Research On Key Technologies Of Relevance Vector Regression Metamodeling And Their Application

Simulation technology is an essential tool for complex system modeling and analysis, which is widely used in the system design, performance evaluation and decision support. Due to the simulation model with high dimension, high nonlinear and multi-modal characteristic, the simulation-based performance evaluation is too expensive to be accepted in the complex system simulation optimization research. Simulation metamodeling technique is often used to model the computation intensive simulation model in order to improve efficiency and has become an emerging hot point in the field of system simulation. The simulation metamodeling can be described as a sort of constrained small-sample machine learning problems, which are able to gain sample points actively. For the metamodeling requirement of simulation optimization, this dissertation introduces Relevance Vector Regression(RVR), currently the Bayesian machine learning theory about small sample statistical learning and forecast, into the simulation metamodel building. Combing the newest technique in machine learning, several key problems in simulation metamodeling are studied. The main contents and innovations of this thesis are as follows:1. The simulation metamodeling is described as a small sample machine learning problem. RVR is introduced into the construction of the simulation metamodel. And the RVR metamodeling method is compared with Polynomial Regression, Kriging, Radial Basis Functions and Support Vector Regression metamodeling. Then its applicability is analyzed.2. The evolutionary algorithm based RVR multiple kernel function optimization method is presented to solve the RVR metamodel kernel function selection. The multiple kernel function is introduced into RVR metamodel. According to the difference of combination, the multiple kernel function can be divided into linear multiple kernel function and nonlinear multiple kernel function. For the linear multiple kernel, the memetic algorithm is used to optimize the weights and kernel parameters simultaneously. The nonlinear multiple kernel function is coded as tree structure, owing to its shape uncertainty. And genetic programming is used to optimize the nonlinear multiple kernel function. Simulation experiment results show linear multiple kernel function and nonlinear multiple kernel function have better prediction performance than the standard kernel function.3. To improve the applicability and generalization performance of metamodel, the ensemble of RVR metamodel is provided. This ensemble method uses Bagging and multiple kernel function selection to increase the diversity of individual learners, achieve the selective ensemble of individual learners by memetic algorithm. Simulation results prove this method can improve its generalization performance with little RVR individual in most circumstances.4. For the sake of online metamodeling requirement, a RVR metamodel incremental learning method using ensemble algorithm is presented. In the train phase, every new sample set is incorporated relevance vectors from the last RVR training for training new RVR individual. In the ensemble predicting phase, the k points in the training set nearest to the predict point are found and the performance of each RVR metamodel on these k points is used as measurement. Dynamic selection and dynamically weighted schemes are employed to combine the ensemble. Simulation results show that the proposed method can speed up the training while maintaining the prediction accuracy, but its prediction time is longer.5. Based on the above research, the RVR metamodel-based simulation optimization method is proposed, including the RVR metamodel based on the single-objective genetic algorithm and the multi-objective genetic algorithm. The Wi-Fi network throughput optimization problem and the multi-objective optimization problem prove its validity preliminarily. Taking the UAVs communication network simulation system as context, the proposed RVR metamodel based simulation optimization method is applied in the UAV communication network protocol optimization problem. The results prove its effectiveness and feasibility.