Today, modern digital devices can be customized significantly to the individual user by adjusting or optimizing multiple parameters affecting the output of the devices. Such personal optimization of devices is referred to as personalization. In the case of hearing aids, personalization is not only a possibility offered to the user, but a requirement that must be performed carefully and precisely in order for the user to utilize the full potential of modern multi-parameter hearing aids. Today though, personalization is still based on a manual timeconsuming trial-and-error approach performed by the user himself or, in case of hearing aids, by a hearing-care professional based on typically ambiguous oral feedback from the user. This often results in sub-optimal or even inappropriate settings of multi-parameter devices. This dissertation presents research on a machine-learning based interactive personalization system to improve the personalization of devices and, in particular, of hearing-aid devices. The proposed personalization system iteratively learns a non-parametric probabilistic model of a user’s assumed internal response function over all possible settings of a multi-parameter device based directly on sequential perceptual feedback from the user. A sequential design based on active learning is used to obtain the maximum of the user’s unknown internal response function in as few iterations as possible. Experiments were conducted where the proposed personalization system obtained a significantly preferred setting for individual users within ten to twenty iterations in scenarios with up to four parameters. Following a short introduction that includes a summary of results and contributions, the first main chapter focuses on the probabilistic modeling framework in which a Gaussian process is used to model the user’s unobserved internal response function. The first main challenge addressed in this context is to account for inconsistent and thus noisy user feedback. The second main challenge addressed is to support feedback which closely reflects the user’s perception while providing maximal information about it without imposing a high cognitive load. In the second main chapter, active learning and sequential design are discussed in relation to the challenge of obtaining the setting that maximizes the user’s unobserved internal response function in as few iterations as possible. For the Gaussian process framework, an active learning criterion is proposed specifically suitable for this type of optimization. The final chapter contains an overall discussion and conclusion of the present work and research based in part on the results from eight scientific paper contributions contained in the appendices.