Khmilovyy S. Informational technology of knowledge discovery for time series forecasting by the example of carrier loading.

A dissertation for obtaining the academic degree of Doctor in Technical Science (Ph.D. Thesis) in specialty 05.13.06 - Information Technologies - Donetsk National University, Donetsk, 2009. An important scientific and practical problem of knowledge discovery for time series forecasting is being considered in the Thesis. At the preliminary data processing stage the significant factors selection is found the most important. In conditions of the factor set valuation stochasticity (the one to be obtained at the time series forecasting using neural networks) the methods known are seldom applicable and require modifying. Here proposed a compact genetic algorithm (CGA), its probabilistic vector alteration step is modified. The modification changes the value of CGA probabilistic vector depending on comparison reliability of paraphernalia subset. The knowledge discovery process has been conducted by means of automatically evolving fuzzy rules database construction. Automatic rules construction is being made with evolutionary algorithm. There was modified the fitness-function of evolutionary algorithm to improve accuracy of the rules constructed. There were preserved the parts of initial function responsible for the rules diversity and for the short covering of mistaken points. The J-measure is used as main accuracy part. The Linguistic Database (LDB) was modified to improve accuracy of derived knowledge database. Modification of fuzzy inference system by means of Mamdani algorithm substitution for simplified algorithm allows to implement any kind of membership function. LDB parameters optimization on base of optional membership function with assistance of (1+1) - evolutionary strategy allows practical reaching of forecast accuracy produced with NN (neural networks). There for the first time on genetic fuzzy systems was created parallel evolutionary algorithm to make the knowledge database. It allows to increase productivity almost linearly under a little amount of clients generating rules. There was modified the stage of postprocessing where applied measures both for accuracy improving and rules' interpretability: multisimplification, deliberation of rules, tuning. Here proved effectiveness of 1+1 - evolutionary strategy implementation into tuning process. Here created algorithms for information technology software, designed hierarchy of objects for object-oriented software realization. The software was tested at benchmarks, here proved efficiency of the methods and modifications proposed. Here designed technical and organizational technology providing. There conducted an approbation of technology at Promtelecom JSC for a task of amount forecasting of connections to the company automatic telephone exchange. Key words: information technology, significant factors selection, genetic algorithm, fuzzy inference system, knowledge database, time series forecasting, automatically evolving fuzzy rules construction, rules postprocessing, interpretability.