Dispatch rules are commonly used to schedule lots in the semiconductor industry. Earlier studies have shown that changing dispatch rules that react to a dynamic manufacturing situation improves the overall performance. It is common to use discrete event simulation to evaluate dispatch rules under different manufacturing situations. On the other hand, machine learning method is shown to be useful in learning the relationship of a manufacturing situation and the dispatch rules to generate dispatching knowledge. In this work, we use simulation and machine learning methods to generate dispatching knowledge and define features that are relevant in a dynamic product mix situation. However, more features will increase the risk of overfitting the machine learning model. Hence, dimension reduction methods are explored to reduce overfitting and improve generalization of the model. Simulation results show that this approach can adapt the dispatch rule combination and achieve a comparable factory performance measurement.