Genetic Algorithm. The genetic algorithm that is used in this article is based on algorithms used in Xxxxxxxx et al. [29, 30] and Xxxxxxxxxx et al. [11, 7]. The genetic algorithm for all experiments is a (10, 100) strategy where 100 stands for the number of individuals in the pool and 10 for the number of parents that is selected from that pool. Parents are selected by choosing the ten fittest individuals from the population. Every parent is then copied ten times every generation (making 100 children) and all individuals are mutated using a mutation operator. Unlike previous experiments, no crossover was used in these experiments. Not only have earlier results by Xxxxxxxxxx et al. [11, 7] (see Chapter 4) shown that the impact of crossover on the Majority Problem is minimal, but investigating the effect of self-adaptation is also a lot easier without taking into account crossover. The algorithm uses a comma strategy. This is also different from previous experiments, which used a plus strategy. Preliminary experiments seem to suggest that self-adaptation of mutation rates in GA do not work very well in a plus strategy. This probably has to do with the success rate in the al- gorithm and the fact that the mutation rate only changes if the individual is mutated. In a plus (or elitist) strategy the parents are not mutated and copied to the next generation without any changes. If the success rate in the algorithm goes down, these elite individuals will drown out any form of diversity while at the same time keeping the mutation rate identical and probably too high. Using a plus strategy with self adaptation is probably possible if this effect can be countered somehow and this is worth investigat- ing in the future, but by choosing a comma strategy this problem is evaded. Every surviving individual is mutated every generation, which results in a dynamic mutation rate that evolves at the same time as the individuals object values, thus forcing the algorithm to select the best mutation rate at different stages in the optimization.
Genetic Algorithm. The second script contains two variants of a genetic algorithm [Deb12] that have two parameters to tune their behaviour. The variations o er choices of both how the initial population is generated and how parents are selected to produce each subsequent generation. The parameters a ect the size of the initial population and how long the algorithm will continue if no better results are being found. The strength of a closed loop search, such is this genetic search, is that they will perform a search of the design space with- out testing each design and so require less CPU time than an exhaustive search [FGPL17]. This strength comes at the cost of guaranteeing of nding globally optimal designs, the search may instead nd some set of local opti- mums and return those. The following subsections outline the steps take by the genetic algorithm. Initial population generation The rst step in the genetic algorithm is to generate an initial population of designs. The size of this initial set is a parameter the user may set and there is ongoing work that is described in D3.2a [FGPP16] that aims to provide guidance on what this size should be. It is on the generation of this initial set that the two genetic algorithms di er. One version of the script produces an entirely random set of designs and then proceeds to the next step, while the other version attempts to produce a set of designs that is evenly distributed across the design space. Again the experimentation in D3.2a aims to provide guidance about which of these options should be used and when.
Genetic Algorithm. The second script contains two variants of a genetic algorithm [Deb12] that have two parameters to tune their behaviour. The variations offer choices of both how the initial population is generated and how parents are selected to produce each subsequent generation. The parameters affect the size of the initial population and how long the algorithm will continue if no better results are being found. The following subsections outline the steps take by the genetic algorithm. in D3.2a [FGPP16] that aims to provide guidance on what this size should be. It is on the generation of this initial set that the two genetic algorithms differ. One version of the script produces an entirely random set of designs and then proceeds to the next step, while the other version attempts to produce a set of designs that is evenly distributed across the design space. Again the experimentation in D3.2a aims to provide guidance about which of these options should be used and when. Evaluation and ranking The second step in the genetic algorithm is to evaluate the new designs according to the objectives in the DSE config file (section 4.2.2) and then to place them in a partial order according to the ranking defined (section 4.2.4).
