RU29597U1 - Connected Expert System - Google Patents

Abstract

A connected expert system (ES), including a knowledge base (KB), a database (VD), a processor with dynamic memory connected to a logical inference machine (MLV), which includes an interpreter and selector, an explanation block, a unit for organizing user interaction, which differs by the fact that a duality check block is introduced in the MLV, and the base of values is made in the form of a pertron-based neural network connected to the MLV selector, the input of the duality check block is connected to the MLV selector, its output is connected to the interpreter m MLV.

Description

CONNECTED BlvGliijiPTHiin CHCT ivlii

The utility model relates to the field of computer technology and automation, designed to study the characteristics of pattern recognition systems, can be used in equipment to increase the information content of expert systems (ES). There is a well-known expert system () an expro-pilot for managing the quality of flight activity (see link 30 SP expro-PELOT FOR managing the quality of flight activity. SELECTIVE TECHNOLOGY. QB systems. Management. Issue 2. 1290. Moscow-Sofia), which imitates discussions of highly qualified experts in the process of analysis and evaluation of flights performed by crews and provides information on the most probable reasons for the permitted crew and the corresponding control actions :. The expo-pilot system presents an ES focused on the diagnostics of crew activities when analyzing flight information recorded by flight recorders.

with the help of an expert, the knowledge engineering hypotheses yfvio conclusions regarding different: aspects of crew activities. Hypothesis - there is concession on a certain set of facts or other hypotheses of the lower level of the hierarchy; according to knowledge1, the investigator introduces the formulated hypotheses into the knowledge base, which also entails the formation of a questionnaire for the subject area, the addition of new texts to the explanation tables. After that, the system is ready to diagnose data coming from the storage medium.

The input of the program-converter receives data read from a magnetic tape and processed by a set of software tools. Ilporpamia converter converts numerical flight data into a multitude of facts of the betrayal area. The facts thus formed are fed to the input of the ES, after which an analysis of the feasibility of the hypotheses of the betrayal region is carried out. Missing facts for hypothesis testing can be entered by the user in a dialogue mode. If the hypothesis is true, the user can view the chains of explanations of this hypothesis.

The expert is given the opportunity to use all the directives inherent in the system1: add and delete facts, selectively test hypotheses, work with libraries where user responses are written.

KB is a set of inference hypotheses. Each hypothesis can be defined as E- ()., Where F is the set of facts of the subject area (obtained on the basis of user responses; Н - MHOsiecTBo hypotheses of the lower level (in relation to this one); b - a lot of literals that determine quantitative relations between the ranks of facts (ranks are entered by the user in the mode of working with ranks); P is a lot of slots.

x / 3 / fb

V - rank from the user's beta (point values are al interval). The literals that determine the relationship between the ranks allow you to set thrush relations equal to, more, B less than I.t. Slots-procedures allow you to calculate the values of arbitrary parameters specified in the hypothesis, based on the available answers of the user or by presenting additional polling programs. ES is implemented in the Prolog language.

.However, in this ES, conclusions based on incomplete anshormation are not realizable.

Known automated system for acquiring knowledge in an expert system in the management process. W Ptoc-Cbi pi - / :) r rftif W {) (.,

T. 1989, 136 is 6. Express inc., calculated T. Technology B 36, 1990, taken as a prototype. The ES contains a processor, an inference machine (No. 18), a knowledge base (KB), a communication unit, and an ES training unit are productive. In it, the working memory of the processor contains the observed (controlled) variables of the object, and the production memory contains a list of rules for concluding the settings of the control variables. The knowledge base includes lists of rules, and each list is associated with the conclusion about setting one control variable. Each list of rules may contain diagnostic rules and object management rules. IIoaToiviy important ordering rules. The rule replenishment strategy is to replace existing rules with more esophistic ones that are added to the end of the list.

In working conditions; : ES continuously takes actions aimed at maintaining the operation of the object within the given boundaries. In this case, after each action of the ES evaluates the essence of this action.

  . led to wrong actions.

The knowledge acquisition mechanism uses this rule to search for an error rule. The knowledge about the rules contained in the BS of the ES is entered into the information module on the rules, which uses a frame representation. Using frames allows you to search for new rules based on those data that are available in the UZ.

In the process control environment, it is possible to select the mode of operation of an object that contains information on whether the set of control variables has a definite value; after the process, they are classified as real values of the controlled variables. Therefore, the greatest aspect of the derivation procedure is the selection of examples of the operation of the object that contain a maximum of information regarding the setting of the values of the control variables and the state of the controlled variables. Ind: 1uch1aie rezhigzh selected mechanism acquisition values obektakBZ DB; A genetic algorithm for deriving rules is used.

However, in this ES there is no module of knowledge replenishment on the basis of incomplete information in conditions of damage.

The purpose of the utility model is the development of an expert system, the knowledge base of which is formed on the basis of incomplete reliability of knowledge, symbolic representations and transformations.

To solve this problem in a connected expert system, including a knowledge base (KB), a database (DB), a processor with dynamic memory connected to a logical output indent (d1LV), including an integrator and selector, an explanation block, an organization block , a block was introduced in the Major League Baseball

duality check, right the knowledge base is made in the form of a perceptron-based neural network connected to the SHSh selector, the input of the duality check block is connected to the MLB selector, and its output is connected to the MLV interpreter.

4.

Block diagram of the proposed connected ES; prEVEdena in figure 1, GDZ depicted:

1- CPU

2- logical EVODBB machine (ushv)

3- neural network (NS) - knowledge base (KB)

4- block of organizations of ezazmotekstva with user

5- communication unit

6- training unit

7- interpreter

8- explanation block

9- selector

10- duality check block

11,12,13 - output and intermediate cells of the neural network (NS) 14 - output cells (NS)

Information confirming the continuity of the implementation of a useful modem A connected expert system includes a knowledge base (KB) 3, a database (DB), a processor with dynamic storage device I, a logical output machine (MLV) 2, including an interpreter 7, a selector S, an explanation block 8, a block organization of interaction with the user 4, a block for checking duality 10 has been introduced in SE 2.

The processor with Dynamic Storage Unit I is connected to a logical inference machine 2. Knowledge base 3 is made in the form of a neural network based on a persentron, connected to a selector 9, the input of the duality checking unit 10 is connected to a selector 9, and its output is connected to an interpreter 7.

The system operates as follows.

-explained by the number of ANALOGUES obtained and the consequences, if necessary,

-interactive query seen necessary (additional) information,

-modularity of the structure.

Computing system - ES should correspond to the level of a professional expert, decision-making methods (judgment method) should be reproduced in S form understandable to the expert and user due to the ability to change the shorthanding of requests and tasks and the sequence of their occurrence. The system has the ability to use, acquire and store general and private schemes of deforestation, built not on completely reliable data and symbolic transformations.

Such requirements are satisfied by a hardware-software system capable of solving classes of unformalized problems.

The main components of such a system are considered

1- processor that communicates the component with the user or element

2- Machshra logical inference, providing the construction of schemes of reasoning.

3 knowledge bases for storage and processing of knowledge represented by logical, production or semantic models

4- subsystem of explanations or administration of the knowledge base

Processor I provides the interaction of the user and the element with the hardware and software of the ES by converting (translating, converting, interpreting) sentences in a natural language into sentences in an internal language (metalanguage) and vice versa. The KB is involved in these transformations. Entering character string information, the process of forming the interface is implemented

The user is explained the reasons for the requests E to decrypt the contents of the database. The peculiarity of presenting knowledge (ES is the condition of dividing into declarative (facts, data), procedural, governing and meta-knowledge - (for the purpose of structuring knowledge). Declarative knowledge is considered as facts presented by the user, procedural - as rules presented by an expert, controlling knowledge - a set of strategies that determine the MLF operation. Meta-knowledge is presented to the user and the examiner during the functioning of the ES, with the help of which its state, structure and reasoning scheme are revealed.

The main property of ZS is to change its structure and content in the process of functioning. The principle of ES learning by changing the structure and content is combined with the principle of an invariable set of tailored procedures.

iVfeiMHa inference 2 is intended to generate recommendations for solving an applied problem based on information concentrated in KB 3 and is based on the theory of the Post machine. The system produces rules based on aullification: R-; 1 R where is the production rule extracted by the KB, If is the condition for applying the rule Ei, i is the generated rule that can be placed or not placed in the BZ-3. In the process of solving a problem in the system, an interpretation (block 7) is carried out — an appropriation of one or another rule and the execution of actions determined by this rule. The selection (identification) of a rule is based on a comparison of conditions 1; and in the general case leads to several rules at the same time, i.e. generates a conflicting set. Conflict resolution is carried out by the procedure - selector 9. For the operational storage of intermediate data according to the condition 1 l Ij, a working memory is provided - a blackboard. The logical conclusion is carried out with the use of heuristics implemented in the environment. The logical inference machine 2 (MB) recognizes the rule, the E conclusion of which contains the solution closest to the problem. the premise of the rule is not defined; unknown variables are enumerated: in the premise of the rule, as applied to new conditions The operation is repeated cyclically until a solution is found, or until a problem is solved. Settlements to the rules are formed using the duality check block 10.

In ES, in block 10, the next 11th lattic model is considered. By S denotes the EU, setting

(. (I)

where / And - some set of sets; y is a family of inference rules. Believes that if

, Mb№ (j

(then for rule bf his work can be represented as

derivation of a corollary of the form (iglations If A, then B)

} ()

Then the dual model ES (i) means writing down the necessary -shk and sufficient conditions for the fulfillment of the following (3). These conditions are of the form:

m

Here P is an image (class of elements), indicated by Deity

moreover

P {11m) Pm / 5th

oso

When constructing the dual problem im immersed the original problem in some neighborhood in the space of problems, we have the properties on all the problems of this neighborhood.

You can look at the ES as an existing output of SLEDSTESH data and know, in addition, 2 as the task of making an effective choice, i.e. finding an item

 (Cj

J

Finally, as the task of diagnosing objects and situations, i.e. as

on the task of pattern recognition. This shows that all these problems can be looked at from the point of view of a unified model (I) and, therefore, dual tasks - as a verification of the fulfillment of the need and with sufficient conditions for the solvability of inclusion (6),

30 - a set of methods for deriving consequences of the form KGM from the System of relations j: / ,,,, (7j

The dual task: writing down the necessary and sufficient conditions that the inclusion is a consequence of the system (7).

Let - be a complement to the set M. Then either 0 - 0 and then / / 7 (/ У, but then, by virtue of

We have the same corollary: M: iiM (J

3 In any case, the matter boils down to sufficient and sufficient

conditions of the emptiness of the intersection of a finite number of -. For convex sets, a real linear HopMirooBaHjjeiE space is considered. Let / / - m be open and convex sets that are nonempty and do not coincide with the whole space, and - be an arbitrary convex rzzoszet, nonempty and incomplete in b. Let {(I) Then there exist open half-spaces,, ..., У, such that / И P; ..mj. (9)

where And i P is the closure of the set P, and f f

A specific information is stored in the knowledge base regarding the corresponding subject - facts, cisors, characteristics, practical rules and conditions. MLV - 2 addresses the knowledge base and, on the basis of the data stored in it, develops advice, recommendation or proposed solution to the task. ES contains a learning tool - block

  training 6, explanatory block 8, which informs the user about what knowledge, rules and the course of reasoning led to the data, but it turned out that the system asked the user some specific question. An ES contains all subsystems with its interfaces, providing: communication with the various functional blocks of the complex.

The basis of the ES is the knowledge base, which is a set of facts related to the competence of that expert or another expert, a set of rules, and the SH gives conclusions based on the analysis of combinations of these rules. For the formation and debugging of the knowledge base, the following information is required:

-name of all cells corresponding to the variables being expanded

- questions that allow you to receive information from the user to clarify the value of ka; by which variable

-information about the relationships (dependencies) me; ed intermediate variables and outputs of variables

- Lots of training in two examples that provide network tuning. ES is built in the form of a connected network and performs mellow tasks;

-realization of conclusions based on incomplete information,

-search for unknown input variables necessary to implement additional conclusions,

- justification of the findings.

In the methods for explaining the results of the formal conclusion, two approaches are used:

- an explanation using the generation of a chain of rules of type §; if ..., then ... that connect the initial data with the results of conclusions,

Demonstration to the user of the chain of structures used to obtain a specific conclusion.

-obtaining input values E moment T ((

- calculation of output values using EO based on weight coefficients, (cells II-I3),

- obtaining the correct output .SCHE of inputs related to the previous time G-s ((c {non-urgent value characterizing the time delay. for data input-input),

-performing one iteration of the learning algorithm (with a given example), ensuring the correctness of the corresponding perceptron weights. At the same time, the main advantage of the system is observed: the possibility of on-line processing of the physical training tools appeared in terms of taking into account all the noisy noises.

A network of autonomously functioning cell elements is connected by directional / 12 connections arcs / w cd .., where are indicated - input cells 14 intermediate II-I3 located higher than cell 9. Each cavity has a weight that characterizes the influence of the cell Wj on i / i (positive values correspond to excitation negative to inhibition). Weights determine the behavior of the network as a whole. Each cell carries out activation - a numerical value as its output; all cells implement the same activation algorithm. .Activation of any cell is determined by the activation of cells directly associated with it and corresponding to the scales. The inputs and activation of the cells take the values I, O, - I. The activity of an arbitrary cell Ui (with the exception of the inputs) is calculated according to shory1 glam:

 1-)

.y

+ I if i. ABOUT

- 1, if: $ С О О, if Sc О

The training is connected with the threat of a live network by presenting to it ody iaiochiye primerye - chained pairs of input-output.

Cells of type (I4-I3-I2) - logical. Its elements, iteration of the network operation otoot from the recalculation (As a result of activation) of the activity values of all network cells. Such a network model is a linear discriminant network.

To overcome the limitations inherent in the persentron, it’s easy, turning to the use of systems where between the inputs of the output there is a certain number of layers of hidden 12,13,14 elements. The learning algorithm for such a generalized perceptron is the error propagation method.

The reviewing system consists of layers, the elements 14 of the first layer are INPUT1 - they are subjected to J.I. images to be analyzed. The last layer II is the output, the image / 0 will be removed from it; . input 14 and output II are a layer of hidden elements. 3, the system has connections only between the element belonging to the adjacent layer 171; the number of elements in kaichdom

the layer may be different.

When the signal X arrives at the input of the Lth element in the layer with number 1g, this element generates an output signal according to the law Y; Г i eicp / / J (I The specific type of dependence. M is not very significant. It is necessary that when the signal changes) (from -eo to -do, the value of the signal v monotonically increases from O to I. This is to simplify the calculations.

In the FIRST turn, the signal arriving at element j in a common, h-fl layer is formed as a weighted sum of output signals from elements of the previous Pt layer:

 j; V.

I am presented with IlSoopasieH e, i.e. 1 j.

The output signal of the elements of the last layer gives the result of processing,

I / g

T.E. J i j

If the weighting coefficients of the bonds are random, the reaction of such a system will be random - the image formed at its output. The task of training is to achieve the desired response to the presented image by selecting the weighting coefficients of the bonds.

ys; w available M different reference images I / for each of which the desired reaction is known output image And // | . You need to choose the values

(J d)}

weight co-customers | Jyi V and the number of progladstvennyh layers, which when applying to the input of the system any of the reference images, the corresponding picture we set appeared at its output. Define 1u1 error as

 zA - ,,, /. 0)

/ And / j

where SO // M i //. w are output, 1 element signals

V / j f

the last layer II when applying to the first 14 layer of the reference image j 1 |, gi | . This function reaches its minimum value only if, for all M reference images, the result

Processing is the same as desired. The purpose of training is to find

t)

values of the coefficients minimizing the drying of the function (E) (0); learning turns out to be a simple optimization problem.

To minimize E by the fastest descent method, it is necessary to calculate the partial derivatives of E with respect to the weight coefficient of the repentant connection. Any such derivative will be given by the sum of the values relating to the various reference images. Separately for each reference image, all partial derivatives of the Error function are calculated by the back propagation method.

at 1 7

Claims (1)

A connected expert system (ES), including a knowledge base (KB), a database (VD), a processor with dynamic storage device connected to a logical inference machine (MLV), including an interpreter and selector, an explanation block, a unit for organizing user interaction, differing by the fact that a duality check block is introduced in the MLV, and the base of values is made in the form of a pertron-based neural network connected to the MLV selector, the input of the duality check block is connected to the MLV selector, its output is connected to the interpreter m MLV.