Chapter 4. Self-Organizing Process

The working process of an intelligent inference system, like NARS, is a self-organizing process, in which all components at the object-level adapts to the experience of the system, managed by the innate mechanism at the meta-level.

Section 4.1. Learning as self-organization

For a system like NARS, its behaviors are determined by two factors: its design (i.e., built-in inference rules and control algorithms) and its experience (its history of interaction with the environment).  More concretely, the former factor determines the possibility (what the system could become), and the latter choose one possibility, and turns it into reality (what the system has become).

The self-organization of a system like NARS is driven by two fundamental conflicts within the system persistently:

1. The system has to predict future with past experience, which is always insufficient for the purpose.
2. The system has to achieve its goals with its constant resources, which is always insufficient for the purpose.

Therefore the success of the self-organization should not be judged according to the correctness or optimum of the system's behaviors, because that will need to compare the behaviors with future experience, and ignore the knowledge and resources constraints under which the behaviors are produced.  Instead, the aim of self-organization is to produce "rational behaviors" which have the best expectation to achieve the current goals according to available knowledge and resources.

Different from the common understanding in the "Machine Learning" community, learning in a system like NARS is not a computation process following an algorithm. Instead, it is an open-ended process influenced by many factors in the system. It is not an "improving routine" that is additional to the "working routine" of the system, neither. Instead, in a system like NARS, "learning" is just the long-term effect of the inference process.

Since experience by definition is different from system to system, even among the systems that co-existent in the same environment, the result of self-organization is fundamentally subjective.  What is rational to one system may be not so to another system, and rationality cannot be defined by taking the average among many systems.

Self-organization is an irreversible process.  In NARS, there is no way to "unlearn" from a given section of experience.  For a given piece of knowledge, the system can "neutralize" it by revising it with its negation, so the positive evidence and negative evidence weight the same.  However, the total amount of evidence increased in this process, so to make future revisions harder.  Even when a old belief has been forgot, it may have contributed to the derivation or revision of other beliefs in the system.

Section 4.2. The self-organization of goals

For a system like NARS, all the goals of the system are derived from a group of primary goals built into, or imposed upon, the system, and to them the system has no way to choose or to restrict, as far as they are in a format recognizable to the system.

While some of the goals can be directly satisfied or achieved by system's actions, most of them need to be achieved indirectly via the achieving of derived goals. Due to the insufficiency of knowledge and resources of the system, the derived goals may turn out to conflict with their parent goals, a phenomenon called "alienation". Though often seen as undesired, this phenomenon is also responsible for the autonomy, initiative, and originality of intelligent systems.

Also because of the insufficiency of knowledge and resources, the system cannot process goals one by one, but have to process many of them in parallel, in a time-sharing manner. At any moment, the system's decisions are usually made by taking multiple goals into consideration, rather than determined by a single goal. In the competitions among goals, the primary goals do not always win over derived goals. The system attempts to unevenly distribute resources among the goals to achieve the highest overall satisfaction.

Another important aspect of self-organization of goals is to resolve the conflicts among the goals by selecting actions based on compromise among goals. When a goal can be achieved by multiple actions, the impacts of those actions on other goals must be taken into account. 

Section 4.3. The self-organization of actions

The actions of the system at a given moment is the set of operations the system can perform on its internal structure or external environment.  The system is born with a set of primary actions, as well as operators to compose compound actions from existing ones.  The self-organization of action is mainly in the process of selectively building and maintaining compound actions.

The meaning of an action, either primary or composed, is mostly revealed by its sufficient and necessary conditions indicating the cause and effect of the action. For a system with insufficient knowledge and resources, the meaning of an action is never fully given, but gradually acquired from the experience of the system.

To be able to reason about actions allows the system to predict the effect of an action without actually execute it.  This is the "internalization" of actions, that is, simulations of the corresponding actions within the system.

If computational cost is not an issue, the system can only keep the primary actions, and combine them by inference to reach whatever goal when required.  However this is impossible, because the system must work in real time, and it does not always have the time to think before action.  On the other hand, it is also impossible for the system to try out all combinations of primary actions, and keep the useful compound actions. To use the resources efficiently, an adaptive system should use its experience to guide the self-reorganization of operations, which means to create a compound action only when it solves a existing task, to adjust the related truth-values and priority according to their past performance, and to gradually forget the compound actions that are not very useful. During such a process, the system reorganizes its actions according to the repeatedly appeared goals in its experience, so as to improve its overall efficiency.

When the system is equipped with (external) tools, similar procedure happens within the system, because with a tool, an existing action changes it meaning.  The system needs exercises and practice to use a tool skillfully.

Section 4.4. The self-organization of beliefs

In an intelligent system like NARS, the self-organization of beliefs become necessary because (1) with insufficient knowledge, the system must extend its past experience into current and future situations, and (2) with insufficient resources, the system must compress its experience, so that to use time and space efficiently.

As a result, the extreme aim of belief self-organization is not to get a "true description" of the world, but to efficiently connect the system's goals to its actions, according to past experience.  For this purpose, there are three requirements that the system should try to satisfy when organizing its beliefs:

• Correctness: The beliefs generated through self-organization should be compatible with the experience of the system. 
• Compactness: Instead of directly recording the "raw" experience, it is much more efficient to generalize them. 
• Concreteness: Beliefs should be organized in such a way that can provide concrete guidance to the system in the current and future situations.

These three requirements are independent to each other.  When there are conflicts among them, no one is logically superior to the others in general, though for a given situation one may be more important then the others.

The self-organization of beliefs are carried out by constantly generating derived beliefs from existing ones, adjusting truth-values of beliefs by combining evidence for the same statements collected in different way, and adjusting priority of beliefs according to their usefulness and relevance.

There are two co-existing tendencies in the self-organization process.  To achieve the Compactness, Concreteness, and Correctness, the system's target is to turn the knowledge structure into something as close to a pure-axiomatic system as possible, which can be called "axiomatization".   On the other hand, there is always a "de-axiomatization" process, triggered by the coming of new experience, which often challenge the knowledge structure built so far.  

Section 4.5. The self-organization of concepts

Concepts provide an intermediate level of structure between the whole memory and the individual goals, actions, and beliefs. In a system like NARS, a concept is both a storage unit to keep all directly related items (goal, action, and belief), and a processing unit for inference activities, because entities only directly interact with each other if they are semantically related by sharing a common term.

Within a concept, self-organization is responsible for the evolving of the meaning. A concept is not an internal representation of an external object (or a set of them), but an identifiable ingredient in the system's experience. A concept has no "true" or "real" meaning, but a meaning that depends on the system's experience with it. To improve efficiency, self-organization in concept attempts to give a concept a relatively clear and stable meaning, which is also part of the axiomatization process. However, with the changing of context and the coming of new experience, the process usually does not converge to a final meaning of the concept.

To improve the efficiency of summarizing experience, the inference process constantly compose new concepts new concepts, as novel ways to cluster related items. This process is not random, but data-driven, in the sense that a new term (and the related concept) is built only when it provide a preferred way to organize some experience for a certain situation. Whether this concept has long-term value is to be determined by the following experience of the system. It is a common mistake to assume concepts are created by going through all possible concepts and using a fixed criterion to harvest the desired ones.

The quality of a concept is not a matter of true or false, but good or bad. Desired properties of concepts are similar to those of beliefs:

Among existing concepts, the self-organization process evaluates and adjusts the priority distribution among them, to form a conceptual hierarchy by arranging concepts according to their inheritance relations. When the system encounters a new situation, a perception/categorization process puts it under existing concepts, which suggest possible behaviors according to the system's experience in similar situations. The triggered concepts are selected both according to their similarity to the new situation, and to their intrinsic quality.

Section 4.6. Learning and intelligence

According to the working definition of intelligence advocated in this theory, "intelligence" is not "problem-solving capability", but "learning capability". For a system like NARS, its initial problem-solving capability can be anything (determined by its innate goals, actions, and beliefs), while its "learning capability" is comparable to that of a human adult. After the system begins to communicate with its environment, its problem-solving capability usually increases as the result of self-organization, while its learning capability remains more or less the same through its life time.

In this way, the level of intelligence of a system like NARS is reflected in the expressing power of its knowledge representation language, the inferential power of its inference engine, and the resource efficiency of its management mechanism. These factors are independent to the goals, actions, beliefs, and concepts the system has at a given moment. It will be possible to define a quantitative measurement of a system's intelligence by comparing the relevant factors of a system to those in a reference class, though this measurement is only indirectly related to measurements of the system's problem-solving capability.

To increase the system's intelligence, it is possible to allow the system to learn at the meta-level, by making experience-driven changes in its representation, inference, and control mechanism, though the benefits of this type of learning may be far below what is believed by many people. Meta-level learning is highly restricted, and often dangerous.

To automatically produce systems with higher intelligence may be the job of artificial evolution, not artificial intelligence. Evolution and intelligence are two different forms of adaptation. Though the two processes share many common features, their key differences are often underestimated. Intelligence happens in individual level, and produces experience-dependent, gradual, and justified changes; evolution happens in species level, and produces experience-independent, radical, and random changes. The two processes can be combined, though not mixed, in the context of AI research, using intelligence for object-level learning, and evolution for meta-level learning.