This thesis presents an approach for optimizing in-car audio equalizer settings that prioritizes user preferences. Traditional equalizer interfaces overwhelm users with technical parameters, hindering intuitive sound customization. This work proposes a preference learning system that leverages iterative user feedback through pairwise comparisons of equalized audio samples. The core challenge addressed is the optimization of a function where user evaluation is indirect. Instead of directly evaluating an objective function, users express preferences by comparing different configurations. This scenario is particularly relevant in audio equalization, where users provide feedback on which of two equalized audio samples sounds better. The goal is to identify the most preferred (optimal) configuration based on these subjective evaluations while minimizing the number of comparisons required. The proposed approach employs an iterative process actively querying users for preference comparisons between equalizer settings. Inspired by active learning principles, an acquisition function balances exploitation and exploration. A radial basis function (RBF) acts as a surrogate model, predicting user preferences for explored configurations (exploitation). An inverse distance weighting (IDW) function encourages exploration of less-sampled regions within the decision space. This acquisition function guides the selection of new equalizer settings for comparison with the current best option. The thesis demonstrates the effectiveness of this approach for audio equalizer optimization, achieving near-optimal results with a limited number of user interactions. Furthermore, the robustness and generalizability of the algorithm are explored beyond the audio equalizer case study. Global optimization problems are utilized to analyze the algorithm’s performance under various conditions, including sensitivity to user feedback noise, behavior with increasing dimensionality, dependence on the chosen surrogate model, and its ability to adapt (self-calibrate) to a hyperparameter during the optimization process. This comprehensive evaluation establishes the efficacy of the proposed method for solving optimization problems where user feedback is limited to preference comparisons.