October 13th

14:30 , R2014 Digiteo Shannon (660) (see location):

Marcus Gallagher

Title : Engineering Features for the Analysis and Comparison Black-box Optimization Problems and Algorithms

Abstract :

Given two black-box optimization algorithms (e.g. A and , it is of interest to quantify what kinds of problems A outperforms B on and vice-versa. However, unless the mechanisms of A and B are very well understood and/or strong assumptions made about the structure of the problems, this is a very difficult task. One possible approach is to measure features of optimization problems which capture and measure salient properties of the problems, such that the features can be used to distinguish or categorize problem instances. Subsequently, algorithm performance might be predictable based on such problem features. In this talk I will discuss a recently proposed feature called "length scale". Length scale is based on data available in the black-box setting and makes few, if any, apriori assumptions on the nature of the problem or algorithms to be analysed. Experimental results will be presented showing the potential utility of the analysis across continuous and discrete problems.