Russia. Moscow. Research and development institute on information and computer-aided technologies
NEUROLOGICAL MODELS IN INTELLECTUAL TECHNOLOGIES
The problems on integration neural networks and expert technologies for problem solving of increased complexity, including problems of a creative character are considered. The determination of a creative problem is given. Are considered force and weak parties neural networks and expert technologies and is concluded a possibility of association of their strengths. Theoretical basis of a research is the information approach.
Россия. Москва. Российский научно-исследовательский институт информационных
технологий и систем автоматизированного проектирования
НЕЙРОЛОГИЧЕСКИЕ МОДЕЛИ В ИНТЕЛЛЕКТУАЛЬНЫХ ТЕХНОЛОГИЯХ
Рассматриваются вопросы интеграции нейросетевой и экспертной технологий для решения задач повышенной сложности, в том числе задач творческого характера. Дается определение творческой задачи. Рассматриваются сильные и слабые стороны нейросетевой и экспертной технологий и делается вывод о возможности объединения их сильных сторон. Теоретической базой исследования является информационный подход.
The intellectual technologies largely based on hypothesizes of a symbolical system and search /1/. The symbolical system is: a collection of symbols forming symbolical structures; a collection of processes which are capable to create, to delete and to modify symbolical structures. The symbolical structures have two main properties: can designate objects, processes and other symbolical structures; and if they designate processes, they can be interpreted. If the symbolical system can realize a behavior, defined by the given essence, or can influence this essence therefore it designates some essence (object, process or other symbolical structure). Symbolical systems have necessary and sufficient conditions for a realization of intellectual operations (hypothesis about symbolical systems). The symbolical systems decide problems with the help of search, i.e. they generate potential solutions and a gradually modify them to satisfy to specific conditions of a solution (hypothesis about search). It seems that the majority of existing intellectual systems confirm these hypotheses. However more deep analysis shows a deficiency of a logic paradigm for a solution of practically problems of a creative character. It needs of something else.
We consider a creative problem is such problem, which conditions are contradictory and incompatible: the intersection of decisive sets assigned by conditions, is empty, does not contain any submission. The solution of such problem means the sanction of an inconsistency. We treat a solution of a problem as shaping of submission on a base of specific properties. The properties are set by conditions of a problem. Logically each such condition is a predicate, which defines some property and set of objects, for which this property is true (“a decision set”). In this case the solution of a problem is conjunction of predicates, intersection оf appropriate decision sets. If the intersection is not empty, a solution can be obtained by usual regular methods. But if any two subsets are not intersected, the set of solutions appears empty. It means, that the appropriate conditions of a problem contradict one another, there is no submission, which would answer all these conditions. But it just most interesting situation, because a majority of nontrivial creative problems (for example, invention) is contain inconsistencies in the initial axioms.
One of received results /2/ is that the inconsistency in conditions of a creative problem is not absolute. It is a consequence of the limit of an initial set of submissions of the subject. The decision sets assigned by conditions of a problem, are not intersected only within the limits of an initial set, but quite can be intersected out of his limits. Therefore, it is necessary to leave these limits for a solution of a problem. The difficulty is that the limit of an initial set of submissions is not usually realized by the subject. It is a consequence of the non critical acquired knowledge, traditions, customs, prejudices etc. It is represented expedient to apply hybrid models and systems for a solution of creative problems. The joint application of neural networks technology and technology of dynamic expert systems (ES) of real time which are supplying one another have a good perspective. Strengths of dynamic ES are a possibility of a solution of poorly formalized problems, which have one or several of the following performances /3/: problems can not be given in the numerical form; the purposes can not be expressed in terms of the precisely certain goal function; there is no algorithmic problem solving; the algorithmic solution exists, but it cannot be used because of limit of resources (time, memory). However the availability at least of one expert for use of ES technology is necessary. He can on a base of his knowledge choose the main concepts, relations and known to him methods of a solution of a problem.
Neural network technology supplements ES technology. This technology ensures: a possibility of a solution even of those problems, for which there are no verbalized expert knowledges and/or algorithms of a solution; a possibility of adaptation to conditions of operation and tutoring. This technology demonstrates an absence of necessity in detail programming of process of a solution of a problem. Neural networks technology gives to the user a universal nonlinear adaptive element with a possibility of a broad modification and set- up it of parameters. Possessing some kind of “designer” containing of such elements, the user receives a possibility to simulate the diversified processes, connecting them in a network. The user does not have necessity to think about processes happening in a network. The purposefulness and the optimality are guarantied them beforehand.
Using an analogy to a human brain, it is possible to compare ES traditional with left hemisphere working on a base of deductive methods. With an appearance of neural networks ES acquires “ the right hemisphere ”, i.e. a possibility of figurative thinking, intuitions. But in a human brain the functions of both hemispheres have developed during a long evolution and integrally are connected. Just this connection ensures effective interaction of the logic and figurative forms of information processing, mind and intuition, creative thinking. Anything similar yet in modem intellectual systems is not observed.
Theoretical basis of a research is the information approach which has received a reflection in a number of the publications /4,5/. The main hypothesis is that a behavior of adaptive systems (biological, technical, social etc.) submits to an extreme maximum principle of an information and has by the purpose the best adaptation of a system to an environment. A quantitative measure of adaptation is the average mutual information between conditions of environment (“ by stimulus ”) and responses of a system. Thus maximum principle of an information is a maximum principle of adaptation expressed in the approaching quantitative form. It allows applying a principle to such phenomena as thinking and creativity considered as adaptive processes. In particular, this approach allows to formulate in language of an information theory the difference between conceptual and figurative thinking, to offer new algorithms of tutoring and receptions of a solution of creative problems, to optimize processes of information processing and to receive a number of design results, which can be used with reference to intellectual systems /4.5,6/.
Let's consider it in details. The interaction of neural networks with a logic (expert) level of a system gives the following preferences. The own sensual experience is limited both at the person and at neural networks. Many conditions in space of properties appear basically inaccessible: or they seldom happen actually, or the stay in them is connected to danger and possibility of loss for a system, or that there are no sensors and engineering of measurements for the certain range of magnitudes etc. A circulation to logic methods can be useful there: the models constructed on a base of physical and mathematical regularities allow to imitate such situations. Neural networks, as well as the person, not at once and not in the general format acquires the laws of a nature and restrictions existing in a reality. Therefore intuitive methods can reduce to conditions, which actually are prohibited. Here again information incorporated in a knowledge base at a logic level can be used for adjustment of intuitive conclusions received on neural networks level. If there is an information about the laws and rules, connecting network variables, at a logic level, the learning process can be reduced by before tuning of a network. It signifies, that we are set by not casual initial significances of synapse weights of neural networks (as it is usually done), but we can add to them significances appropriate to the rules, known to us.
We know what difficulties arise in tutoring, if the condition of a network in weight space hits on flat plots of performance. The gradient of goal function appears zero and The network loses a reference point for further movement. There is a so-called paralysis of a network and the learning process is blocked. The circulation to a logic level can help a system to overcome such “ a dead zone ”. One from component of creative process is a generating of new submissions, new combinations of properties. However blind search of combinations is not fruitful - because a lot of them have not a physical sense and can not be realized. The inventor or scientist usually intuitively feel, what combinations are perspective, and what are certainly fruitless. By the before tuning it is possible to set and to keep constants of significances of some variables of a network. By that it is possible to receive at giving of a noise on an input not completely casual samples, but samples having some fixed properties. For example, conditions of a creative problem can be such properties,. The network will generate not any possible combinations (majority from them will be senseless), but only what answer conditions of a problem and restrictions, existing in an environment. It will be a noise, but it will be a directed noise, which is subordinated to definite purposes. Probably just so the creative person imagination for the works, when a person creates not any combinations, but only that relate to business.
1. Newell A., Simon M.A. Computer science fs empirical enquiry: Symbols and search//Communications of the ASM.-1976.- V.10.- №3.
2. Golitsyn G. A., Petrov V. M. Information and Creation. — BaseLBirkhauser Verlag, 1995.
3. Попов Э.В., Фоминых И.Б., Кисель Е.Б., Шапот М.Д. Статические и динамические экспертные системы. Учебное пособие, M.: Финансы и статистика, 1996
4. Голицын Г.А., Петров В.М. Информация- поведение - творчество-М.:Наука,1991
5. Голицын Г.А., Фоминых И.Б.. Нейронные сети и экспертные системы: перспективы интеграции. Новости искусственного интеллекта, M, АИИ, .№4, 1996.
6. Фоминых И.Б. Некоторые формальные аспекты информационного подхода к построению нейроэкспертных систем. Известия РАН. Теория и системы управления, №5,1999.