|Institution:||Universität Heidelberg ; Thes|
|Full text PDF:||http://archiv.ub.uni-heidelberg.de/volltextserver/1428/|
In the present thesis, a connectionist model (called EVA) was developed for the simulation of human planning behavior in the domain of Plan-A-Day. Plan-A-Day is a diagnostic instrument for the assessment of planning capabilities (Funke & Krüger, 1995), which provides a computer scenario in which the subjects must find an optimal sequence for several scheduled appointments by applying operators from a predefined set. EVA is designed to produce sequences of these operators which are applied within Plan-A-Day. In this way, EVA is intended to show the same kind of operationally defined planning behavior as human subjects do. EVA comprises three sub-networks of the backpropagation-class, two of which are hierarchical Elman networks. Based on the situated action account (Clark, 1997; Suchman, 1987), the input of EVA is for the most part a representation of the current state of the Plan-A-Day scenario. The data obtained by simulation runs is compared to empirical data obtained by a study with 45 human subjects. The comparison is carried out with regard to indicators reflecting planning performance, as well as characteristics of the process of planning. The fit between simulated and human data is good enough to allow certain theoretical claims. EVA provides evidence that simple pattern transformation devices like connectionist networks are capable of performing human-like planning, if these networks are thoroughly embedded into their environment, as it is demanded by the situated action account. EVA demonstrates that even conventional connectionist models are applicable to tasks from the field of high-level cognition like planning or problem-solving, which were mostly reserved for symbol-processing models until now.