Is it possible to program randomness?

What is the difference between a person and a program

Neural networks, which now make up almost the entire field of artificial intelligence, can take into account much more factors than people in decision-making, do it faster and in most cases more accurately. But programs only work as they have been programmed or trained. They can be very complex, take into account many factors and act very variably. But still they cannot replace a person in decision making. What is the difference between a person and such a program? Here, 3 key differences should be noted, from which all the others follow:

1. A person has a picture of the world, which allows him, in terms of information, to supplement the picture with data that are not prescribed in the program. In addition, the picture of the world is structurally structured so that it allows us to have at least some idea of ​​everything. Even if it is something round and glows in the sky (UFO). Usually, ontologies are built for this, but ontologies do not have such completeness, poorly take into account the ambiguity of concepts, their mutual influence, and so far apply only to strictly limited topics.
2. A person has a logic that takes into account this picture of the world, which we call common sense or common sense. Any statement has meaning, and takes into account hidden undeclared knowledge. Despite the fact that the laws of logic are many hundreds of years old, nobody still knows how the ordinary, non-mathematical, reasoning logic functions. We essentially do not know how to program even ordinary syllogisms.
3. Arbitrariness. Programs do not have randomness. This is perhaps the most difficult of all three differences. What do we call randomness? The ability to build new behavior that is different from what we performed under the same circumstances earlier, or to build behavior in new circumstances not previously encountered. That is, in essence, this is the creation on the go of a new program of behavior without trial and error, taking into account new, including internal, circumstances.

Arbitrariness is still an unknown field for researchers. Genetic algorithms that can generate a new program of behavior for intelligent agents are not an option, since they generate a solution not logically, but by means of “mutations” and the solution is found “randomly” during the selection of these mutations, that is, by trial and error. A person finds a solution right away, building it logically. A person can even explain why such a solution is chosen. The genetic algorithm has no arguments.



It is known that the higher the animal on the evolutionary ladder, the more arbitrary its behavior can be. And it manifests the greatest arbitrariness precisely in a person, since a person has the ability to take into account not only external circumstances and his learned skills, but also hidden circumstances - personal motives, previously reported information, the results of actions in similar circumstances. This greatly increases the variability of human behavior, and, in my opinion, it is consciousness that is involved. But more on that later.

Consciousness and Arbitrariness

And then consciousness? In the psychology of behavior, it is known that we carry out habitual actions automatically, mechanically, that is, without the participation of consciousness. This is a remarkable fact, meaning that consciousness is involved in the creation of new behavior, associated with orientational behavior. It also means that consciousness is connected precisely when it is necessary to change the usual pattern of behavior, for example, to respond to new requests taking into account new opportunities. Also, some scientists, for example, Dawkins or Metzinger, pointed out that consciousness is somehow connected with the presence of a self-image in people, that the model of the world includes the model of the subject itself. What then should the system itself look like that would have such arbitrariness? What structure should be possessed so that it can build new behavior to solve the problem in accordance with new circumstances.



To do this, we first need to recall and clarify some well-known facts. All animals with a nervous system, one way or another, contain in it a model of the environment integrated with the arsenal of their possible actions in it. That is, this is not only a model of the environment as some scientists write, but a model of possible behavior in a particular situation. And at the same time it is a model for predicting changes in the environment in response to any actions of the animal. Cognitive scientists do not always take this into account, although open mirror neurons of the premotor cortex directly indicate this, as well as studies of the activation of macaque neurons, in response to the perception of banana in which not only the banana region in the visual and temporal cortex is activated, but also the hands in the somatosensory that the banana model is directly connected to the hand, since the monkey is interested only in that fruit that it can take it and eat it. We simply forget that the nervous system did not appear to reflect the world of animals. They are not sophists, they just want to eat, so their model is more a model of behavior, not a reflection of the environment.



Such a model already has a certain degree of arbitrariness, which is expressed in the variability of behavior in similar circumstances. That is, animals have some arsenal of possible actions that they can carry out depending on the situation. It can be more complex temporary patterns (conditioned reflexes) than a direct reaction to events. But still, this is not completely arbitrary behavior, which allows us to train animals, but not humans.



And here there is an important circumstance that we need to take into account - the more well-known circumstances met, the less varied the behavior, since the brain has a solution. Conversely, the newer the circumstances, the more options for possible behavior. And the whole question is their selection and combination. Animals do this simply by displaying the entire arsenal of their possible actions, as Skiner had shown in his experiments.



This is not to say that arbitrary behavior is completely new; it consists of previously learned patterns of behavior. This is their recombination, initiated by new circumstances that do not fully coincide with those circumstances for which there is already a ready-made pattern. And this is precisely the point of separation of arbitrary and machine behavior.

Randomness modeling

The creation of a program of arbitrary behavior, able to take into account new circumstances, would make it possible to create a universal “program of everything” (by analogy with the “theory of everything”), at least for a specific task domain.



What could make their behavior more arbitrary, free? My experiments showed that the only way out is to have a second model that models the first and can change it, that is, act not with the environment as the first, but with the first model to change it.



The first model responds to environmental circumstances. And if the pattern activated by her turned out to be new, the second model is called, which is taught to look for solutions in the first model, recognizing all possible behaviors in the new environment. Let me remind you that in the new environment, more behaviors are activated, so the question is precisely in their selection or combination. This is because, in contrast to the familiar environment, in response to new circumstances, not one behavior pattern is activated, but several at once.



Each time the brain meets with something new, it carries out not one, but two acts - recognition of the situation in the first model and recognition of already implemented or possible actions by the second model. And in this structure there are many opportunities similar to consciousness.

1. This two-act structure allows you to take into account not only external, but also internal factors - in the second model, the results of the previous action, distant motives of the subject, etc. can be remembered and recognized.
2. Such a system can build new behavior right away, without long training initiated by the environment according to evolutionary theory. For example, the second model has the ability to transfer solutions from one submodel of the first model to its other parts and many other metamodel capabilities.
3. A distinctive property of consciousness is the presence of knowledge about its action, or autobiographical memory, as shown in article (1). The proposed two-act structure just possesses this ability - the second model can store data about the actions of the first (no model can store data about its own actions, since for this it should contain consistent models of its actions, and not environmental reactions).

But how exactly is the construction of a new behavior in the two-act structure of consciousness? We do not have at our disposal the brain, and even its plausible model. We began to experiment with verb frames as prototypes of models that are contained in our brains. A frame is a set of verb actant options for describing a situation, and a combination of frames can serve to describe complex behavior. Frames for describing situations are frames of the first model, a frame for describing their actions in it is a frame of the second model with verbs of personal actions. We often have them mixed, because even one sentence is a mixture of several acts of recognition and action (speech act). And the very construction of long speech expressions is the best example of arbitrary behavior.



When the first model of the system recognizes a new pattern for which it does not have a programmed response, it calls the second model. The second model collects the activated frames of the first and searches for a shorter path in the graph of connected frames, which in the best way “closes” the patterns of the new situation with a combination of frames. This is a rather complicated operation and we have not yet achieved a result claiming to be the “program of everything”, but the first successes are encouraging.



Experimental studies of consciousness by modeling and comparing software solutions with psychology data provides interesting material for further research and allows us to test some hypotheses that are poorly tested in experiments on people. This can be called modeling experiments. And this is only the first result in this direction of research.



Bibliography



1. Two-act structure of reflective consciousness, A. Khomyakov, Academia.edu, 2019 .