Dynamische Symbolsysteme (Dynamic Symbol Systems)

This thesis introduces dynamic symbol systems (DSS). The approach combines a symbolic format of information with a self-organizing dynamics. It is primarily intended for the modeling of intelligent, situated agents. The basic information processing module is a self-organizing stream. In computational terms, this is an anytime-algorithm for the processing of information that comes in a quite general stream format. In terms of thermodynamic systems, it is an open, dissipative, rapidly selforganizing system. In terms of cognitive science, a self-organizing stream is a module that can appear at any place from the peripheric sensomotoric interface to the central conceptual level, performing tasks of pattern completion, noise filtering, and gestalt formation. Different self-organizing streams can be coupled, yielding complex, self-organizing information processing systems. These associeties can span the entire periphery-centre axis of an agent. Top-down and bottom-up influences mutually support each other, with none of them having causal or temporal precedence over the other. At the most central, conceptual level, the DSS representation format is in many aspects comparable to classical semantic networks. The approach proposes answers to several controversial issues concerning the nature of concepts, e.g., context sensitivity, nonmonotonic inheritance, the nature of concepts vs. attributes, and conceptual cycles. DSS is a concrete formal instantiation of a general, structuralistic epistemology for situated information processing, namely, dynamic symbol structures. The key assumption of this perspective is to consider symbols as empirical, physical observables, which can be reliably detected in an information processing system. This epistemological frame yields a unified view on two paradigms that are often considered incompatible, i.e., the physical symbol systems paradigm and situated action. The dynamics is described in algorithmic terms. It should be relatively straightforward to make the formalism run on a computer. As yet, however, the approach is not implemented. The thesis first establishes and explores the epistemological frame. It then develops, in a rigorous fashion, the concrete DSS formalism. An application is proposed, where DSS serves as an auxiliary mechanism for memory access in an otherwise classical system. The DSS representation format for symbolic knowledge is compared with classical methods for concept representation. General insights concerning the combination of a self-organizing dynamics with symbolic representation formats are derived from a comparison of DSS with related connectionist models.