This paper considers an important, yet poorly studied, aspect of genetic algorithms (GAs)-mixing of building blocks. Analytical models of building-block exchange in a linear problem are developed and verified through computational experiments. First, a simple model of allele-wise mixing without selection is analyzed to verify the prediction of mixing time alone. Thereafter, this result along with a previously derived model of selection are used to predict a GA control map involving two important GA parameters-selection pressure and crossover rate. The paper concludes by discussing how these methods can be extended to problems of bounded difficulty.

View full abstract

This paper presents a continuous generation genetic algorithm. Most genetic algorithms use a discrete generation model in which all individuals in a population synchronize mating period. The discrete generation model, however, wastes processor time in parallel implementations when the fitness of each individual (proportionally or reversely) correlates with the computational cost of its evaluation. An example of such a task is neural network design and training. In some cases, over 80% of total CPU time has been wasted. The continuous generation model mitigates this problem by introducing asynchronous mating. the continuous generation model increases the number of reproduction per a unit-time over 500% over the traditional discrete model. CPU idle time has been minimized to 1/25. Also, a significant improvement in convergence speed has been estimated.

View full abstract

[in Japanese]

View full abstract