On Risky Methods for Local Selection under Noise

The choice of the selection method used in an evolutionary algorithm may have considerable impacts on the behavior of the entire algorithm. Therefore, earlier work was devoted to the characterization of selection methods by means of certain distinguishing measures that may guide the design of an evolutionary algorithm for a specific task. Here, a complementary characterization of selection methods is proposed, which is useful in the presence of noise. This characterization is derived from the interpretation of iterated selection procedures as sequential non-parametric statistical tests. From this point of view, a selection method is risky if there exists a parameterization of the noise distributions, such that the population is more often directed into the wrong than into the correct direction, i.e., if the error probability is larger than 1=2. It is shown that this characterization actually partitions the set of selection methods into two non-empty sets by presenting an element of eac...

• Type: Conference
• Year: 1998
• Where: PPSN
• Authors: Günter Rudolph
• Keywords: Characterization, Distributed And Parallel Computing, Evolutionary Algorithm, PPSN 1998, Selection Methods