Method of troubleshooting in the neural network environment of intellectual decision supporting systems

Abstract

The issues of error detection in artificial neural networks are considered. They are related to conflicting opinions of experts and / or limited (imperfect) description of the subject area. Approaches to their debugging are analyzed. Ways to improve existing approaches to debugging errors such as “forget about exclusion” are shown. Possible ways of application of the received decisions for debugging of errors “intersection of critical events” are shown. A formalized definition of the error of the neural network intelligent system is proposed, taking into account the requirements for efficiency and accuracy of information presentation. The issues of the influence of incorrect organization of machine learning on the accuracy of classification of elements of the neural network intellectual system are considered. The possibility of applying network contrast methods at the preparatory stage for testing the knowledge base of the intelligent decision support system is proved. This reduces the likelihood of errors of this type for these systems. The classification and analysis of algorithms for sampling knowledge from an artificial neural network are given. It is shown that to detect these types of errors it is advisable to use a modified GLARE algorithm with the adaptation procedure. Block diagrams of algorithms for debugging the knowledge base of an intelligent decision support system using the obtained theoretical solutions are presented. The scheme of the organization of testing process on levels of detailing for integration and modular testing is offered. This approach can be used to implement testing processes of Agile methodology, in particular: Agile Modeling, Agile Unified Process, Agile Data Method, Essential Unified Process, Extreme Programming, Feature Driven Development, Getting Real, Open UP, Scrum, Kanban.