Conceptual spaces as a modeling system for Information Fusion.

摘要

In the world of Information Fusion, there are many algorithms and techniques utilized to help understand situations occurring within a user system. Traditionally these mathematical models follow either a Symbolic-type (rule-based) or an Associationist-type (feature-based) cognitive model such as a logical statement-based system or a neural network respectively. Although often successful, each of these modeling procedures has both their merits and their drawbacks. In his work on Conceptual Spaces, Peter Gardenfors offers a means to "bridge the gap" between Symbolic and Associationist models. He suggests that these two models can and should be utilized together; however, in order to do so they must be connected by another model. Conceptual Spaces represent the way in which humans understand concepts within their world by way of convex geometric spaces.;We offer a new approach to information fusion systems through a hybrid model joining Conceptual Spaces and Mathematical Programming. Conceptual spaces are the modeling piece while mathematical programming is the tool by which the models are quantified. We achieve a novel system through a single mathematical program that can solve various fusion related problems including association of observations to objects, classification of observed objects, determination of changes in objects over time, relationships between the observed objects and detection of overall situations, based exclusively on feature-based sensory reports. The system handles multiple observations of multiple objects by multiple sensors within a single integer programming model.;In this thesis, we first introduce the Conceptual Space--Mathematical Programming Hybrid Model for classification of observed objects and discuss its computational complexity. We then provide an example system in the field of Emotional Recognition wherein we consider facial images to understand which emotion is truly felt or being faked by the person in each image. The hybrid model proves highly successful in both classification accuracy and computational time as compared to the widely utilized support vector machine modeling approach. We continue building the hybrid model by adding further capabilities in considering observed changes over time, relationships between objects and classifying situations, thus providing a single model with the ability to capture both level one and level two fusion.