This chapter defines intelligent system as adaptive information system that works with insufficient knowledge and resources. This working definition is accepted, because it properly covers several types of intelligent systems, as well as excludes the ones that are usually considered as lacking intelligence.
Section 2.1. Intelligence vs. instinct
Information systems can be divided into "intelligent systems" and "instinctive systems". The basic difference is that intelligent systems are adaptive, while instinctive systems are not.
In an instinctive system, all of its beliefs, or knowledge, take the form of instinct, innate relations that link goals to actions. The system behaves by reacting to current needs or signals in predetermined manner, and the experience of the system has little impact on the stimulus-response connection. Consequently, the capabilities of an instinctive system remains constant over time. Most of animals and traditional computer systems can be put into this category.
In an intelligent system, most of its beliefs, or knowledge, are summarized from the system's experience, as the result of adaptation. Even if there are still innate stimulus-response connections, they are typically modifiable by the system itself. Consequently, the capabilities of an intelligent system changes, and normally increases, over time. Most human beings and some computer systems can be put into this category.
Please note that, defined in this way, an intelligent system is not necessarily more capable than an instinctive system. Intelligence does not indicate capability, but change in capability. Whether (or how much) a system is intelligent is not determined by what the system can do, but what it can learn to do.
Section 2.2. Intelligence as adaptivity
For an adaptive system, the attempts to increase capability are made according to experience, therefore there is no guarantee that the attempts will be successful, because future situation may be different from past situation.
Adaptation will improve the system's capability only when the future is indeed similar to the past. It means that adaptive systems only live better in relatively stable environments, where things may change, but not too much. In environments where changes happens randomly (that is, no pattern or law can be recognized), all systems are equally bad, adaptive or not.
Designing an adaptive system is quite different from designing an instinctive system. In the latter case, first the designer need to identify the goal of the system, then build necessary ability and knowledge that is sufficient to achieve the goal. The basic requirement for the design is correctness, that is, the system is indeed able to achieve the goal with the given knowledge and ability. If the goal changes, it is the responsibility of the designer to adjust the knowledge and ability of the system accordingly.
To design an adaptive system is another story. Since by definition the designer has no way to provide sufficient knowledge and ability for the system to fully achieve its goal, what the designer build into the system is not only knowledge and ability to achieve concrete goals, but knowledge and ability that allow the system to obtain new knowledge and ability by itself from its experience (and therefore better satisfy its goal).
Also, an adaptive system is evaluated differently. To check if a system is truly adaptive, we should not check whether (or to what extent) its goal is achieved, but to check whether (or to what extent) its goal will be achieved if the future situation is consistent with the system's past experience.
Being adaptive is closely related to being able to work with insufficient knowledge and resources. A system with sufficient knowledge and resources (with respect to its goals) has no need to adapt; a non-adaptive system treats its knowledge and resources as sufficient, for all practical considerations.
Section 2.3. Properties of intelligent systems
An intelligent system often needs to deal with goals for which it has insufficient knowledge and resources. These goals may indicate the direction of behaviors, not concrete internal or external states that can be reached or kept. What the system does is to derive achievable goals according to its knowledge, then achieve them. In this way, the unachievable goals are partially achieved. An implication of this situation is that an intelligent system usually has no "final state" where it stops itself.
A traditional computer system, as an instinctive system, has a well-defined "solvable tasks" set, and anything outside it should be rejected. Also, each concrete solvable task will take a constant amount of computational time and space resources, so that a task whose resource request goes beyond the system's supply cannot be solved. Since an intelligent system adjusts its beliefs and actions all the time, what the system can do and what the system cannot do no longer have a clear distinction. When facing a novel task, sometimes neither an observer nor the system itself has a sure prediction on whether it can be done before the system actually work on it. Furthermore, what counts as a "solution" becomes fuzzy. For a question that no sure answer can be obtained, a tentative answer is better than no answer --- the correctness of the answer cannot be guaranteed anyway. Related to this, the system's resource spent on a task is not a constant anymore.
While an instinctive system has the same actions all the time, an intelligent system can develop new actions, either by composing existing actions into macro actions, which can be directly invoked altogether, or by using external tools, which gives an action novel effects. The usefulness of a new operation need to be determined by the system's experience. With insufficient resources, relatively useless actions will be forgot.
In an intelligent system, beliefs are constantly revised according to the new experience of the system, to eventually record whether a certain goal is achieved by a certain action. However, to summarize complicated experience with limited resources, many beliefs no longer directly link goals to actions, but link goals to derived goals, actions to composed actions, or beliefs to other beliefs.
Section 2.4. Human and artificial intelligence
The above working definition fits the situation of human beings very well. Human beings are clearly adaptive, and have to work with insufficient knowledge and resources.
This working definition does not describe what is called "AI" in the current research community, but it may be exactly why this research has failed to meet people's expectations. To most people, "intelligence" means problem solving capability, and the goal of AI is to make computer systems to solve the problems solvable by human only at the moment. Though this opinion is intuitive, and leads to useful results, to establish AI as a branch of science on this foundation may be a bad idea.
Since all human beings start with similar innate capability, their capability at a certain age is highly correlated to their learning ability, and that is why the notion of IQ makes some sense. However, since a computer system can be built with various levels of innate capability, what the system can do at a moment and what it can learn at the time have little correlation.
To measure intelligence by problem solving capability effectively identify AI with computer science. Though practically fruitful, this approach will never give AI an identity, nor will it reduce the huge difference between human and computer at the current time, which is on adaptivity or leaning ability, not on problem-solving capability.
On the other hand, artificial intelligence is different from artificial human intelligence or artificial person, meaning that the goal should not be accurately duplicate human brain, mind, or behavior, but be identical to human at a more abstract level, that is, the information processing/transferring level.
Section 2.5. Other forms of intelligence
The above working definition of intelligence can be applied to other fields like animal intelligence, alien intelligence, and community intelligence.
When words like "intelligent" or "smart" are used on animals, it is almost always about their acquired ability, not on their innate ability. There are animals with complicated and highly efficient innate abilities, but they are usually not called "intelligent" for that reason. Furthermore, whether, or how much, an animal is intelligent is usually evaluated with respects to its own capability, not that of a human being.
Though we have not confirmed the existence of any alien intelligence, the notion is accepted by many people as a possibility. However, when we encounter an alien object, how can we decide whether, or how much, it is intelligent? Obviously, we will try to generalize some problem-solving capability from its stimulus-response patterns, but we clearly cannot make the judgment according to whether they are similar to human problem-solving capability, or according to how complicated these capabilities look to us. Instead, it will be judged according to whether these patterns are adaptive or not.
The topic of community intelligence will be addressed in Chapter 6.
All these fields support the opinion that "intelligence", as an ability abstracted from multiple fields, is about whether, or how much, a system can improve its problem-solving capability by adaptation, that is, learning from its past experience.
Section 2.6. What is not intelligent
If we really want to use the concept "intelligence" as a generalization of concepts include "human intelligence", "artificial intelligence", "animal intelligence", "alien intelligence", "community intelligence", and so on, and build a scientific theory around this concept, we cannot let it cover all kinds of things.
Intelligence does not means identical to human mind in all aspects. Instead, "human intelligence" and "artificial intelligence" become identical only after the content of goals, actions, and beliefs of the systems are omitted, and what is left are the relationship among them.
Intelligence is not identified by a special class of problem-solving capability. Intelligent systems may have completely different problem-solving capabilities, though their ability to acquire these capabilities is similar to each other.
Intelligence does not means always finding the correct or optimal solution among all possible ones. Every concrete adaptive system is restricted by its experience and resources. To assume otherwise leads to completely different models.
Though each of the alternative understanding of intelligence has theoretical and practical value, they lead the research to fundamentally different directions and results.