Optimization algorithm based on densification and dynamic canonical descent (Q2488893)

The authors propose a fast and efficient derivative-free global optimization algorithm. The search domain \(X\) is assumed to be compact and convex, and the objective function \(f(x)\) to be continuous almost everywhere on \(X\). The algorithm aims at approximating the minimizer of \(f\) by finding out, at each iteration, a point along the coordinate direction which is better than the previous one, and by cycling through coordinate directions. Besides the number of variables of the original problem is reduced by the use of densification curves enabling to spare the evaluation of functional values regarding the same precision when exploring the space. A new approach for the variable reduction without any differentiability assumption is proposed which is based on a multidimensional reduction instead of a monodimensional one, i.e., the authors propose to reduce a multivariable function to a new multivariable function, but with much less variables than the original function. Finally, the performance of the method is tested on several known optimization problems and compared with its competitive methods like simulated annealing and genetic algorithms. Results of numerical experiments show a faster convergence for the algorithm proposed due to less the number of function evaluations.