Personalization of Hearing Aids through Bayesian Preference Elicitation
Adriana Birlutiu, P. Groot, Tjeerd M. H. Dijkstra, Tom Heskes · 2010
Hearing aid devices contain many tunable parameters. Setting parameters in order to maximize user satisfaction is a complex problem because the determinants of hearing-impaired user satisfaction are unknown, and because the evaluation of this satisfaction through listening tests is costly and unreliable. This raises two challenges. Firstly, how should we fit the hearing aid given partial information about user’s preferences? And secondly, what experimental design leads to the fastest and most accurate elicitation of user’s preferences? The goal of our work is to improve the hearing aid fitting procedure. The approach that we take is that of Bayesian incremental utility elicitation. The satisfaction of a patient with a particular setting of hearing aid parameters is captured by a utility function. A fundamental feature of the Bayesian approach is that uncertainties about the utility function are explicitly modeled through a probability distribution. By means of listening experiments we update this probability distribution using Bayes’ rule. Patient specific information, like the audiogram (threshold of hearing for pure tones), lifestyle parameters, can be incorporated in the prior ([1]). Given a probability distribution over the utility function, one approach to find the optimal parameters is to optimize them based on the expected expected utility: it contains two expectations, taking into account a library of sound samples and the uncertainty about the utility function. We use the Bayesian experimental design for choosing the next experiment which maximizes the expected gain in utility ([2]). Fig.1 illustrates the Bayesian updating of the utility model in a paired-comparison experimental design.