Model Assisted Evolution Strategy - MAES


In the pre-selection concept lambdaPlus>lambda individuals are generated from mu parents. All lambdaPlus individuals are evaluated by a surrogate model of the fitness landscape and the estimated fitness values are used to pre-select the lambda best individuals, which will be evaluated with the real fitness function. The model is trained at the beginning with a randomly created initial population and is updated after each generation step with lambda new fitness cases. The idea behind this approach is that only the most promising individuals with a good fitness prediction are evaluated with the true fitness function. Every generation a new offspring lambda is evaluated with the real fitness function, the model is updated with this information of \lambda fitness cases.