Assessment of Simulated Cognitive Maps: The Influence of Prior Knowledge from Cartographic Maps

Real cognitive maps encoded by humans are difficult to study using experimental methods because they are a product of complex processes whose content and timing cannot easily be known or controlled. This paper assesses the value of using neural network model simulations for investigating cognitive maps. The study simulated the learning of mapped city locations in South Carolina from reference sites in the three primary regions of the state using Kohonen self-organizing maps. The learning performances of models were considered based on available prior knowledge. Bi-dimensional regression analyses were used to assess the congruity of the simulated cognitive maps with a cartographic map and with sketch maps produced by human subjects. Error analyses indicated differences between central and peripheral reference sites. The cities known by subjects living at a central location were more evenly distributed in space and associated with significantly smaller errors. Models that learned combined state boundary and interstate highway information as prior knowledge or simultaneously with city locations consistently produced the best simulation results. The results indicated simulated cognitive maps could be used effectively to study the acquisition of spatial knowledge.