Empirical analysis of diversity-preserving mechanisms on example landscapes for multimodal optimisation

Many diversity-preserving mechanisms have been developed to reduce the risk of premature convergence in evolutionary algorithms and it is not clear which mechanism is best. Most multimodal optimisation problems studied empirically are restricted to real-parameter problems and are not accessible to theoretical analysis, while theoreticians analyse the simple bimodal function TwoMax. This paper looks to narrow the gap between both approaches. We perform an extensive empirical study involving 9 common diversity mechanisms on Jansen-Zarges multimodal function classes (Jansen and Zarges, PPSN 2016) that allow to control important problem features while still being amenable to theoretical analysis. This allows us to study functions with various degrees of multimodality and to explain the results in the light of previous theoretical works. We show which mechanisms are able to find and maintain a large number of distant optima, escape from local optima, and which fail to locate even a single peak.