Rank based variation operators for genetic algorithms

We show how and why using genetic operators that are applied with probabilities that depend on the fitness rank of a genotype or phenotype offers a robust alternative to the Simple GA and avoids some questions of parameter tuning without having to introduce an explicit encoded self-adaptation mechanism. We motivate the algorithm by appealing to previous theoretic analysis that show how different landscapes and population states require different mutation rates to dynamically optimize the balance between exploration and exploitation. We test the algorithm on a range of model landscapes where we can see under what circumstances this Rank GA is likely to outperform the Simple GA and how it outperforms standard heuristics such as 1/N. We try to explain the reasons behind this behaviour. ACM Primary Classification: I.2.8 Problem Solving, Control Methods and Search Subjects: Heuristic methods. ACM Additional Classification: J.2 Subjects: Engineering. J.3 Subjects: Biology and Genetics...

• Type: Conference
• Year: 2008
• Where: GECCO
• Authors: Jorge Cervantes, Christopher R. Stephens
• Keywords: Different Mutation Rates, Explicit Encoded Self-adaptation, GECCO 2008, Optimization, Simple Ga