The optimization of manufacturing systems is computationally expensive in most cases. A meta-heuristic (simulated annealing) is considered here to control the overall optimization process. Stochastic Petri nets are used for the modelling and evaluation part. The basic idea is to split the optimization in two phases. In the first one a "near" optimal parameter set is quickly computed, which is improved in a second phase. This strategy has shown its ability to reduce the computational effort substantially in some cases in previous papers. Several additional heuristics are developed in this work which aim at reducing the optimization effort even further. In a first improvement, the results of the approximation phase are analyzed further to gain deeper knowledge about the optimization parameter space. This knowledge is then used to control the algorithm parameters of the second optimization phase. The solutions obtained with these new techniques are comparable to the ones obtained in the original two phase optimization work, but the computational effort is reduced by 50 percent on average. In a second approach a new optimization scheme is proposed, which can be applied to models for which the fast approximation technique used in the two-phase approach cannot be used. This scheme takes advantage of the possibility of executing parameterized simulations of the Petri net models.