Models to classify the difficulty of genetic algorithms to solve continuous optimization problems

What constitutes a hard optimization problem to an Evolutionary Algorithm (EA)? To answer the question, the study of Fitness Landscape (FL) has emerged as one of the most successful techniques. FL measures the landscape depicted by the problem’s cost function. Fitness Landscape Analysis (FLA) uses a set of metrics to try to determine the hardness of problems; FL metrics can be divided in descriptive and dynamic. Descriptive metrics measure the intrinsic problem features, examples of these measures are ruggedness, neutrality, basins of attraction, and epistasis. Dynamic metrics measure the evolvability of EA, examples of these measures are fitness distance correlation, and negative slope coefficient. This contribution presents a procedure called Performance Classification Models (PCM) which creates learnings models to predict the performance exhibited by Genetic Algorithms (GA) in the solution of optimization problems in the continuous domain. PCM classifies the performance in two classes (easy or difficult). The dataset has as predictor variables, FL features, and as target variable the performance exhibited by the GA. The problems used in experiments are benchmark optimization functions. A product of this approach, is a procedure to Recommend Population Size (RPS): given an optimization problem, RPS recommends the minimal population size to get an efficient level of performance. This work can be easily extended to use other metrics, or a different set of problems, or the use of other EA. Developing performance models for other EA, we can solve an instance of the algorithm selection problem.