Directions in AI

Knowledge representation (knowledge representation) is one of the most established areas of artificial intelligence. Traditionally, it included the development of formal languages ​​and software for displaying and describing the so-called cognitive structures. Today, studies on descriptive logic, the logic of space and time, ontologies also rank as a representation of knowledge.

Spatial logic allows you to describe the configuration of spatial areas, objects in space; families of spatial relationships are also being studied. Recently, this area, due to its close connection with applied problems, has become dominant in knowledge representation studies. For example, for tasks of robotics, it is important to be able to restore its verbal (formal) description from the image of a scene, in order to use this description further, for example, to plan the actions of the robot.

The objects of descriptive logic are the so-called concepts (basic structures for describing objects in expert systems) and connected into a whole set of concepts (aggregated objects). Descriptive logic generates methods for working with such complex concepts, reasoning about their properties and their derivability. Descriptive logic can also be used to build the explanatory component of the knowledge base.

Finally, ontological studies are devoted to ways of conceptualizing knowledge and methodological considerations about the development of tools for knowledge analysis.

Different ways of representing knowledge underlie the modeling of reasoning, which includes: case-based reasoning modeling (CBR), reasoning or constraints, modeling reasoning with uncertainty, reasoning about actions and changes, non-monotonic reasoning models, etc. Let's stop on some of them.

CBR

Here, the main problems are the search for adaptation algorithms, the “focusing of search” on the use of past experience, a conclusion based on the assessment of similarities and visualization technology.

Let the precedents be given as a set of pairs <CASE, SOLUTION>, a set of dependencies between different attributes of CASES and SOLUTIONS, as well as a target problem AIM. For the emerging new situation (“new case”), it is required to find a pair of <NEW CASE, FIXED SOLUTION>, which solves the target problem.

Algorithms for such problems are usually based on comparison of precedents with a new case (in some metric), using dependencies between case attributes and solution attributes. Such dependencies can be set by a person when constructing a database of cases, or automatically detected in the database of cases. When searching for a solution for a target problem, an adaptation of the solution already existing in the database of use cases is performed. For this adaptation, the indicated dependencies are used.

An important CBR problem is the problem of choosing a suitable precedent. It is natural to look for a suitable precedent in the area of ​​the search space where solutions to similar problems are found. But how to determine which solutions should be considered similar?

There are hypotheses that the similarity of problems imposes restrictions on the similarity of the corresponding hypotheses in the form of a weak connection between them. This circumstance is used to limit the search for solutions.

Let, for example, we are talking about a certain client who (with his wife) wishes to spend two weeks on the Canaries and wants to pay for it no more than one and a half thousand dollars.

The first version of the dialogue with the system is not capable of reasoning (and on the basis of precedents, too):

Client: Hello, I would like to go on vacation for two weeks in July to the Canary Islands. I would like to go with my wife, but I cannot pay more than one and a half thousand dollars.

• System: sorry, but we have no such opportunity.
• Client: good, but maybe you can find something similar in a close region.
• System: Could you please clarify what you mean by talking about a close region.
• Client: I mean the coast of Spain.
• System: sorry, but this is not a close region. It is more than a thousand kilometers from the Canary Islands!
• Client: but the climate is similar ...
• System: sorry, what do you mean, speaking about the climate ...

Most likely, the client will contact another travel agency. The second dialogue option:

• Client: Hello, I would like to go on vacation for two weeks in July to the Canary Islands. I would like to go with my wife, but I cannot pay more than one and a half thousand dollars.
• System: sorry, there is no such possibility now. But maybe you are satisfied with the coast of Spain?
• Client: What about climate, does it look like Canary?
• System: Yes, and besides, tickets are cheaper!
• Client: great, book please.

From this, perhaps, not a very serious example, it is clear that, a) in the second case, the client deals with a system that understands that even though the region is not close to Kanaram, it is closer to the client, in any case, the ticket is cheaper and b) since the client wants to go in July, apparently, he is interested in sea bathing and beaches, and this is both on the Canaries and on the coast of Spain. From this, the system concludes about the similarity of situations (described by the client and the existing vacancy in its base) and, on the basis of this, assumes that the decision “holidays in Spain” is a close solution to “holidays in the Canaries”.

The CBR methods are already used in a variety of applied tasks - in medicine, project management, for analyzing and reorganizing the environment, for developing consumer goods, taking into account the preferences of different consumer groups, etc. One should expect applications of CBR methods for the tasks of intelligent information search, e-commerce (offering goods, creating virtual trading agencies), planning behavior in dynamic environments, layout, design, program synthesis.

Modeling reasoning based on constraints

The most interesting problems here are the modeling of reasoning based on procedural dynamic constraints. They are motivated by complex actual tasks - for example, planning in a real situation.

The problem of satisfying constraints is understood as four sets: a set of variables, a set of corresponding variable domains, a set of restrictions on variables, and a set of relations over domains. The solution to the problem of satisfying constraints is a set of variable values ​​that satisfy the constraints on variables, such that, in this case, the regions to which these values ​​belong, satisfy the relations over the domains.

The task of satisfying dynamic constraints is a sequence of tasks of satisfying constraints, in which each subsequent task is a constraint of the previous one. These tasks are close to the tasks of dynamic programming. They are also associated with interval algebra.

Non-monotonic reasoning models

This includes studies on the logic of defaults (default logic), according to the logic of “canceled” (Defeasible) reasoning, program logic, theoretical-argumentative characterization of logics with cancellations, characterization of logics with preference relations, construction of equivalent sets of formulas for logics with outlines (circumscription) and some others. Such models arise in the implementation of inductive reasoning, for example, by examples; they are also associated with machine learning tasks and some other tasks. In particular, in problems of modeling reasoning on the basis of induction, the source of the initial hypotheses are examples. If a certain hypothesis H originated on the basis of N positive examples (for example, an experimental one), then no one can guarantee that the N + 1 th example refuting the hypothesis (or changing its degree of truth) will not be in the database or in the field of view of the algorithm. ). This means that revisions should be subject to all the consequences of hypothesis H.

Discourses on actions and changes

Most of the work in this area is devoted to the application of situational calculus — the formalism proposed by John McCarthy in 1968 to describe actions, reasoning about them, and the effects of actions. To transform the plan of the robot's behavior into an executable program, reaching with a certain probability of a fixed goal, a special logical calculus is introduced, based on situational logic. For this logic, the proposed implementation options for the pGOLOG language is a version of the GOLOG language containing means for introducing probabilities. Actively investigated the logic of action, the use of modal logic for reasoning about knowledge and actions.

Arguments with uncertainty

This includes the use of Bayesian formalism in rule systems and network models. Bayesian networks are a statistical method for detecting patterns in data. For this purpose, the primary information contained in the network structures or in databases is used. In this case, network structures are understood to mean the set of vertices and the relations on them, which are defined using edges. Substantially, edges are interpreted as causal relationships. Each set of vertices Z, representing all the paths between some two other vertices X and Y, corresponds to a conditional dependence between these two last vertices.

Next, a certain probability distribution on the set of variables corresponding to the vertices of this graph is given, and the resulting, but minimized (in a certain sense) network is called the Byse network.

On such a network, the so-called Bayesian output can be used, i.e. to calculate the probabilities of the consequences of events, you can use (with some stretch) the probability theory formulas.