Research Of System Effectiveness Evaluation Based On Data Farming And Data Mining

Effectiveness evaluation of combat systems is one of the most popular and difficult problems in military operations research. There are a lot of evaluation methods, among which simulation technology plays an important role. As warfare form transforms from mechanization to IT application, traditional evaluation methods are facing many challenges, e.g., dealing with the"curse of dimensionality","curse of complexity", and nonlinear characteristics in information warfare. With the utilization of combat simulation technology, this thesis presents an evaluation method based on data farming and data mining to solve the above problems. The main results can be summarized as follows:1. The High Level Architecture (HLA) based framework of data farming system is studied with its subsystems, like data planting, data growing platform, data collecting, database and data analysis, etc. And a direct data collecting strategy is emphatically discussed, which is proven to be superior to the method recommended by Defense Modeling and Simulation Office (DMSO). Besides, according to Object Model Template (OMT), a scheme of designing HLA result databases, including relational table creation, data loading, index maintenance and data integral recovery etc, is proposed to solve some practical problems during HLA system development and application.2. In view of the multi-dimensional complex tremendous data produced by HLA data farming systems, a result management strategy with the adoption of data mining is systematically studied. The key techniques for data mining applications, e.g., the task understanding, the data extraction and preparation, the algorithm selection, the evaluation and application of mining models etc are analyzed in detail. Besides, in order to evaluate the effectiveness of some special fields within Command, Control, Communication, Computer, Intelligence Surveillance and Reconnaissance (C4ISR) systems, an On Line Analytical Processing (OLAP) based data mining approach is proposed. The problems during data farming processes are also identified and corrected.3. In order to explore the intrinsic complex characteristics of combat systems, such as nonlinearity etc, a data mining approach based on neural networks is presented for analysis. The following results are derived: First, an improved neural networks algorithm named NN-LMBP (Nearest Neighbor Levenberg-Marquardt Back Propagation) is proposed. In particular, the Levenberg-Marquardt optimization algorithm is utilized to improve the convergence speed of neural networks. Besides, a Nearest Neighbor based pruning strategy is adopted to enhance its generalization ability. And the effectiveness of the algorithm is verified by standard testing database. Second, the neural networks based data mining approach is exerted to evaluate the effectiveness of Command, Control, Communication, Intelligence and Electronic Warfare (C3IEW) systems. It is found that towards the evaluation model, neural networks can help to choose the weights of the bottom indexes and to deduce the nonlinear relationships between the input and output parameters. On the other hand, the structure of the neural networks can be optimized according to the evaluation model. The advantage is that the double problems are skillfully solved in the field of artificial intelligence and complex system evaluation.4. As data samples of HLA data farming system are limited with"curse of dimensionality"and"curse of complexity", a data mining approach relied upon Support Vector Machine (SVM) is further analyzed. The main contributions are outlined as follows: First, in accordance with the typical characteristics of data farming systems, the principles of Statistical Learning Theory (SLT) and SVM are discussed. Second, considering the difficulty of choosing SVM model, a strategy based on fuzzy similar matrix is provided to find the proper kernel parameters. The effectiveness of the strategy is proved by standard testing database. Third, referring to the learning experience of human beings, a novel SVM learning algorithm of BV-SVM is proposed. The initial step of BV-SVM is to learn some preparatory knowledge from boundary vectors in kernel space. Then with samples that violate Karush-Kuhn-Tucker (KKT) conditions, the final SVM is acquired by incremental training. Experiments indicate that, compared with conventional SVM, BV-SVM is faster in training speed and has comparable generalization ability for large data samples. Finally, an SVM-based military decision system is carefully investigated. The key elements in the battlefield are found by comparing partial derivatives. In addition, the hidden rules behind the chosen elements are further mined.5. The HLA result database management tool HLA-DATABSE is independently developed. And the SVM based data mining tool is designed with the cooperation of others. Accordingly, the HLA data farming and data mining prototype system is established and applied to project XXX. Besides, the evaluation of information superiority in project XXX is effectively accomplished with the help of the prototype system.