Pyramid-Z: Evolving Hierarchical Specialists in Genetic Algorithms

Pyramid is a hierarchical approach to Evolutionary Computation that decomposes problems by first tackling simpler versions of them before scaling up to increasingly more difficult versions with smaller populations. Previous work showed that Pyramid was mostly as good or better than a standard GA approach, but that it did so with a fraction of individuals processed. Pyramid requires two key parameters to manage the problem complexity; (i) a threshold a as the performance bar, and (ii) b as the container with the maximum number of individuals to survive to the next level down. Pyramid-Z addressed the shortcomings of Pyramid by automating the choice of a (to assure that the top individuals are highly significantly better from the original population at the current level) and makes b less aggressive (to maintain a moderately sized population at the final level). In cases where evolution starts to stagnate at the final level, the population enters into a different form of evolution, driven by a form of hyper-mutation that runs until either a satisfactory fitness has been found or the total evaluation budget has been exhausted. The experimental results show that Pyramid-Z consistently outperforms the previous version and the baseline too.