This thesis describes the F1uid Analogies Engine (FAE), a computational model aimed at exploring the general cognitive mechanisms underlying flexible perception and intelligent deliberation. The architecture of this model was inspired by the work of Douglas Hofstadter and the Fluid Analogies Research Group who have developed a general computational framework for modelling high-level cognitive tasks including planning, creativity and analogy-making (Hofstadter & FARG, 1995). In this approach, high-level intelligent behaviour emerges from the parallel interactions of a multitude of agents that compete with and support each other in manipulating flexible, context-sensitive mental representations.
In addition to the Fluid Analogy models, there have been a range of computational frameworks proposed for modelling cognition using a complex systems approach. Such examples include LEABRA (O'Reilly and Munakata, 2000), DUAL (Kokinov, 1994b) and 3Caps (Just, Carpenter and Hemphill, 1996). As with the Fluid Analogy Models, these approaches are limited however, in that they utilise algorithms or representations that are only appropriate for modelling either low or high-level cognition. For example, connectionist models are limited in that they cannot readily model the manipulation of complex symbol structures required for planning and reasoning, and the Fluid Analogy models are limited in that they do not possess mechanisms for processing the raw distributed information provided by the senses. As is argued within this thesis, both high and low levels of processing are not only crucial for intelligent behaviour, they are heavily co-dependent and inseparable. The aim of the present work is to devise a general framework for cognitive modelling that can equally capture a wide range of both high and low levels of processing using the same set of mechanisms.
In bridging the gap between high and low level cognition, FAE uses a hybrid architecture involving a connectionist production system. All processing in FAE is achieved through a dynamically constructed neural network capable of processing distributed subsymbolic information. The network itself is formed through the actions of symbolically defined productions that afford the manipulation of complex hierarchical representations. F AE' s general architecture includes a user defined set of sensory modules (for processing raw distributed data), a working memory (for the processing of symbol structures), a semantic memory (storing important "facts" that are known to the system) and an episodic memory (that allows for self-watching during mental exploration). FAE differs from other models of high-level cognition in that it utilises representations and algorithms that are equally applicable to both high and low level perceptual tasks, creating a unifying framework for general cognitive modelling.