EDDA-V2 – An Improvement of the Evolutionary Demes Despeciation Algorithm

For any population-based algorithm, the initialization of the population is a very important step. In Genetic Programming (GP), in particular, initialization is known to play a crucial role - traditionally, a wide variety of trees of various sizes and shapes are desirable. In this paper, we propose an advancement of a previously conceived Evolutionary Demes Despeciation Algorithm (EDDA), inspired by the biological phenomenon of demes despeciation. In the pioneer design of EDDA, the initial population is generated using the best individuals obtained from a set of independent subpopulations (demes), which are evolved for a few generations, by means of conceptually different evolutionary algorithms - some use standard syntax-based GP and others use a semantics-based GP system. The new technique we propose here (EDDA-V2), imposes more diverse evolutionary conditions - each deme evolves using a distinct random sample of training data instances and input features. Experimental results show that EDDA-V2 is a feasible initialization technique: populations converge towards solutions with comparable or even better generalization ability with respect to the ones initialized with EDDA, by using significantly reduced computational time.