A. P. Sviridov, A. M. Titov
In this article there algorithms of learning and neural network result explanation, which is used for computer control of knowledge are considered. Neural network technology can be used not only in modeling relations “student – tutor” in computer control of knowledge, but also in consideration other relations of “human – human” type. Questions of practical realization of created algorithms are considered also. The problem of computer control of knowledge is considered as problem of recognition (classification) of images. Image is a sequence of levels of truth of answers on questions, which are given to students. Class is a mark or quality of knowledge of student. To solve this problem it is supposed to use neural networks.
Using of neural networks causes several problems, such as:
it is impossible to describe set of rules which are used by neural network when it makes a decision. Neural network is considered as “black box” - an algorithm with unclear internal logic.
there is no clear algorithm for modeling training sequence for neural network.
the best structure of neural network (quantity of layers and quantity of neurons in each layer) is undefined.
To solve these problems following algorithms were developed:
algorithm for generating of training sequence for neural network and computer control of knowledge;
algorithm for searching of topology of neural network for computer control of knowledge;
algorithm for explanation of decisions which are made by neural network for computer control of knowledge.
In algorithm of generating of training sequence the program iterative generates sequences of answers on questions (solving of problem), and tutor approves them. During one operation iteration is generated (and approved) one sequence of answers. Advantage f this method of generation of learning selection is that system can create images, which increases its representativeness if they are included in this learning selection. In this case there is generated that all space of meanings of signs is covered. In the same time system should work for minimize request to the tutor and give him opportunity to stop generation in any moment.
To analyze (explain work) neural network the set of fuzzy statements, each of which consists 2 types of linguistic variables: context importance and context utility is built. Mentioned algorithms of explanations of decisions of neural network is based on algorithm, which was offended by K. Flaming.
Optimization of topology of neural network is realized with using of genetic algorithm.
These algorithms were realized in developed dialogue system of analysis and synthesis of standardized and neural algorithms of computer control of knowledge “DIANIS”).
The program consists of 2 parts: system of testing and system of control of testing. Developed system has following main features:
Three-level control of knowledge: level of knowledge of question, part and course.
Adaptive choice of questions – economy of time during testing without loosing of quality.
Possibility to use neural network and standardized algorithms of control on level of course and part.
Genetic algorithm of modeling of learning sequence during learning of neural network.
Saving of information not only about level of knowledge of curse, but also about level of knowledge of other parts.
Availability of mechanism o explanation of sequences of neural network.