Knowledge acquisition

Works in the field of knowledge acquisition by intellectual systems have been and remain the most important direction of the theory and practice of artificial intelligence. The goal of this work is to create methodologies, technologies and software tools for transferring knowledge (or, as they sometimes say, competence) into the system’s knowledge base. At the same time, experts (ie, highly qualified subject matter specialists), texts and data, for example, stored in databases, act as sources of knowledge. Accordingly, various methods of acquiring knowledge are being developed.

Machine learning in the world pays great attention. There are many machine learning algorithms, among the most common are algorithms of class C4. One of the algorithms of this class, C4.5, is essentially a decomposition algorithm and builds a decision tree. The initial information for building this tree is a lot of examples. The most frequent (at the current step) frequency class of examples is associated with each vertex of the tree. In the next step, this principle is recursively applied to the current node, i.e. Many examples related to the current vertex are also divided into subclasses. The algorithm terminates its work either when a certain criterion is satisfied or when the subclasses are exhausted (if they are specified).

Actively investigating methods of teaching causes of action. Sometimes they talk about the so-called theory of action, referring to the situational calculus in the spirit of John McCarthy. In this theory, the causes of actions and the actions themselves are described in the form of clausal structures (one of the types of such structures is an implication, the left part of which is the conjunction of atomic formulas, and the right part consists of one atomic formula).

Further, the methods of inductive logic programming are modified in such a way as to be applicable to the search for such structures. When such structures are found, they can be used in logic languages ​​for reasoning about actions and their causes.

Many works of this direction are devoted to the "neural paradigm". The neural network approach is used in a huge number of tasks - for clustering information from the Internet, automatically generating local directories, representing images (in recursive neural networks). Among the topics that have been actively studied recently are inhomogeneous neural models with similarity relations. (Heterogeneous Neural Networks with similarity relation).

This similarity relation is determined on a set of inputs and a set of network states, and the measure of similarity is the scalar product of vectors or the Euclidean distance (where one vector is the vector of inputs, and the other is the weight distribution of neurons describing the current situation).

Works on the automatic generation of hypotheses are mainly connected with the formalization of plausible reasoning, the search for dependencies of a causal type between certain entities. Examples include the generation of hypotheses about the properties of chemical compounds (prediction of biological activities), possible causes of defects (diagnostics), etc.

Intelligent data analysis and image processing

This is a relatively new direction, the basis of which consists of two procedures: the discovery of patterns in the initial information and the use of the patterns found for prediction (prediction). These include the task of selecting informative data from a large set of them, choosing the informative characteristics of an object from a wider set of its characteristics, the problem of constructing a model, allowing to calculate the values ​​of selected informative characteristics from the values ​​of other characteristics, etc.

A significant part of this direction consists of research on various aspects of image recognition, in particular, using neural networks (including pseudo-optical neural networks). Methods for recognizing sequences of video images based on a declarative approach and extracting semantically meaningful information are being studied. To the same direction belong research on graphical programming technology on the Internet.

Multi-agent systems, dynamic intelligent systems and planning

This is a new (however, in theoretical, behavioral aspects — rather, a well-forgotten old) direction that studies intellectual software agents and their teams. An intelligent agent is a software system that has:

• autonomy: agents act without direct human participation and can, within certain limits, manage their actions themselves;
• social traits: agents interact with other agents (and, possibly, a person) through some language of communication;
• reactivity: agents perceive the environment, which can be the physical world, many other agents, the Internet or a combination of all of this, and react to its changes;
• activity: agents can demonstrate purposeful behavior, while showing initiative.

The main tasks in this area are as follows: the implementation of negotiations of intelligent agents and the development of languages ​​for this purpose, the coordination of agent behavior, the development of an agent programming language architecture.

It should be emphasized that agency technologies appeared about 6-7 years ago. During this time, interest in these technologies has moved from the field of academic research to the sphere of commercial and industrial applications, and the ideas and methods of agent technology have very quickly migrated from artificial intelligence to the practice of software development and other computational disciplines.

Behavior planning, or AI — planning is the ability of an intelligent system to synthesize a sequence of actions to achieve a desired target state. Work on the creation of effective methods for such a synthesis is in demand and has been actively conducted for about 30 years. Planning is the basis of intelligent control, that is, automatic control of autonomous, targeted behavior of software and hardware systems.

Among the methods of AI-planning today distinguished classical planning static environment planning, dynamic planning, i.e. planning in a changing environment and, most importantly, accounting for such a change, hierarchical planning, that is, when the actions of the high-level abstract plan are specified by more detailed plans of the lower level, in part - ordered (or monotonous) planning, when the plan is based on partially ordered sets of subplans. At the same time, the general plan (of which the elements are subplans) must be monotonous, and each of the subplans may be non-monotonous. I will add that monotony is such a property of the plan, when each of its actions reduces the differences between the current state and the goal of the behavior. For example, if a robot’s plan for moving to a goal is such that every step it takes approaches a goal, then the plan is monotonous, but if he stumbles upon an obstacle and needs to be circumvented, then the plan’s monotony will be broken. However, if the plan for avoiding an obstacle is singled out in a separate subplan and regarded as an element of the original plan, then the monotony of the latter will be restored.

Work is also actively carried out in the area of ​​recognizing plans, building schedulers and expanding their capabilities, heuristic planning with resource constraints, managing planning through time logic, planning using graphs.

We consider approaches to planning, in which the construction of current plans is carried out continuously for each state of the system in real time. For this, continuous monitoring of the control object is provided.

Natural language processing, user interface and user models

This direction is connected with the development of speech communication support systems, c problem solving of query refinement in information systems, with the problems of text segmentation by thematic topics, with the tasks of managing the dialogue, with the tasks of natural language analysis using various heuristics. This also includes the problems of discourse (sometimes by discourse we mean the totality of speech acts, together with their results).

Still relevant to learning contextual analysis of the text, the task of acquiring knowledge of intellectual systems and extracting information from the texts.

The most important task in the process of extracting information, as well as in the process of acquiring knowledge, is minimizing the role of the expert - participant in the process.

The importance of this direction cannot be underestimated. The reason for this is an increase in the flow of textual information, an existing social order to search for relevant information on the Internet, to analyze textual information, to extract data from texts. Thus, the value of automatic text analysis methods will further increase.

The subject of research is also a dynamic user modeling, in particular, in e-commerce systems, the development of a frame approach to present user requests, an adaptive interface, monitoring and analysis of consumer behavior on the Internet.

Fuzzy models and soft calculations.

This direction is represented by fuzzy “inference by analogy” schemes, a look at the theory of fuzzy measures from probabilistic positions, a fuzzy representation by analytical models for describing geometric objects, evolutionary modeling algorithms with dynamic parameters, such as lifetime and population size, methods for solving optimization problems using technologies of genetic search, homeostatic and synergistic principles and elements of self-organization.

Tool development

This is an extensive field of activity within AI, which has the following objectives:

• creating knowledge acquisition software for automated transfer of competence to knowledge bases. At the same time, not only “direct” carriers can act as sources of such competence - experts in various fields, but also textual materials - from textbooks to protocols, as well as, of course, databases (implicit sources of knowledge). Verbalization, that is, the translation of such sources into an explicit form is the content of methods for detecting knowledge in data, including various teaching methods for examples (including the preprocessing of large data arrays for further analysis);
• implementation of knowledge base software.
• implementation of software support design of intelligent systems. A set of such tools usually contains a text editor, a concept editor, a conceptual model editor, a model library, a system for acquiring knowledge from experts, examples training tools, and a number of other modules.

Perspectives of artificial intelligence

Today we can distinguish a number of areas in artificial intelligence, which in the quite foreseeable future may lead to qualitative changes in technology and technology. Here I intend to present my view on some of them.

Apparently, precedent-based reasoning (CBR) is one of the most promising approaches in artificial intelligence, the introduction of which will lead to significant progress in a number of areas, and a breakthrough in this direction should be expected in the next 3-5 years. Some of the existing and expected CBR applications have been listed above.

Arguments about space are not very new, but now rapidly developing area of ​​artificial intelligence, is of increasing practical importance in connection with the work on the creation of autonomous mobile devices, image analysis (in particular, aerial photographs), the synthesis of textual descriptions of images.

Apparently, using the methods of machine learning and the automatic formation of hypotheses, it will be possible to solve a number of useful practical problems - from detecting patterns in data to increasing the degree of adaptability and the "intelligence level" of various technical devices.

Approaches based on the technology of intelligent agents should be recognized as one of the most promising in the development of large software products, including management tools for large and complex systems (such as telecommunications systems, distributed production, control systems for troops, transport, networks, distributed information retrieval) . It is also possible that this direction will disappear as a separate scientific discipline, dissolving in adjacent areas, but having a decisive technological impact on them.

One should expect an increasing influence of ideas and methods (AI) on computer-aided text analysis (AT) in natural language. This effect is likely to affect semantic analysis and the associated syntactic analysis methods - in this area it will manifest itself in taking into account the model of the world at the final stages of semantic analysis and using knowledge of the subject area and situational information to reduce redundancy at earlier stages (for example, when building parse trees).

The second “communication channel” of AI and AT is the use of machine learning methods in AT; the third “channel” is the use of reasoning based on precedents and reasoning based on argumentation for solving some AT problems, for example, noise reduction and increasing search relevance.

One of the most important and promising areas in artificial intelligence today should include the tasks of automatic behavior planning. The field of application of automatic planning methods is a variety of devices with a high degree of autonomy and purposeful behavior, from household appliances to unmanned spacecraft for deep space research.