Parallel evolutionary algorithm in high-dimensional optimization problem

An implementation of the combined evolutionary algorithm for searching extremum of functions with many parameters is proposed. The algorithm designed to optimize parameters of the molecular-dynamics reactive force field potential ReaxFF also can be efficient in many other extrema-searching problems with arbitrary complex objective function. The algorithm itself is a hybrid of two evolutionary methods: Genetic Algorithm which uses the principle of natural selection in the population of individuals; and Particle Swarm Optimization which imitates the self-organization of the particle swarm. Individuals in population as far as swarm’s particles are treated can be considered as trial solution’s vectors. Combination of these two methods provides an opportunity to work with objective functions with unknown complex structure which often has a composition of specific peculiarities insurmountable by simple algorithms. Genetic Algorithm parameterization regarding choosing its main strategies for computations with different objective functions has been analyzed. Results for classical test functions convergence speed testing presented. Effectiveness of the algorithm working at the computational system with shared memory and at analogous distributed system has been compared and good scalability of implemented algorithm at distributed computational system demonstrated.