Randomized search heuristics like local search, simulated annealing or all kinds of evolutionary algorithms have many applications. However, for most problems the best worst-case expected run times are achieved by more problem-specific algorithms. This raises the question about the limits of general randomized search heuristics.

Here a framework called black-box optimization is developed. The essential issue is that the problem but not the problem instance is known to the algorithm which can collect information about the instance only by asking for the value of points in the search space. All known randomized search heuristics fit into this scenario. Lower bounds on the black-box complexity of problems are derived without complexity theoretical assumptions and are compared to upper bounds in this scenario.