On the Use of Self-Island-based Evolutionary Computation Methods on Complex Environments

Rafael Nogueras, Carlos Cotta

We consider the use of island-based evolutionary algorithms (EAs) on fault-prone computational settings. More precisely, we consider scenarios plagued with correlated node failures. To this end, we use the sandpile model in order to induce such complex, correlated failures in the system. Several EA variants featuring self-adaptive capabilities aimed to alleviate the impact of node failures are considered, and their performance is studied in both correlated and non-correlated scenarios for increasingly large volatility rates. Simple island-based EAs are shown to have a significant performance degradation in the correlated scenario with respect to its uncorrelated counterpart. Resilience is however much improved via the use of self-* properties (self-scaling and self-healing), which leads to a more gentle degradation profile. The inclusion of self-generation also contributes to boost performance, leading to negligible degradation in the scenarios considered.