Skip to Main Content
In this paper we present some experiments coinparing different connectionist dynamic memory models in an interactive context, including a novel version of a second-order architecture, along with an extended discussion of some foundational issues to s'hich they relate. The scenario is a fine-grained simulation of a sensor-based, non-holonomic mobile robot moving in a two dimensional environment in which are set paradigmatic short-term memory experiments involving delayed response tasks for the robotic subject. We set out our position initially in relation to symbolic AI. connectionism and nouvelle AI. In particular. we argue that despite its practical successes, the last named approach does not escape from a foundational error that in different guises undermines the programmatic aspirations of all three. We suggest that this representational problem is inherent in the implementational granularity of systems such as the subsumption architecture and argue for the notion of embedding in time, as a first step around the impasse.