In this paper the concept of user-adaptive assistants is discussed in the context of user-centered automation. It is shown that user behavior, which is needed to control an assistant adaptively, can be identified online from observed operator activities at the interface. The applicability of neural networks for the implementation of operator models is studied. A two-lane car driving task is used as an experimental paradigm for this analysis. Various network architectures are tested. This includes a combination of functional link and backpropagation as a novel, rapidly trainable structure. It is shown experimentally, that individual human driving characteristics are indeed identifiable from the input/output relations of a trained networks. The applicability of such models to an adaptive driver assistant is demonstrated.