Prof. Emo Welzl and Prof. Bernd Gärtner
|Mittagsseminar Talk Information|
Date and Time: Thursday, April 14, 2022, 12:15 pm
Duration: 30 minutes
Location: OAT S15/S16/S17
Speaker: Marc Kaufmann
Parameter control is a classical area of research in the domain of Evolutionary Algorithms. One approach optimizes parameters before execution and leaves the found parameters static throughout the algorithm run. The optimization outcome is then highly dependent on the chosen values. Picking, for instance, the population size too small - as for the (1,λ)-EA - or the mutation rate too large - as for the (1+1)-EA - can distinguish quasilinear from exponential runtime. Another approach attempts to mitigate this by adapting parameters dynamically. For the update rules, we still need to set hyperparameters, in our setting the so-called success rate and update strength, so the question again becomes: Is this strategy more robust than the static case? And do we indeed find locally optimal parameter values? In this talk, we will discuss some results regarding the self-adjusting (1,λ)-EA. All the work presented is joint with Maxime Larcher, Johannes Lengler and Xun Zou.
Automatic MiSe System Software Version 1.4803M | admin login