Studies of resent years in the field of artificial intelligent led to a number of major achievements. The most significant of these was the development of powerful computer systems, known as systems based on knowledge. They are based on the program, designed to represent and apply actual knowledge from subject domain to problem solving. Diagnostic belongs to weakly structured domains, so that the essential (sometimes only) source of information in this domain is experts. In this paper the possibility of computerized construction of diagnostic expert systems' knowledge bases is described on the basis of solving the problem of expert classification. The described in the work approach provides a complete build of consistent knowledge bases of diagnostic expert systems.
achievements. The most significant of these was the development of powerful computer
systems, known as systems based on knowledge. They are based on the program,
designed to represent and apply actual knowledge from subject domain to problem
solving. A general overview of intelligent computer systems is translated into [
is considered for teaching students using a computer, an expert rating system in
commercial banks is considered. In [
factors, and parameters. It is problematic to provide a formal mathematical (analytic)
description of these factors. This leads to the fact that diagnostic belongs to weakly
structured domains [
In systems that are not created by a human (such as biological), there is no goal of functioning that can be formalized. The purpose of the functioning of such systems is their efficiency in general, the support of some parameters in the specified limits, but it is usually impossible to formalize such a goal in the form of a certain criterion. For instance, in the biological system, all factors affecting its functioning are so numerous and the connections between them are so complex and not obvious that it is impossible to set a certain function to describe the purpose of its functioning.
Due to the lack of a functioning goal that is subject to formalization, it is impossible to construct a function whose optimization would provide the best mode of operation of the object. Objective optimization function does not exist, you can specify only some of the factors that can be optimized. However, it is impossible to optimize each of these factors separately, since they are closely interconnected, and their connections will be disturbed during the process of optimization. This can lead to a violation of the process that regulates or maintains the system in a stable condition relative to the changing environment in which the system operates, and may lead to catastrophic events and irreversible changes in the system.
heterogeneity of the criteria that characterize the subject domain.
predict human behavior as an object of control or a component of a system. A person operates in the system, taking into account his personal goals and interests. Therefore, when modeling the object of human behavior management is difficult to consider.
Diagnostic expert systems are essentially designed to solve the classification problem:
each object (situation). From the subject area a sings his diagnosis. Properties, course of
action large role in the experts work of their knowledge is plays with the manual of the
organization of expert inquiry. Success in creating an information system became possible
due to the fact that a certain expert apparatus was developed, allowing to largely
overcome the above problems. The essence of this approach is use of data on the structure
of the problem to be solved, about the relaxation between its other elements. Such
information makes it possible to draw certain conclusions on the possible limits changes
of a number of estimates and thus reduces the number of questions to the expert, control
the consistency of information. Identify and eliminate appearing in it mistakes [
values of the signs have different degrees of characteristics and for diagnosed properties.
used about the varying degrees of the characteristics of the individual values of each
characteristics and for each property (or what is the same class). It is also assumed that,
for each feature, the expert may order its value by their characters and for the
corresponding class and this order doesn’t depend on the value of other characteristics.
The experiment of solving the problem of expert classification suggests that the
assumption of ordering the meaning of the attribute is also valid for many practical
problems, and accordingly, the considered formulation covers a wide class of the problem
of expert classification, the possibility of obtaining valuable information about the objects
membership class makes it possible to construct a rational procedure for expert demand in
order to minimize the number appeal to him. In addition, detailed information allows you
to identify possible errors in the expert’s answers [
= { , , … , }; M – signs, characterizing from different sides the objects of research; set th value of the m-th characteristics: possible mand nm – the number of these values; set of all hypothetical possible states of the object of investigation: = { ,
, … , = × × … × } ; in this state ∈
is characterized by a vector: Formally hypotheses a specifically can be described as follows, ordering characteristics values by their character for the property pi allows to add to transitive and antifriction binary relations (linear order): if qms more characteristics for this property than qmt. On this basis of these relation it is possible to construct a binary relation of dominance by characteristic for each property on the set of the state on the object of investigation which are described as follows: = {( , ) ∈ × | = , , ( , ) ∉
}, = ( ∈ , , , … ,
), = , . (qms, qmi) ϵ rm, ∈ : 1 ≤
≤ , , ∈ .
, such that a, have a reflexive transitive relation of strict dominance:
A = ⋃
k ,
such that, the state ∈ belongs to the class Ki, if the object in this state has the experts
opinion and properties Pi. To class Ko, in this case, there are such states in which the
object doesn’t possess, in the experts opinion, not one consideration. It is customary to
call this problem the main task of classification [
correspondence to each state ∈ to sets – set of the numbers of classes belong to the state ai and – set of class numbers, to which the state ai cannot belong. Let K – multiple numbers of generated classes:
K = {0, 1, … , L}. C C ∩ C
survey ∀ ∈ ∈ a subset of all classified state. Before the beginning of the expert suppose:
C = C = A = 0.
required expert advisor is selected: = . In the course of the survey the next
(object properties, is in this state). Thus for this state ai the set is defined explicitly and an implicit set: a ∈ A\A .
C
= K\C .
A
= A ∪ a . = ( , , … ,
) { , , … ,
}. ( , ) ∈ , if
< = { , … , }.
other states [
may be a need to clarify the severity of the diagnosable properties, which can be sensible ordered. With this definition of groups the state of the objects, in which he has the same degree of severity and diagnosable properties, means the spitting of the state of an object into a set of defined classes. It is customary to call this task the expert classification of the problem of order to classification.
On the scale of each feature ( = 1, (q
). A certain ratio of linear order: , q ) ∈ r , if more characteristic for this property, than object of investigation: , the set of all possible states of the A = Q × Q × … × Q
∀ ∈ and thus to construct a partition of the set A:
A =
Y (Y ∩ Y = ∅, i ≠ j, i, j = 1, n) such that ∈ belong to the class (1 ≤ ≤ ) if the objects in this state have this property of degree .
YB, which correspond to the state ∈ . Before the survey for: the classification can be considered complete, when: ∀ , G = 1. N, ∀ |G | = 1. a ∈ Y , k > a ∈ Y , k < . .
Let the expert determined, what the condition is ∈ corresponds to the value (1 ≤ ≤ ) by the degree of expression of the property under consideration, i.e. ∈ . Consequently, in this case, the state described by a set of characteristics is no less characteristics for this property, it cannot have a lesser degree of its expression that is, if ∈ and ( , ) ∈ , then:
characteristics for this properties, no more characteristics of this property can have a greater degree of its expression, that is, if ∈ and ( , ) ∈ , then:
is also possible to estimate the informative state is the number of indirectly classified states based on the characteristics relationship. For each state, you can find this number with every possible expert answer and calculate in the middle value or the minimum of them.
errors in the answer. These errors can be detected by inattentive, expert fatigue, and also the complexity of the problem being solved. Since the knowledge base must be nonselective, information analysis needed, obtained from an expert, controversy detection.
check the consistency to expert assessments. If there is discrepancy in the indirect and direct evaluation of the status this indicates that there are errors in his answer. It is necessary to present contradictory answers of experts for their comprehension and choosing the right way to assess the series condition.
Possible two strategies for removing the contradiction when building a knowledge base. One of this to continuously compare information, receive from the expert, from the received early and check for inconsistency. If there is contradiction between the last expert answer and the previous information, this contradiction is presented to the expert for analysis and choice of the contradiction of politics. Another strategy is to obtain from the expert or a part, either all the necessary information, and only then the implementation and in it of the search for contradictions and their removal.
In this way. The procedure developed by the expert survey should, on the hand, minimize the work of the expert and on the other hand, allow him to analyze the information received from him, from the point of view of its consistency. If the traditional way of building a knowledge base for diagnostics systems, an expert has to solve the problem of synthesizing of their knowledge, then the proposed method corresponds to its usual task of analysis specific situations. So, he unconsciously uses many of his skills and techniques, which is difficult for him to formulate in an explicit form.
The methods for solving the problem of expert classification have natural limitations, due to their dimensionality. In problems of large dimension (dozens of the signs with a large number of possible boundaries and a large number of diagnosed properties) the laboriousness of the experts work also sharply increases and the computational complexity of the procedure increases the solution of the problem of expert classification, which makes it difficult to use them directly. In practical problems of classification, it is often possible to distinguish separate groups of attributes according to their semantic content in such a way that the characteristics to one group reflect on of the sides of the object under consideration. In these cases one of the possible ways of constructing the coverage of the initial state set is to decompose the original problem into subtasks less than the size, the used classification procedure will consist of several stages.
diagnosed to properties. Therefore, according to the values of the i-th group of characteristics ( = 1, ) the expert can make a preliminary conclusion about the possible classes of states, the characteristics described by these values.
Each group of characteristics gives incomplete, only a partial description of the state of the object of investigation. Accordingly, the evaluation of the expert, it is based only on the part of the object description, can has a probabilistic character and reflects the varying degree of the experts confidence in the availability or the absence of the object in the given state of the diagnosed properties. Consequently, at the first stage of solving the problem, the expert decides its main, and order any classification tasks within each group of characteristics, which allows to build an orderly classification, and for each properties for each group of each characteristics. This give opportunity for each state: a = (a , a ,, … , a ) and for each property Pi define a vector ( , , … , ), where TS – degree of suspicion of i-th property in the state by j-th group of attributes.
The next stage of the solution consists in construction generalized classification for each property based on the other of their classification by separate groups of characteristics. In this case, the signs of the second level are also the degree of the experts confidence in the availability of the i-th property for each of the T group of initial signs. It is necessary for each property ( = 1, ) on the basis of the value of the attributes of second level, construct an order classification, determine whether or the absence of each of the property under consideration and its degree of serenity.
If the expert cannot make a diagnostic conclusion for a certain group of signs. Not having information about the result, received for another group of characteristics, the survey is conducted as follows: if the diagnosis is on i-th group, the expert must build L order of classification for i-th group of characteristics with the addition of the result for property pi by j-th group of characteristics, number of hierarchy levels (as the decomposition levels) can vary in different tasks. It is determined the classification problem.
knowledge bases is described on the basis of solving the problem of expert classification.