Roman Neruda presents Evolutionary Algorithms in Machine Learning

On 2019-04-25 - 2019-04-25 16:15:00 at KN:E-301 Šrámková posluchárna, FEL ČVUT, Karlovo nám. 13, Praha 2
42. Prague Computer Science Seminar

Evolutionary algorithms represent a diverse group of optimization techniques
loosely inspired by biological evolution. Their common characteristics are a
population-based approach and a stochastic nature of optimization heuristics.
Due to their versatility, they constitute an interesting alternative to
traditional optimization algorithms, and they find their use in solutions to
complex problems, such as multi-objective optimization or automated computer
program design.

We will demonstrate several examples of how a specialized evolutionary
algorithm can search for an optimal machine learning model in various
scenarios,and how it can supersede a human expert by an efficient search
algorithm. We will focus on the areas of neuroevolution, which utilizes
evolutionary computing to train neural networks, the evolutionary reinforcement
learning of agents, and meta-learning, where evolutionary algorithms search the
space of hyper-parameters or design complex combinations of models, the
so-called workflows. We will show several original results aiming towards the
silver bullet of the meta-learning algorithms - automated design of complex
data mining systems tailored to given data.

Roman Neruda

Roman Neruda is with the Institute of Computer Science of the Czech Academy of
Sciences (ICS CAS), Department of machine learning, where he is working in the
areas of neurocomputing, evolutionary algorithms, and meta-learning. He
graduated from the Faculty of Mathematics and Physics, Charles University, and
obtained his CSc degree from the ICS CAS. In 1995-1996 he was with the Los
Alamos National Laboratory, he worked on a joint project with colleagues from
Carnegie-Mellon University, Koblenz Universitaet, University of California
Chico, University of St. Etienne, and Universidad Distrital Bogota. He is the
co-author of more than a hundred international publications. He teaches
evolutionary algorithms and multi-agent systems at the Faculty of Mathematics
and Physics, Charles University.
Responsible person: Petr Pošík