Bad codings and the utility of well-designed genetic algorithms
Franz Rothlauf, David E. Goldberg, Armin Heinzl · 2000
This paper compares the performance of the Bayesian optimisation algorithm (BOA) to traditional Genetic Algorithms (GAs) such as the simple GA, or an (μ + λ) Evolution Strategy for a real-world telecommunication problem. Users often notice that GAs perform well for real-world problems, but when the problem is slightly scaled up or modified, they sometimes fail unexpectedly. Competent GAs, such as BOA, however promise to overcome this problem more efficiently, and to behave more robustly on demanding problems. In this practical case study we use the pruefernumber encoding as an example of a bad encoding, that causes GAs difficulty in finding a good solution. The results of the experiments show that traditional GAs sometimes succeed and sometimes fail for different parameter settings or modifications of the encoding. The behaviour could not be predicted. The BOA however is able to perform as well or better than the best traditional GA, and more importantly does not fail once in this case study. It seems that the BOA is a step along the long road towards more robust and competent GAs, that are easier to use by real practitioners on problems with unknown complexity.