The development of highly automated driving functions is currently one of the key drivers for the automotive industry and research. In addition to the technical constraints in the implementation of these functions, a major challenge is the verification of functional safety. Conventional approaches aiming at statistical validation in the sense of real test drives are reaching their economic limits. On the other hand, there are simulation methods that allow a lot of freedom in test case design, but whose representativeness and relevance must be proven separately. In this paper an approach is presented that allows to generate critical concrete scenarios and test cases for automated driving functions by means of a reinforcement learning based optimization using here the example of an overtaking assistant. For this purpose, a Q-Learning approach is used that automates the parameter generation for the test cases. While pure combinatorics of the variable parameters leads to an unmanageable amount of test cases, the percentage of actually relevant critical test cases is very low. In this work we show how the share of critical and thus relevant test cases can be increased significantly by using the presented method compared to a purely combinatorial parameter variation.