AGI: INSTINCT OF CURIOSITY

RISK vs CRAVING FOR SOMETHING NEW

Any autonomous system requires some resources to maintain its activity. In other words, there is a need for these resources. And if the system is autonomous, it must independently take care of the timely replenishment of resources. In turn, this means that the system must be able to measure the current amount of each resource and consider these values when deciding what actions to take. For example, animals need water, food, and rest; if the available level is below a certain threshold, the situation is characterized as thirst, hunger, and fatigue.

As the nervous system develops in the process of evolution and the brain appears as a fully-fledged separate component capable of accumulating information/knowledge, a new type of resource appears that is necessary for the functioning of a particular power of intelligence: information. Information acquires the status of another resource that must be somehow obtained. Accordingly, the brain's ability to measure the amount of information received arises, and the instinct of curiosity is a mechanism for replenishing the deficit of information required for rational behavior in constantly changing conditions. The specificity of the need for new information is that the negative effect of its deficiency (information hunger) usually has a much greater time lag than other needs.

In technical autonomous systems, everything is similar. Still, the set of required resources may, naturally, differ from what is observed in the animal world - gasoline or electricity instead of food, and so on. However, information as a resource is required by an intelligent autonomous technical system in exactly the same way as animals. And just as in nature, the more developed the intellect, the more knowledge is required to realize its intellectual potential.

There are two ways to obtain new information. Passive one (from the point of view of information needs) is reduced to using information obtained to satisfy other needs. In the process of any activity, new situations for the system arise, which are used to replenish/update knowledge, that is, to "on the side" satisfy information hunger. An alternative to this is an active search for new information, that is, certain actions aimed specifically at learning something new.

A natural feature of the active (exploratory) method of obtaining new information is increased risk. If risk is present, then the level of risk is higher, the situation is less known, and new information - including about risk factors - can be obtained. The instinct of curiosity (as a synonym for the need for information) is therefore inextricably linked with the level of risk: satisfying curiosity is associated with an increase in the chances of getting into a dangerous situation, and vice versa, avoiding everything unknown reduces risks by reducing the opportunity to learn something new (and thereby creates a situation of information hunger for one's own intellect).

In nature, an innate mechanism of regulation by a compromise between the level of risk and the intensity of learning associated with age has been developed: while the offspring is under the care of parents, the acceptable level of risk is high, and when the time comes to become parents, the parameter of the level of acceptable risk decreases - suppression of the instinct of curiosity to one degree or another is observed. In technical autonomous intelligent systems, an analog of this can be the presence of two operating modes: conservative, in which actions aimed at exploring unknown places/states/actions are prohibited, and exploratory, in which this is allowed.

In any case, an autonomous intelligent system needs new information and requires ways to measure the amount of new knowledge, assess the degree of risk, and consider the need to satisfy curiosity when making decisions.