P.D. Basalin, K.V. Bezruk
In the article principles of organization and functioning of the intellectual support systems are studied. The systems based on models and methods of the knowledge based concept and neural net technologies for the decisions making are under consideration for problem solving in the area of hardly-formalizable or non-formalizable tasks. The architecture of hybrid intellectual system is proposed which combines the advantages of classical knowledge based approach and neural net problem solving methods and reduces the disadvantages incidental to each of the mentioned approaches. The first approach underlies analytical system component, performing classical symbolic deduction on the fuzzy knowledge representation models. The second approach defines the synthetic component, based on figurative situations perception and detection incidental to the artificial neural nets. Hybrid intellectual system shell consists of six main components: the knowledge database, which is originally empty. It combines the set of long term persistent data in the form of condition-action rules and dynamic working memory, where the facts describing current deduction state are stored; deduction Algorithm, representing the analytical core of the system, performing planning of the working scenario of the problem solving in the form of straight set of logical expressions using the search in deep strategy based on condition-action rules set from the knowledge database; knowledge acquisition subsystem, which creates the set of condition-action rules based on the input data in the form of solution graph (the infological representation of the problem are knowledge) and equivalent artificial neural net (for figurative perception and detection of standard situations); neural net learning mechanism, which uses as a learning examples the vectors of the initial system state as inputs and solution vectors determined by the analytical core of the system based on the rules from the knowledge base and which are used as a desirable outputs; explanation subsystem, which is capable of showing the sequence of condition-action rules used for solution making by the user request; intellectual interface, combining linguistic and software means, providing access for the user and knowledge engineer for the appropriate components of the system. Interface is capable of adjusting itself to use the terminology of the user creating the conditions of usage as comfortable as possible for the user. The implementation of the proposed hybrid intellectual system can be done with the use of both hardware and software components and can be used as a core of hypothetic neurocomputer, which is capable of flexible (in real-time) adjustments to the various problem areas for the hardly-formalizable problems solving, by using fuzzy analytical models.
One of the area where this approach is considered for usage is the problem of various problems representation in neural net basis, related to the selection of the artificial neural net architecture, topology, learning procedures and their adaptation to the to the particular conditions of usage in particular problems solving. All these require significant efforts for collecting and generalization of wide experience in usage of neural technologies in various problem areas and various technology areas.