Benchmarking real-coded genetic algorithm on noisy black-box optimization testbed

Originally, genetic algorithms were developed based on the binary representation of candidate solutions in which each conjectured solution is a fixed-length string of binary numbers; however, real-valued representation scheme is basically superior and frequently utilized in addressing hard optimization tasks, particularly for the optimization in continuous domains under a black-box scenario. In this paper, we implement a generational real-coded genetic algorithm (RCGA)--which is composed of tournament selection, arithmetical crossover, and adaptive-range mutation--with a multiple independent restarts mechanism and benchmark it on the BBOB-2010 noisy testbed. The maximum number of function evaluations for each run is set to 50000 times the search space dimension. For 40-dimensional search space, the algorithm shows promising results with 6 functions being solved up to the precision of 10^-8.