You get a bonus - 1 coin for daily activity. Now you have 1 coin

Knowledge representation in intelligent systems

Lecture



Foreword
Data and knowledge. Basic definitions.
Features of knowledge. Transition from Database to Knowledge Base.
Models of knowledge representation. Informal (semantic) models.
Formal models of knowledge representation.

Foreword

Currently, in the research on artificial intelligence (AI), six areas stand out:

  1. Representation of knowledge.
  2. The manipulation of knowledge.
  3. Communication
  4. Perception.
  5. Training.
  6. Behavior.

Within the framework of the "Knowledge Representation" direction, the tasks associated with the formalization and representation of knowledge in the memory of an intellectual system (IP) are solved. For this purpose, special models of knowledge representation and languages ​​for describing knowledge are developed, various types of knowledge are highlighted. Sources are studied from which IP can draw knowledge, and procedures and techniques are created by which knowledge can be acquired for IP. The problem of knowledge representation for IP is extremely relevant, since IP is a system whose operation relies on knowledge of the problem area that is stored in her memory.

Data and knowledge. Basic definitions.

The information with which a computer is dealt is divided into procedural and declarative . Procedural information is embodied in programs that are executed in the process of solving problems, declarative information is in the data with which these programs work. The standard form of information representation in a computer is a computer word consisting of the number of binary digits defined for a given type of computer - bits . A machine word for data representation and a machine word for representing commands that make up a program may have the same or different numbers of digits. Recently, computer words with the same number of digits have been used to represent data and commands. However, in some cases, machine words are divided into groups of eight bits, which are called bytes .

The same number of digits in machine words for commands and data makes it possible to treat them in a computer as identical information units and to perform operations on commands as data. The contents of the memory form the information base .

In most existing computers it is possible to extract information from any subset of computer word bits up to one bit. In many computers, it is possible to combine two or more computer words into words with longer lengths. However, the machine word is the main characteristic of the information base, since its length is such that each machine word is stored in one standard memory cell, provided with an individual name — the cell address. This name is used to extract information units from the computer memory and to write them into it.

In parallel with the development of the structure of a computer, the development of information structures for data presentation took place. There are ways to describe the data in the form of vectors and matrices, list structures have appeared, hierarchical structures. Currently, high level programming languages ​​use abstract data types whose structure is defined by the programmer. The emergence of databases (DB) marked a further step towards organizing work with declarative information. Databases can simultaneously store large amounts of information, and special tools that make up a database management system (DBMS) allow you to effectively manipulate data, if necessary, extract it from the database and write it in the correct order to the database.

With the development of research in the field of IP, a concept of knowledge emerged that combined many features of procedural and declarative information.

In a computer, knowledge , like data , is displayed in symbolic form - in the form of formulas, text, files, information files, etc. Therefore, it can be said that knowledge is a specially organized data. But that would be too narrow an understanding. Meanwhile, in AI systems, knowledge is the main object of formation, processing and research. The knowledge base , along with the database, is a necessary component of the AI ​​software package. Machines that implement AI algorithms are called knowledge-based machines, and a subsection of the theory of AI, associated with the construction of expert systems, is knowledge engineering.

Features of knowledge. Transition from Database to Knowledge Base.

  Features of knowledge:

  • Internal Interpretability . Each information unit must have a unique name by which the IP finds it, and also responds to requests in which this name is mentioned. When the data stored in the memory were deprived of names, there was no possibility of their identification by the system. The data could be identified only by a program that retrieves them from memory at the direction of the programmer who wrote the program. What is behind this or that binary code of a machine word was unknown to the system.

Table 1.1

Surname

Year of birth

Specialty

Seniority, number of years

Popov

1965

Locksmith

five

Sidorov

1946

Turner

20

Ivanov

1925

Turner

thirty

Petrov

1937

Plumber

25

If, for example, in memory of the computer it was necessary to record information about the employees of the institution, presented in Table. 1.1, without an internal interpretation, a set of four computer words corresponding to the rows of this table would be entered into the computer memory. Moreover, information about which groups of binary digits in these machine words encoded information about specialists is absent. They are known only to the programmer who uses the data table. 1.1 to solve his problems. The system is not able to answer questions like "What do you know about Petrov?" or "Is there a plumber among the specialists?"

In the transition to knowledge, information about a certain proto-structure is entered into the computer memory.   information units . In this example, it is a special machine word in which it is indicated in which categories the information about surnames, years of birth, specialties and experience are stored. In this case, special dictionaries should be given, which list the surnames, the year of birth, the specialty, and the length of experience available in the system. All these attributes can play the role of names for those machine words that correspond to the rows of the table. By them you can search for the necessary information. Each row of the table will be an instance of the proto-structure. Currently, DBMSs provide the implementation of the internal interpretability of all information units stored in the database.

  • Structured . Information units should have a flexible structure. For them, the "matryoshka principle" should be followed, i.e. recursive embeddability of some information units into others. Each information unit can be included in any other, and from each information unit you can select some of its information units. In other words, there should be the possibility of arbitrarily establishing relations between separate information units of the type "part - whole", "gender - kind" or "element - class".
  • Connectivity The information base between information units should provide for the possibility of establishing links of various types. First of all, these links can characterize the relationship between information units. Relationship semantics may be declarative or procedural. For example, two or more information units can be connected by a "at the same time" relationship, two information units - by a "cause-effect" relationship, or by a "be near" relationship. These relationships characterize declarative knowledge. If an argument-function relationship is established between two information units, then it characterizes procedural knowledge associated with the calculation of certain functions. Next, we distinguish between structuring relations, functional relations, causal relations, and semantic relations . With the help of the first, hierarchies of information units are set, the second carry procedural information allowing to find (calculate) some information units through others, the third set cause-effect relationships, the fourth correspond to all other relationships.

Other links can be established between information units, for example, determining the order in which information units are selected from memory or indicating that two information units are incompatible with each other in the same description.

These three features of knowledge allow you to enter a general model of knowledge representation, which can be called a semantic network , which is a hierarchical network, at the vertices of which there are information units. These units are provided with individual names. The semantic network arcs correspond to different connections between information units. In this case, hierarchical relationships are determined by structuring relations, and non-hierarchical relationships, by relations of other types.

  • Semantic metric . On a set of information units, in some cases it is useful to set a relation characterizing the situational proximity of information units, i.e. the power of associative communication between information units. It could be called the relevance relation for information units. Such an attitude makes it possible to highlight in the information base some typical situations (for example, “purchase”, “traffic control at an intersection”). The relevance relation when working with information units allows you to find knowledge close to those already found.
  • Activity Since the advent of the computer and the division of the information units used in it into data and commands, a situation has arisen in which the data is passive and the commands are active. All processes occurring in a computer are initiated by commands, and the data is used by these commands only when necessary. For IP this situation is not acceptable. Like a person, in the IP the actualization of one or another action is promoted by the knowledge available in the system. Thus, the execution of programs in the IS should be initiated by the current state of the information base. The appearance in the database of facts or descriptions of events, establishing links can be a source of system activity.

These five features of the information units define the boundary beyond which the data are transformed into knowledge, and the databases develop into knowledge bases (BR). The set of tools that work with knowledge forms the knowledge management system (SUBZ). Currently, there are no knowledge bases in which internal interpretability, structuring, connectivity, semantic measure are introduced and knowledge activity is fully implemented.

Models of knowledge representation. Informal (semantic) models.

There are two types of knowledge representation methods (PPs):

  1. Formal models of PZ;
  2. Informal (semantic, relational) models PZ.

Obviously, all methods of knowledge representation that are discussed above, including products (this is a system of rules on which the production model of knowledge representation is based), belong to informal models. Unlike the formal models, which are based on a rigorous mathematical theory, informal models do not adhere to such a theory. Each informal model is suitable only for a specific subject area and therefore does not have the versatility that is inherent in formal models. The logical conclusion - the main operation in the AIS - in formal systems is strict and correct, since it is subject to strict axiomatic rules. Inference in informal systems is largely determined by the researcher himself, who is responsible for its correctness.

Each of the methods of the PP corresponds to its own way of describing knowledge.

  1. Logical models. The basis of models of this type is a formal system defined by a quadruple of the form: M = <T, P, A, B> . The set T is a set of basic elements of different nature, for example, words from a certain limited vocabulary, details of a children's designer, which are part of a certain set, etc. It is important that for the set T there is some way to determine whether an arbitrary element belongs to this set. The procedure of such a check can be any, but in a finite number of steps it should give a positive or negative answer to the question whether x is an element of the set T. We denote this procedure by P ( T ).

The set P is a set of syntactic rules . With their help, the elements of T form syntactically correct sets . For example, syntactically correct phrases are built from the words of a limited vocabulary, new constructions are assembled from the details of the children's designer with the help of nuts and bolts. The existence of the procedure P ( P ) is declared , with the help of which, in a finite number of steps, one can get an answer to the question whether the set X is syntactically correct.

In the set of syntactically correct sets, a subset of A is distinguished . Elements A are called axioms . As for the other components of the formal system, there must be a procedure P ( A ), with which for any syntactically correct population one can get an answer to the question of its belonging to the set A.

Set B is a set of inference rules . By applying them to elements A , you can get new syntactically correct sets, to which you can again apply the rules from B. This is how a set of populations derived in this formal system is formed. If there is a procedure P ( B ), with the help of which it is possible to determine for any syntactically correct set, whether it is derivable, then the corresponding formal system is called solvable . This shows that it is the inference rule that is the most complex component of the formal system.

For knowledge included in the knowledge base, we can assume that the set A is formed by all information units that have been entered into the knowledge base from outside, and with the help of the rules for deriving new knowledge from them. In other words, the formal system is a generator of the generation of new knowledge, forming a set of knowledge derived in this system. This property of logical models makes them attractive for use in knowledge bases. It allows you to store in the database only the knowledge that forms the set A , and all other knowledge is obtained from them according to the rules of inference.

2. Network models . The basis of models of this type is the construction, previously called the semantic network. Network models can be formally defined as H = <I, C 1 , C 2 , ..., C n , G > . Here I have a lot of information units; C 1 , C 2 , ..., C n - a set of types of connections between information units. The mapping Γ specifies between information units included in I , links from a given set of link types.

Depending on the types of connections used in the model, there are classifying networks, functional networks and scenarios . Classification networks use structuring relationships. Such networks allow knowledge bases to introduce different hierarchical relationships between information units. Functional networks are characterized by the presence of functional relationships. They are often called computational models , because they allow one to describe the procedures for “computing” some information units through others. The scenarios use causal relations, as well as relations of the types "means - result", "instrument - action", etc. If different types of connections are allowed in the network model, then it is usually called the semantic network.

3. Production models . In models of this type some elements of logical and network models are used. The idea of ​​inference rules, here called products , is borrowed from logical models, and the description of knowledge in the form of a semantic network is borrowed from network models. As a result of applying the rules of inference to fragments of a network description, the semantic network is transformed by changing its fragments, expanding the network and eliminating unnecessary fragments from it. Thus, in production models, procedural information is explicitly identified and described by other means than declarative information. Instead of the logical conclusion, characteristic for logical models, in the production models there appears a conclusion on knowledge .

4. Frame models . Unlike other types of models, the rigid structure of information units, which is called a proto- frame, is fixed in frame models. In general, it looks like this:

(Frame Name:

Slot 1 Name (Slot 1 Value )

Slot name 2 (value of slot 2)

. . . . . . . . . . . . . . . . . . . . . .

Slot name K (slot value K)).

The value of a slot can be almost anything (numbers or mathematical relationships, texts in natural language or programs, rules for output or links to other slots of a given frame or other frames). The slot value can be a set of lower-level slots, which allows to implement the "matryoshka principle" in frame representations.

When a frame is instantiated, specific names are assigned to it and the slots, and the slots are filled. Thus, from the protoframes we get frames - instances. The transition from the original protoframe to the frame - the instance can be multi-step, due to the gradual refinement of the values ​​of the slots.

For example, the structure of the table. 1.1, recorded as a protoframe, has the form

(List of employees:

Surname (meaning slot 1);

Year of birth (value of slot 2);

Specialty (value of slot 3);

Experience (value of slot 4)).

If the values ​​of the slots to use the data table. 1.1, the frame is an instance

(List of employees:

Surname (Popov - Sidorov - Ivanov - Petrov);

Year of birth (1965 - 1946 - 1925 - 1937);

Specialty (mechanic - turner - turner - plumber);

Experience (5 - 20 - 30 - 25)).

Connections between frames are specified by the values ​​of a special slot named "Communication".Some IP specialists believe that there is no need to specifically distinguish frame models in the knowledge representation, since they combine all the main features of the models of other types.

Formal models of knowledge representation.

Система ИИ в определенном смысле моделирует интеллектуальную деятельность человека и, в частности, - логику его рассуждений. В грубо упрощенной форме наши логические построения при этом сводятся к следующей схеме: из одной или нескольких посылок (которые считаются истинными) следует сделать "логически верное" заключение (вывод, следствие). Очевидно, для этого необходимо, чтобы и посылки, и заключение были представлены на понятном языке, адекватно отражающем предметную область, в которой проводится вывод. В обычной жизни это наш естественный язык общения, в математике, например, это язык определенных формул и т.п. Наличие же   языка предполагает, во - первых, наличие алфавита (словаря), отображающего в символьной форме весь набор базовых понятий (элементов), с которыми придется иметь дело и, во - вторых, набор синтаксических правил, на основе которых, пользуясь алфавитом, можно   построить определенные выражения.

Логические выражения, построенные в данном языке, могут быть истинными или ложными. Некоторые из этих выражений, являющиеся всегда истинными. Объявляются аксиомами (или постулатами ). Они составляют ту базовую систему посылок, исходя из которой и пользуясь определенными правилами вывода, можно получить заключения в виде новых выражений, также являющихся истинными.

Если перечисленные условия выполняются, то говорят, что система удовлетворяет требованиям формальной теории. Ее так и называют формальной системой (ФС). Система, построенная на основе формальной теории, называется также аксиоматической системой.

Формальная теория должна, таким образом, удовлетворять следующему определению:

всякая формальная теория F = (A, V, W, R) , определяющая некоторую аксиоматическую систему, характеризуется:

наличием алфавита (словаря), A ,

множеством синтаксических правил, V ,

множеством аксиом, лежащих в основе теории, W ,

множеством правил вывода, R .

Исчисление высказываний (ИВ) и исчисление предикатов (ИП) являются классическими примерами аксиоматических систем. Эти ФС хорошо исследованы и имеют прекрасно разработанные модели логического вывода - главной метапроцедуры в интеллектуальных системах. Поэтому все, что может и гарантирует каждая из этих систем, гарантируется и для прикладных ФС как моделей конкретных предметных областей. В частности, это гарантии непротиворечивости вывода, алгоритмической разрешимости (для исчисления высказываний) и полуразрешимости (для исчислений предикатов первого порядка).

ФС имеют и недостатки, которые заставляют искать иные формы представления. Главный недостаток - это "закрытость" ФС, их негибкость. Модификация и расширение здесь всегда связаны с перестройкой всей ФС, что для практических систем сложно и трудоемко. В них очень сложно учитывать происходящие изменения. Поэтому ФС как модели представления знаний используются в тех предметных областях, которые хорошо локализуются и мало зависят от внешних факторов.


Comments


To leave a comment
If you have any suggestion, idea, thanks or comment, feel free to write. We really value feedback and are glad to hear your opinion.
To reply

Presentation and use of knowledge

Terms: Presentation and use of knowledge