AI methods

Many types of human mental activity, such as writing programs for a computer , doing math, reasoning at the level of common sense, and even driving a car, require “intelligence”. Over the past decades, several computer systems have been built that can perform similar tasks.

There are systems that can diagnose diseases, plan the synthesis of complex synthetic compounds, solve differential equations in symbolic form, analyze electronic circuits, understand the limited amount of human speech and the natural language text. We can say that such systems have to some extent artificial intelligence. Work on the construction of such systems is carried out in the area, called artificial intelligence (AI). When implementing intellectual functions, information called knowledge is by all means present. In other words, intelligent systems are at the same time knowledge processing systems. Currently, six main areas have emerged in artificial intelligence research.

1. Representation of knowledge. Within this direction, the tasks associated with the formalization and representation of knowledge in the memory of the AI ​​system are solved. For this purpose, special models of knowledge representation and knowledge description languages ​​are being developed, various types of knowledge are being introduced. The problem of knowledge representation is one of the main problems for the AI ​​system, since the functioning of such a system relies on knowledge of the problem area that is stored in its memory.
2. The manipulation of knowledge. In order for knowledge to be used when solving a problem, it is necessary to teach the AI ​​system to operate with them. Within this framework, ways of developing knowledge based on incomplete descriptions are being developed, methods of reliable and plausible inference based on existing knowledge are being created, models of reasoning based on knowledge and imitating features of human reasoning are proposed. The manipulation of knowledge is very closely connected with the representation of knowledge, and these two directions can only be divided conditionally.
3. Communication The range of tasks in this area includes: the problem of understanding and synthesis of coherent texts in natural language, understanding and synthesis of speech, the theory of communication models between man and the AI ​​system. Based on research in this direction, methods of constructing linguistic processes, question-answer systems, dialogue systems and other AI systems are being formed, the purpose of which is to provide a comfortable environment for a person to communicate with the AI ​​system.
4. Perception. This direction includes the development of methods for presenting information about visual images in the knowledge base, creating methods of transition from visual scenes to their textual description and methods of reverse transition, creating means for generating visual scenes based on internal representations in IP systems.
5. Training. To develop the ability of AI systems to learn , i.e. To solve problems that they have never met before, methods are being developed to formulate the conditions of tasks by describing a problem situation or to monitor it, methods for moving from a well-known solution of particular problems (examples) to solving a general problem, creating methods for decomposing the original problem into smaller and already known for AI systems. In this direction, the AI ​​has done very little.
6. Behavior. Since AI systems must operate in a certain environment, it is necessary to develop some behavioral procedures that would allow them to interact adequately with the environment, other AI systems, and people. This direction in AI is developed very poorly.

This tutorial covers only the basics of the first two areas - knowledge representation and inference methods.

Knowledge and their presentation

Knowledge-based systems are software systems, the main structural elements of which are the knowledge base and the mechanism of inference. First of all, they include expert systems that can diagnose diseases, evaluate potential mineral deposits, process natural language, recognize speech and images, etc. Expert systems are the first step in the practical implementation of research in the field of AI. Currently they are already used in industry.

An expert system is a computational system that includes the knowledge of specialists about a particular problem area and, within this area, is able to make expert decisions. The structural elements that make up the system perform the following functions. Knowledge Base - implements the functions of knowledge representation in a specific subject area and their management. The mechanism of logical conclusions - performs logical conclusions based on the knowledge available in the knowledge base. The user interface is necessary for the correct transmission of answers to the user, otherwise it is extremely inconvenient to use the system.

The module of knowledge acquisition is necessary to obtain knowledge from an expert, support the knowledge base and supplement it if necessary. The module of answers and explanations - forms the conclusion of the expert system and presents various comments attached to the conclusion, and also explains the reasons for the conclusion.

The listed structural elements are the most characteristic, although in real expert systems their functions can be appropriately strengthened or expanded.

Knowledge in the knowledge base is presented in a specific form and the organization of the knowledge base allows them to be easily identified, modified and added. Solving problems using inference based on knowledge stored in a knowledge base is implemented by an autonomous inference mechanism. Although both of these components of the system are independent from the point of view of its structure, they are closely related to each other and the definition of the knowledge representation model imposes restrictions on the choice of an appropriate mechanism for logical conclusions. Thus, when designing expert systems, it is necessary to analyze both of these components. To manipulate real-world knowledge using a computer, it is necessary to simulate it. The main models of knowledge representation include:

• logical models ;
• production models;
• network models;
• frame models.

Before proceeding to a more detailed examination of the typical models of knowledge representation in AI systems, let us get acquainted with each of them.

Logic models

The basis of models of this type is the concept of a formal system. The formulation and solution of any problem is associated with a specific subject area. So, solving the task of scheduling the processing of parts on machine tools, we involve in the subject area such objects as specific machines, parts, time intervals and the general concepts of "machine", "part", "machine type", etc.

All subjects and events that form the basis of a common understanding of the information necessary for solving a problem are called the subject area. Mentally, the domain is represented as consisting of real objects called entities. The entities of the domain are in a certain relationship to each other. Relationships between entities are expressed through judgment. In the language (formal or natural) sentences respond.

Languages ​​intended for describing subject areas are called knowledge representation languages. The universal language of knowledge representation is natural language. However, the use of natural language in systems of machine representation of knowledge encounters a number of obstacles, the main of which is the lack of formal semantics of natural language. In order to represent mathematical knowledge in mathematical logic, logical formalisms are used - calculus of statements and predicate calculus. These formalisms have clear formal semantics and inference mechanisms have been developed for them. Therefore, the predicate calculus was the first logical language that was used for the formal description of subject areas related to the solution of applied problems.

Subject domain descriptions made in logical languages ​​are called logical models. Logic models built using logic programming languages ​​are widely used in knowledge bases and expert systems.

Production models

Products (along with network models) are the most popular means of knowledge representation in AI systems. In general, products are understood as an expression of the form A 'B. The usual reading of a product is as follows: IF A, TH is B. Implication can be interpreted in the usual logical sense as a logical sign of B from true A. Other interpretations of the product are possible, for example, A describes some the condition necessary to be able to perform act B.

If a set of products is stored in the system's memory, then they form a production system. In the production system, special product management procedures should be set, with the help of which the products are updated and one or another product is updated.

The production system includes a rule base (products), a global database and a management system. The rule base is a memory area that contains a body of knowledge in the form of rules of the form IF - THEN. A global database is a memory area containing factual data (facts). The management system generates conclusions using a rule base and database. There are two ways of forming conclusions - direct conclusions and reverse conclusions.

In direct outputs, one of the data elements contained in the database is selected, and if this element is matched with the left part of the rule (sending) during matching, the corresponding conclusion is derived from the rule and placed in the database or the action determined by the rule is executed and accordingly changes the contents of the database. In the inverse of the process begins on the goal. If this goal is consistent with the right side of the rule (conclusion), then the premise of the rule is taken as a sub-goal or hypothesis. This process is repeated until a sub-match with the data is obtained. With a large number of productions in the production model, it becomes more difficult to check the consistency of the production system, i.e. sets of rules. Therefore, the number of products with which modern AI systems work, as a rule, does not exceed thousands.

Network models

The basis of models of this type is the construction, previously called the semantic network. The semantic approach to the construction of AI systems is used in systems for understanding natural language, in question-answer systems, in various subject-oriented systems.

In the most general case, the semantic network is an information model of the domain and has the form of a graph, the vertices of which correspond to the objects of the domain, and the arcs correspond to the relations between them. Arcs can be defined by different methods depending on the type of knowledge presented. Typically, arcs used to represent a hierarchy include arrays of the type set, subset, and element. Semantic networks used to describe natural languages ​​use arcs of the type “agent”, “object”, “recipient”. In semantic networks, it is possible to represent knowledge in a more natural and structured way than in other formalisms.

Frame models

Unlike other types of models, the rigid models of information units called frames are fixed in frame models. A frame is a form of representing a situation, which can (or should be) described by a certain set of concepts and entities. The frame name is assigned as the frame identifier. This name must be the only one in the entire frame system.

The frame has a certain internal structure consisting of a set of elements, called slots, which are also assigned names. Each slot in turn is represented by a certain data structure. Specific information relating to the object described by this frame is substituted into the slot value.

A slot value can be almost anything: numbers, formulas, natural language texts or programs, output rules, 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.