| Summary: | 博士 === 雲林科技大學 === 工程科技研究所博士班 === 99 === To analyze a system with many attributes, we always need to recognize how a specified objective attribute depend on other attributes. It is essential a frequent problems in information fusion. Traditionally, the most common aggregation tools are the weighted average method and the linear regression. These methods are all linear and must make a basic assumption that there are no interactions among predictive attributes. However, in many real-world problems such as a behavior-based intelligent mobile robot, the weather forecast, etc., the inherent interaction among predictive attributes must be considered circumspectly; meanwhile, these kinds of problems are essential non-additive systems. Hence, a nonlinear integral model with respective to a non-additive set function is suitable for dealing with these kinds of problems. In general, these kinds of models constitute over-determined systems with nonlinear integrals and then, it is extremely difficult to acquire the analytic solution. Therefore, an efficiently intelligent computing technique is necessary for these models to perform precise estimations of model’s parameter.
In this dissertation, different nonlinear multi-regression models based on the Choquet integral (NMRCI) are considered for modeling non-additive systems with different type of predictive attributes, respectively. Meanwhile, the parameters estimation for the simplest model is also performed via a modified particle swarm optimization with quantum-behavior (MQPSO) algorithm, in which the mechanism of multi-elitist crossover is included and then, concepts of elitist reproductions and adaptive decay are introduced to improve the MQPSO algorithm such that the generalized NMRCI model can be estimated well. That is, the genetic algorithm (GA) and the simulated annealing (SA) algorithm are embedded in the improved algorithm which is named MQPSO with nonlinear creative coefficient (MQPSO-NLB) algorithm. Furthermore, the effects which are caused from outliers are also considered for extending these proposed NMRCI models such that the extended model can model a real-world non-additive system well. Meanwhile, to efficiently estimate the extended model parameters, the robustness of the proposed MQPSO-NLB algorithm must be gained. Therefore, the high breakdown regression estimator, least trimmed squares (LTS) is then introduced to eliminate the deviations caused by the observations contaminated with outliers. That is, the LTS estimator is instead of the LS estimator in the MQPSO-NLB algorithm and then, the proposed intelligent computing algorithm which combines the mechanisms of the GA, the SA algorithm and the LTS estimator into the QPSO algorithm named LTS-MQPSO-NLB algorithm can deal with the generalized NMRCI model with contaminated observations for a non-additive system with outliers. From the numerical simulations, satisfied results can be achieved.
|
|---|
