A hearing device has a signal processor which has an adjustable parameter that has a given setting at a given time. The parameter is set depending on the situation by selecting a setting, depending on an environmental situation and by a learning machine. A current setting of the parameter can be rated by feedback from a user. In a first training procedure the learning machine is passively trained by negative feedback signals, by rating feedback from the user as dissatisfaction with the current setting and by assuming the user's satisfaction with the current setting as long as no feedback is given. In a second training procedure the learning machine is trained by changing the current setting independently of the feedback from the user and in spite of an assumed satisfaction with the current setting, so that the user is offered a different setting which can then be rated by feedback.