct application.’ We guarantee a good theory. And a significant part of it is aimed at facilitating the correct application, namely self-tuning, self-adaptation, hybridization and automation. Enjoy it, colleagues! If you have any questions, please contact us. We are not only scientists and researchers, but also teachers. Within the framework of the project, educational programmes of the appropriate direction (and level!) will be created.