En poursuivant votre navigation, vous acceptez l'utilisation de cookies destinés à des fins de mesure d'audience, à améliorer la performance de ce site et à vous proposer des services et contenus personnalisés. En savoir plus


Natural selection applied to machines

Three École Polytechnique students will present their research on genetic and evolutionary algorithms at the GECCO World Conference that will be held in Kyoto, Japan, in July. These algorithms have numerous applications in artificial intelligence, but their theoretical understanding is still limited.

Gautier Izacard, Raphaël Dang-Nhu and Thibault Dardinier, three École Polytechnique students, have carried out their third year Computer Science research project on genetic and evolutionary algorithms. Inspired by Darwin’s theory evolution, these algorithms imitate biological mechanisms of reproduction, mutation and genetic recombination.

Concrete applications for March exploration or Sony robot

These evolutionary algorithms have led to numerous concrete applications. “Very often, machine learning problems can easily be solved with these algorithms” explains Raphaël Dang-Nhu, an Ingénieur Polytechnicien student. They can be very useful in many contexts: for example, to improve the Pathfinder robot’s movements on Mars or to learn how to walk to Aibo, Sony’s robot. Other useful use: the algorithms enable to develop artificial intelligence in the video game industry.

A theoretical project to better understand their efficacy

Nevertheless, if the use of theses algorithms is widespread, the theoretical approach of this method is still limited. According to Raphaël, “It’s very important to deepen results that exist regarding the evolutionary algorithms, in order to obtain more guarantees of their efficacy. The aim of our project is to study how these algorithms react when problems they have to optimize are based on imperfect data, or when the problem changes over time”. To this end, students developed a general approach to study this kind of problems and identify the average time needed for a genetic algorithm to find the best solution.

Their work is supervised by Benjamin Doerr, researcher at the École Polytechnique Computer Science Laboratory and will be presented at the GECCO, the Genetic and Evolutionary Computation Conference in Kyoto from July 15th to July 19th, 2018. Thibault and Raphaël will attend the conference and hope that their research will enable evolutionary algorithms’ users to better understand the application performance