You will not find a single algorithm which will solve all the objective functions with equal efficiency. An algorithm may solve a particular problem very efficiently while fetching poor results (optimum points) in another problem.
Without knowing the characteristics of your fitness function, extent of non linearity, it is impossible to comment upon the perfect algorithm.
The best you can do is to select 3-4 algorithms and test these on your objective function.
Efficiency depends on how many computations are necessary per successful iteration of the search process (fitness evaluation, whatever). For example, in tabu search a neighborhood of potential solutions is assessed prior to selecting one. If the assessment of the neighbors is very complex and computationally expensive, maybe this isn't the best search process for the problem being solved. As another example, if the GA you are using results in many, many infeasible children due to violations of the constraints of the problem as a result of the splitting and recombination of parental DNA and resulting mutation of the children, perhaps this isn't the best search process either.
Other way to evaluate the solution using an meta-heuristic model is by the analyses of confusion matrix. A confusion matrix brings information about the number of correctly classified intances in a dataset and how many was not very good classified, in general. With this the model can provide some key metrics for evaluation of the proposal solution, such as: accuracy, medium square error and other. For problems of optimization is very important to consider the optimal hyper-paramethers of the algorithms before train it.