Finding the optimum of a black-box function is a challenging task. Optimizing these functions is often time consuming and computationally expensive. In such situations, surrogate models are often used. Surrogate models are mathematical models that approximate the behavior of a black-box function using sample data. There are several works that present methods to optimize these surrogate models. However, the influence of the different surrogate models and their properties on the solution of the underlying optimization task remains unclear. In our work, we focus on a learn-to-optimize approach, where an optimization algorithm is trained using reinforcement learning. We study different surrogate models and analyze their properties. As surrogate models, we use well-known machine learning algorithms such as k-nearest neighbors, decision trees, and deep neural networks. We optimize these models to fit our dataset and compare them using different metrics. We evaluate the models’ ability to predict the maximum of our function using Monte Carlo sampling. We then analyze the influence of their different properties in the learn-to-optimize process. Finally, we identify the best model for our task and propose ideas for further improvement based on our observations. All this is done using real data from performance tuning of semiconductor circuits in post-silicon validation. Performance tuning is a complex optimization task to determine the optimal configuration of various circuit parameters to achieve maximum performance. Semiconductor circuits are used in a wide range of applications, including computing and power electronics. Therefore, their optimization is of general interest, which underlines the importance of the topic.