The increasing complexity of automated driving functions (ADFs) necessitates efficient data collection strategies due to the significant resources required for real-world data acquisition. Traditional methods rely on real-world vehicles, which implies costly and time-consuming collection. On the other hand, modern synthetic data generation methods do not explicitly account for the variation in complexity between the scenario classes and the data balance in relation to the training performance of the model. This paper presents a novel approach for generating synthetic automotive driving scenario datasets using a Reinforcement Learning (RL) -based framework. The system is designed to optimize data distribution dynamically at runtime, thus enhancing the stability of the training procedure. The proposed architecture iteratively guides the scenario generation process by identifying the most difficult scenarios in terms of model training F1 score performance. Preliminary experimental results demonstrate that this approach effectively stabilizes the training procedure for different scenario classes by introducing a controlled data imbalance. The study provides promising insights into the potential of RL-driven data generation for learning modules, also highlighting areas for future research to further enhance the capabilities of the system.