After describing an intelligent inference system's static structure (Chapter 3) and dynamic process (Chapter 4), in this chapter the focus becomes the interaction between the system and its environment. Different from instinctive systems, the behavior of an intelligent system is determined both by its innate features and its experience.
Section 5.1. Types of experience
The experience of an information system is the history of its interaction with its environment, described as information transferring. In this sense, every information system has experience.
For a system like NARS, all types of experience are eventually represented as Narsese sentences, obtained by the execution of certain actions, either triggered by internal events or external events.
All these actions are implemented by mechanisms below the information transferring level of description, as physical, chemical, biological, or other processes. They are usually referred to as the sensorimotor mechanism.
The sensorimotor mechanism on the external environment is usually covered by notions like sensation, perception, and motion; the sensorimotor mechanism on the internal environment is usually covered by notions like self-awareness and self-control.
Though all interactions among intelligent systems are carried out by the sensorimotor mechanisms of the systems involved, the systems usually develop a signal mechanism to transfer complex information at the level of concepts by convention. This special type of interaction is called "communication", and the signal mechanism is called "language".
All these types of experience provide an origin of the beliefs of the system, and decide the derived goals and composed actions to be formed by the system.
Section 5.2. Sensorimotor mechanism
Every information system, intelligent or not, interacts with its environment. In this sense, it has sensorimotor mechanism, though it may be completely different from human sensorimotor mechanism.
A system may have several sensory channels, each can recognize a certain type of signal (light, sound, ...). In each channel, the recognizable signals are determined by the system's hardware/wetware, and usually remain unchanged after the system is born/implemented. To extend this kind of capability, the system can use tools.
With the coming of each recognizable signal, an action of the system is triggered to generate an internal representation of the signal. The underlying process of such an action is specific to the concrete (physical/chemical/...) property of the signal, though the overall cause and effect can be described abstractly as information transferring.
Since the system's beliefs and concepts eventually depend on the available actions, systems with different sensorimotor mechanisms may perceive the same environment in different ways, and therefore form different world view. On the other hand, they may still have very similar "intelligence", that is, how the experience is processed and used. Given the difference between human body and computer hardware, we should not expect an AI system to have identical beliefs and concepts, therefore behavior, as a typical human being.
Since low-level perception (in sensorimotor) and high-level perception (in categorization) basically face the same problem, and work under the same restriction, we can expect them to follow similar principles, though the details of the processing may be very different.
An intelligent system needs to learn about when an action can be executed, and what effects it will have. This learning is usually achieved through a sensorimotor feedback loop: when an action is executed, its observed effect and its expected effect are compared, so as to revise the system's beliefs about the action. On the other hand, the sensory capacity of the system usually depends on the motor capacity, because many observations require the execution of certain actions. Therefore, sensorimotor should be treated as one mechanism, with a sensation aspect and a motion aspect.
Section 5.3. Self-awareness and self-control
When talking about the "environment" of an information system, it should include the "internal environment". The sensory ability represents internal states, structures, and their change as beliefs about self, and the motor ability carry out self-control. As with outside environment, the system's knowledge and control on its internal structure and activity are limited and selective.
Inside-oriented sensorimotor mechanism follows the same principles as outside-oriented sensorimotor mechanism, though the two use different sensors and motors. Consequently, the system develops different concepts and beliefs when describing internal and external events, which is where the "mind-body problem" starts. Since the internal events are only observable to the system itself, their descriptions are inevitably from a first-person point of view. On the contrary, the external events happen in the shared environment, so can be described from a third-person point of view.
Since not all internal events are represented in the system's beliefs, we can distinguish consciously sensed/controlled internal processes from "automatic" processes. The former can be manipulated by the system's information processing mechanism, while the latter cannot. This is also where the "Self and Other" distinction comes --- some events are perceivable while the others are not; some events can be controlled while the others cannot.
Self-consciousness is developed in advanced intelligent systems for complicated adaptive behaviors. It is not something that comes as "additional" or "optional" to those behaviors. Some AI systems will be self-conscious, but because of its intrinsic "first-person" nature, we cannot directly perceive it, but have to recognize it in the system's behaviors.
Section 5.4. Communication and language
Communication happens between two information systems, and provides shared experience for them, as well as increases their capabilities via cooperation, in which one system can serve as sensor, processor, or effector of the other, in a way similar to tool usage.
Communication happens in a language, which is a sign system associated to the concepts of the systems. Though language comprehension and production are supported by sensorimotor, the conventional nature of language allows the systems to interact with each other at conceptual level, and to ignore the details of sensorimotor. Using language, systems can directly describe beliefs, goals, and actions. A language usually provides approximate many-to-many mappings between signs in the language and concepts in the systems, and the mapping is established in history by an evolving convention.
Communication is a goal-directed process between two or more information systems, though their goals for the process may not be the same. For a communication to be successful, the signals involved should correspond to similar concepts in all the systems, though "perfect mutual understanding" is usually impossible. Similar to sensorimotor, the language comprehension/production ability of a system is highly language-specific, and is acquired from language-specific experience.
A secondary function of language, derived from the function of communication, is representation. By setting up an external and materialized version of beliefs and concepts, it is much more efficient for a system to organize and process its knowledge. Language makes the content of thought become its object that can be explicitly manipulated.
Language usage presumes a categorization and inference mechanism. Historically, language capability starts at pragmatics, since communication is goal-directed activity in the systems participated. The stable conventions on the word-concept relation formed in communication becomes semantics. Finally, syntax and grammar appear to express complicated semantic structure. Language processing happens at these three levels simultaneously. Language acquisition and processing are carried out by the same mechanism responsible for intelligence and cognition in general.
Section 5.5. Socialization
As soon as two or more intelligent systems begin to communicate with each other, they start to have common experience, which will shape their beliefs, goals, and actions. In the long run, a system's behavior will be strongly influenced by the society it lives in, via such a socialization process.
Though living in a society is not a precondition for intelligence, social experience consists of a major part of experience for advanced intelligent systems. If an intelligent system only gets its knowledge from its own sensorimotor mechanism, its capability will be highly restricted.
The common beliefs accepted by most of the members of the society at the current time provides an "objective world view". To an individual system, a large part of this common knowledge is directly accepted, and becomes the system's beliefs. In places where common knowledge conflicts with personal beliefs, the result is usually a compromise.
As a special case, language-related conventions are the common knowledge of a community of users of a given language. To effectively communicate with the others, an individual must follow the common usage of the language. On the other hand, given the special experience and need, violations of the common usage are inevitable, which are the forces behind language evolution.
Socialization not only provides the knowledge of the system, but also regulates the development of the goal structure of individuals. A system will obtain reward or punishment during socialization, depends on the compatibility of its goals and the goals of the other system. Morality and ethics knowledge is also acquired in this process.
Furthermore, socialization extends the system's available action set, by allowing individuals to participate in social cooperation.
Section 5.6. Education
Education, or training, is a special type of socialization. In this process the system being educated or trained is provided with a predetermined partial experience, to get a desired result, as the obtained goals, beliefs, and actions of the system. It is a semi-compulsory socialization.
For a society, education is an efficient way to pass certain experience to new members in the society. It is necessary for a social system to adapt to its environment, and to keep its internal consistency, though very often various biases are spreaded in this process, too.
With the born of truly intelligent computer systems, "education of AI" will become a necessary step. Actually, this is the stage where domain-specific requirements are taken into consideration, which should not be hard-wired into the system. Unlike human mind, for an AI system it is possible to "implant" structure into it, but it cannot completely replace education.
The education of AI will to a large extent follows the same principles and procedures of human education. Just load a huge amount of "common sense facts" into a system is not the right way to educate it, because a proper knowledge structure should also include knowledge about the relative priority among the beliefs, as well as the related questions and goals that make the beliefs useful to the system.
An important part of education is to make sure that an AI system will properly co-exist with human beings and other AI systems. What makes things complicated is that for an AI system, its behavior not only depends on its design and education, but also depends on its "personal experience", which usually cannot be accurately predicated. Therefore, its "possible behavior space" of a system is determined by the three factors:
- how the system is designed,
- how the system is educated before it is "on its own",
- what possible situation the system may face.