Известия Томского политехнического университета: Инжиниринг георесурсов (May 2019)
Automatic meta-learning system supporting selection of optimal algorithm for problem solving and calculation of optimal parameters of its functioning
Abstract
The relevance of the work is caused by necessity of increasing efficiency of automatic data mining systems based on meta-learning. The main aim of the study is to design an automatic meta-learning system supporting selection of optimal algorithm for problem solving and calculation of optimal parameters of its functioning. The methods used in the study: inductive modeling, methods of statistical analysis of results. Results: The known meta-learning systems were integrated based on produced classification features taking into account internal structure of systems. The author has stated the requirements for implementation of the automatic meta-learning system and has offered the way to build a meta-learning system satisfying all stated requirements and accumulating meta-knowledge, building meta-models on its basis, selecting optimal algorithm from a set of available ones and calculating optimal parameters of its functioning. The object-oriented architecture of a software framework for implementation of any meta-learning system presented in the systematization was developed. The efficiency of the implemented automatic meta-learning system using algorithms of group method of data handling was experimentally examined being applied to solution of problems related to the short-term time series forecasting (1428 time series from the testing set known as "M3 Competition").
