AI approaches

How to think in a human way: an approach based on cognitive modeling

Before stating that a particular program thinks like a person, it is necessary to have some way of determining how people think. It is necessary to penetrate into the actual process of the work of the human mind itself. Two ways can be used for this: introspection (an attempt to follow the course of one’s own thoughts) and psychological experiments.

Only after creating a sufficiently accurate theory of thinking will it be possible to present the formulas of this theory in the form of a computer program. And if the input and output data of the program, as well as the distribution of the actions performed by it over time, exactly correspond to the behavior of a person, this may indicate that some of the mechanisms of this program can also act in the human brain.

For example, Allen Newell and Herbert Simon, who developed the GPS program (General Problem Solver), did not seek to ensure that this program correctly solved the tasks. They were more concerned that the recording of the stages of the reasoning carried out by it coincided with the recording of the reasoning of people solving the same problems.

In the interdisciplinary field of cognitive science, computer models taken from artificial intelligence and experimental methods taken from psychology are jointly used to develop accurate and valid theories of the human brain.

A field of knowledge such as cognitive science is quite fascinating and so extensive that a separate encyclopedia may well be devoted to it. However, real scientific cognitive science must necessarily be based on an experimental study of real people or animals.

At the initial stages of the development of artificial intelligence, there was often confusion between the approaches described above, for example, sometimes some programmers had to deal with such statements that the algorithm they proposed copes well with a particular task and therefore is a good model of human abilities, or vice versa.

Modern engineers present the results of their research in these two areas separately; This separation allows for the development of both artificial intelligence and cognitive science faster.

But these two scientific areas continue to enrich each other, especially in areas such as visual perception and understanding of natural language. Recently, particularly significant progress has been achieved in the field of visual perception through the use of an integrated approach, in which both neurophysiological experimental data and computational models are used.

How to think rationally: an approach based on the use of "laws of thinking"

The Greek philosopher Aristotle was one of the first to try to define the laws of "correct thinking", i.e. processes of formation of irrefutable reasoning. His syllogisms became a model for creating proof procedures that always allow one to come to the right conclusions if the right prerequisites are given, for example, “Socrates is a man; all men are mortal; therefore Socrates is mortal. " These studies were based on the assumption that such laws of thought govern the work of the mind; on the basis of them developed a scientific direction, called logic.

In the 19th century, scientists working in the field of logic created an exact system of logical notation for statements about objects of any kind that are found in the world and about the relations between them. (Compare it with the usual system of arithmetic notation, which is intended mainly to form statements about equality and inequality of numbers.)

By 1965, programs had already been developed that could, in principle, solve any solvable problem described in the system of logical symbols. Researchers in the field of artificial intelligence, adhering to the so-called traditions of logicism, hope that they will be able to create intelligent systems based on such programs.

But in the implementation of this approach, there are two serious obstacles. First, it is rather difficult to take any informal knowledge and express it in the formal terms required for a system of logical symbols, especially if this knowledge is not completely reliable. Secondly, the possibility of comparatively easy to solve the problem “in principle” does not mean at all that it can really be done in practice.

Even such tasks, which are based on several dozens of facts, can exhaust the computing resources of any computer, unless certain methods are used to control which stages of the reasoning should be tested in the first place. Although we encounter both of these obstacles in any attempt to create computing systems to automate the process of reasoning, they were first discovered within the framework of the traditions of logicism.

How to think rationally: a rational agent based approach

An agent is considered everything that acts (the word agent comes from the Latin word agere - to act). But it is assumed that computer agents have some other attributes that distinguish them from ordinary “programs”, such as the ability to function under autonomous control, to perceive their environment, to exist for a long period of time, to adapt to changes and to have the ability to take on the achievement of goals. put by others.

A rational agent is an agent who acts in such a way that one can achieve the best result or, in the face of uncertainty, the best expected result.

In the approach to the creation of artificial intelligence based on the “laws of thinking,” emphasis was placed on the formation of correct logical conclusions. Of course, sometimes the formation of correct logical conclusions becomes part of the functioning of a rational agent, since one of the ways to rationalize your actions is to rationally conclude that this particular action allows you to achieve these goals, and then act in accordance with the decision . On the other hand, the correct logical conclusion does not exhaust the concepts of rationality, since situations often arise in which it is impossible to unequivocally choose any correct actions, but you still need to do something. In addition, there are ways of rational organization of actions in respect of which it cannot be argued that they use logical inference.

For example, withdrawing a finger from a hot stove is a reflex action that is usually more successful than a slower move made after careful consideration of all the circumstances.

Thus, all the skills required to pass the Turing test, also allow for rational action. So, first of all, it is necessary to be able to represent knowledge and reason on the basis of it, because it allows us to develop acceptable solutions in various situations.

It is necessary to have the ability to form understandable sentences in a natural language, since only those who are able to express their thoughts correctly are accepted into a complex society. It is necessary to learn not only for the sake of acquiring erudition, but also due to the fact that a better idea of ​​how the world works allows us to develop more effective strategies for action in this world. You need to have the ability to visual perception, not only because the process of visual observation allows you to have fun, but also because vision tells you what can be achieved with a certain action, for example, whoever can see the tidbit faster than anyone, will get a chance to get to it before the others.

For these reasons, the approach to the study of artificial intelligence as the field of designing rational agents has at least two advantages. Firstly, this approach is more general than the one based on the use of the “laws of thought”, since the correct inference is simply one of several possible mechanisms for achieving rationality. Secondly, it is more promising for scientific development in comparison with approaches based on the study of human behavior or human thinking, since the standard of rationality is clearly defined and fully generalized. Human behavior, on the other hand, is well adapted to only one specific environment and is partly a product of a complex and largely unexplored evolutionary process, which, as it turned out, does not allow to form creatures that are ideal in all respects.