Metaheuristic Start for Gradient based Optimization Algorithms

摘要

Due to the complexity of many real-world optimization problems, better optimization algorithms are always needed. Complex optimization problems that cannot be solved using classical approaches require efficient search metaheuristics to find optimal solutions. Recently, metaheuristic global optimization algorithms becomes a popular choice and more practical for solving complex and loosely defined problems, which are otherwise difficult to solve by traditional methods. This is due to their nature that implies discontinuities of the search space, non differentiability of the objective functions and initial feasible solutions. But metaheuristic global optimization algorithms are less susceptible to discontinuity and differentiability and also bad proposals of initial feasible solution do not affect the end solution. Thus, an initial feasible solution gauss for gradient based optimization algorithms can be generated with well known population based metaheuristic Genetic Algorithm. The continuous genetic algorithm will easily couple to gradient based optimization, since gradient based optimizers use continuous variables. Therefore, Instead of starting with initial guess, random starting with genetic algorithm finds the region of the optimum value, and then gradient based optimizer takes over to find the global optimum. In this paper the hybrid of metaheuristic global search, followed with gradient based optimization methods shows great improvements on optimal solution than using separately.