Extending Cognitive Architectures with Spatial and Visual Imagery Mechanisms

This research presents a computational synthesis of cognition with spatial and visual imagery processing by extending a symbolic cognitive architecture (Soar) with mechanisms to support reasoning with quantitative spatial and visual depictive representations. Inspired by psychological and neurological evidence of mental imagery, our primary goals are to achieve new functional capability and computational efficiency in a task-independent manner. We describe how our theory and the corresponding architecture derive from behavioral, biological, functional, and computational constraints and demonstrate results from three different domains. Our evaluation reveals that in tasks where reasoning includes many spatial or visual properties, the combination of amodal and perceptual representations provides an agent with additional functional capability and improves its problem-solving quality. We also show that specialized processing units specific to a perceptual representation but independent of task knowledge are likely to be necessary in order to realize computational efficiency in a general manner. The research is significant because past research in cognitive architectures primarily views amodal, symbolic representations as being sufficient for knowledge representation and thought. We expand those ideas with the notion that perceptual-based representations participate directly in the thinking rather than serving simply as a source of sensory information. The new capabilities of the resulting architecture, which includes Soar and its Spatial-Visual Imagery (SVI) component, emerge from its ability to amalgamate symbolic and perceptual representations and use them to inform reasoning. Soar’s symbolic memories and processes provide the building blocks necessary for high-level control in the pursuit of goals, learning, and the encoding of amodal, symbolic knowledge for abstract reasoning. SVI encompasses the quantitative spatial and visual depictive representations and processes specialized for efficient construction and extraction of spatial and visual properties.PhDComputer Science and EngineeringUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/60876/1/slathrop_1.pd