Prof. Emo Welzl and Prof. Bernd Gärtner
|Mittagsseminar Talk Information|
Date and Time: Thursday, September 23, 2021, 12:15 pm
Duration: 30 minutes
Location: CAB G51 and Zoom: conference room
Speaker: Johannes Lengler
Evolutionary Algorithms (EAs) are simple optimization heuristics. They maintain a population of solutions, and evolve this population over time. EAs are particularly useful when no gradients of the objective/loss/fitness function are available, and when the objective function does not have exploitable structures, like being quadratic or convex.
I will present results on the behaviour of EAs in dynamic environments. In dynamic linear functions, the objective value of a bit string x is \sum W_i x_i, where the W_i are positive weights that are i.i.d. redrawn periodically. Surprisingly, the performance of EAs can deteriorate dramatically in such situations. In particular, they may need exponential time (in the dimension) to find the optimum. Our main objective is to understand when exactly this slowdown happens. The answer is rather complex, and involves some surprising interaction between population size, mutation strength, and the speed of environmental changes. Our understanding is only partially, and I will sketch some known results.
Automatic MiSe System Software Version 1.4803M | admin login