Processing Symbols at Variable Speed in DUAL: Connectionist Activation as Power Supply

Abstract:

This article explores the advantages and one potential implementation of a new style of computation in which multiple lines of symbolic processing are pursued at different speeds within a hybrid multi-agent system. The cognitive architecture DUAL consists of small hybrid computational entities called DUAL agents. Each agent has a symbolic processor capable of simple symbol manipulations. There is also an activation level associated with each agent. Activation spreads according to connectionist rules. The speed of each symbolic processor is proportional to the activation level of the corresponding DUAL agent and varies dynamically. Thus multiple candidate-solutions to a given problem can be explored in parallel. More computational resources are dedicated to the more promising candidates and the degree of 'promise' is re-evaluated dynamically. This allows for flexible and efficient behavior of the system as a whole. The exact relationship between symbolic speed and connectionist activation is based on an energetic analogy. The symbolic processor is conceptualized as a machine converting connectionist activation into symbolic work. A language for implementing variable-speed symbol manipulations using delayed evaluation is introduced: S-Lisp. A small example from a DUAL-based cognitive model, illustrates variable-speed marker passing in a semantic network.

Preprint (pdf) Full text