Late Bloomers, First Glances, Second Chances: Exploration of the Mechanisms Behind Fitness Diversity

Fitness diversity is an idea in the field of evolutionary algorithms, which calls for supporting the evolution of solutions at all fitness levels simultaneously. In some cases, this idea may even extend to cultivating the worst solutions. While this may seem counterintuitive, fitness diversity has shown its promise in algorithms such as Hierarchical Fair Competition and Convection Selection. Although these algorithms share many similarities, the role fitness diversity serves in each of them is different. In Hierarchical Fair Competition, fitness diversity facilitates a constant incorporation of novel genotypes into the solutions that are already good - a mechanism we dub First Glances - and discovery of solutions through the exploration of neutral networks of different fitness levels - which we name Late Bloomers. On the other hand, Convection Selection uses fitness diversity techniques to give broken solutions time and shelter necessary to cross larger valleys in the fitness landscape - a mechanism we call Second Chances. In this work, we compare these two algorithms and their respective mechanisms over a range of numerical and 3D structure design optimization problems. We analyze the extent to which their mechanisms are utilized, and measure the impact of these mechanisms on finding good solutions.