| Hi-index | 0.00 |
This dissertation addresses problems encountered in combining information from multiple sources. Novel methods for learning parameters for information aggregation are proposed. In practical applications of pattern classification, multiple algorithms are often developed for the same classification problem. Each algorithm produces confidence values by which each new sample may be classified. We would like to aggregate these confidence values to produce the best possible confidence for the given sample. This can be seen as a particular instance of what is called information fusion. In addition to learning parameters of aggregation operators to assign the best confidence for a given sample, we would also like the aggregation operators to use a subset of the algorithm confidences and achieve the same level of performance as the entire set of confidences. Using a subset of the algorithms implies lower cost for applications. Choquet integrals are nonlinear operators based on fuzzy measures that can represent a wide variety of aggregation operators. Previous research has demonstrated the utility of Choquet integrals for this problem compared to other methods such as neural networks and Bayesian approaches. However, one of the novel results of this research is that the measures learned can be very sensitive to the choice of desired outputs. In response to this problem, we propose an alternative training methodology based on Minimum Classification Error (MCE) training that does not require the use of desired outputs. A problem with this method is that it depends on a constrained type of fuzzy measure, the Sugeno measure. There is a need for additional approaches to learning unconstrained fuzzy measures that are more computationally attractive and provide more robust performance. We propose an approach to learning unconstrained fuzzy measures that relies on Markov Chain/Monte Carlo sampling methods. The use of such approaches for learning measures for Choquet integral fusion is completely novel. In addition, we propose the inclusion of the Bayesian approach of imposing sparsity promoting prior distributions on the measure parameters during sampling as a way of selecting subsets of the algorithms for inclusion in the aggregation. This approach is completely new for learning fuzzy measures.