Optimization processes are used in a wide range of areas. This thesis presents a method that allows to predict optimization results early in the optimization process by using machine learning. During optimization, a lot of data are created. Especially the costs and constraint violations of the current iteration can be utilized. Moreover, parameters like the total change of the variables in one timestep or internal parameters of the optimizer can be used. During each optimization timestep these data are created. They can be put into a machine learning model to make predictions. Here, Recurrent Neural Networks are used to deal with the time-series input. The main contribution is the creation of a new stopping criterion that can replace traditional stopping criteria to speed up the optimization process. The goal is to detect early whether the current optimization run leads to a feasible solution or not. The training and test data that are used for training and evaluation of the networks are created by an optimizer solving robotics tasks. Long-Short Term Memory Networks (LSTMs) are then trained on different features to predict feasibility of the optimization run. When the network predicts that the probability of the run being feasible is sufficiently low, it can be aborted. With the new stopping criterion, the number of steps the optimizer performs unnecessarily can be reduced by over 85% compared to the standard stopping criterion. It is almost impossible to completely avoid that a new stopping criterion also stops optimization runs that would lead to a feasible solution. For the stopping criterion that is proposed here, the amount of runs that were aborted erroneously is below 5%. This stopping criterion is actually integrated in an optimizer as an additional stopping criterion. The time the optimizer needs to perform a set of tasks can be reduced by about 75%.