In this work, a framework based on maximum likelihood estimation and mutual information is proposed to design a metaheuristic. A multilevel decomposition of metaheuristics is proposed that allow to have a unified vision on this optimization approach. Then, a new layer based on machine learning is added to take profit from the evolution of the algorithm to adapt it to the considered problem to alleviate users efforts interested in designing and implementing metaheuristics. In other terms, the goal is to proved an alternative to implement metaheuristics. Surprisingly, when results were compared to those of other classical metaheuristics, in most of cases the proposed approach provided good results.