Neural Network Model Synthesis Based on a Regression Tree
- PDF / 581,371 Bytes
- 10 Pages / 612 x 792 pts (letter) Page_size
- 87 Downloads / 265 Views
eural Network Model Synthesis Based on a Regression Tree S. Subbotin* Head of the Department of Software Tools National University “Zaporizhzhia Polytechnic,” Zaporizhzhia, 69063 Ukraine *e-mail: [email protected] Received September 10, 2019; revised December 30, 2019; accepted January 14, 2020
Abstract—The problem of dependency modeling by experimentally obtained observations is considered. The objective is to develop methods for neural network model synthesis allowing to automatize, simplify and speed-up model building. The mathematical support for neural network model synthesis is developed. It contains set of methods that transform the sample into a decision tree or a regression tree, on the basis of which the neural network structure is formed and the parameters are adjusted. The experiments on practical problems solving were carried out. Their results were confirmed the efficiency of the proposed methods. The results of the experiments allow to recommend the developed methods for solving the problems of constructing neural network models on precedents. Keywords: regression, decision tree, regression tree, neural network DOI: 10.3103/S0146411620040100
1. INTRODUCTION Artificial neural networks [1–4] are widely used in the problems of technical and biomedical diagnostics [2], pattern recognition [3], and forecasting [4], due to such their advantages as generalization of observations, learning, universality of approximation, compactness of the model, high accuracy of the model, as well as massive parallelism of computations. However, the problem of choosing the optimal model structure arises in practice at neural network model building. One of the ways to solve this problem is to involve a human user who must define the network structure. The disadvantage of this way is its dependence on a human, the non-optimality of the chosen model structure and its possible inadequacy to the task. Another way is to solve the optimization problem of iterating over various network configurations with the choice of the best configuration in automatic mode. The disadvantage of this method is large time spent on finding the optimal structure due to the high iteration of the search and the need to train the model for each structure variant. Therefore, to build a neural network model, it is advisable to perform a preliminary data analysis, which the results will make it possible to form the neural model structure. As a data analysis tool, it is possible to use decision and regression trees. The decision tree [5, 6] is a model of a dependency, represented in the form of a tree, which nodes contain checks for compliance of the feature values of recognized instance. The regression trees [7, 8] are used in problems with a real output. They replace the estimation problem by the classification problem usually loosing accuracy. Decision and regression trees have such advantages as interpretability of the model and its decisions, the possibility of simultaneous use of categorical and quantitative features, the universality for classific
Data Loading...
