Accelerating Real-Valued Genetic Algorithms Using Mutation-with-Momentum

In a canonical genetic algorithm, the reproduction operators (crossover and mutation) are random in nature. The direction of the search carried out by the GA system is driven purely by the bias to fitter individuals in the selection process. Several authors have proposed the use of directed mutation operators as a means of improving the convergence speed of GAs on problems involving real-valued alleles. This paper proposes a new approach to directed mutation based on the momentum concept commonly used to accelerate the gradient descent training of neural networks. This mutation-withmomentum operator is compared against standard Gaussian mutation across a series of benchmark problems, and is shown to regularly result in rapid improvements in performance during the early generations of the GA. A hybrid system combining the momentum-based and standard mutation operators is shown to outperform either individual approach to mutation across all of the benchmarks.