Central Issues in A. I.
Intelligent behavior in a specific problem-solving application requires substantial knowledge of the domain in which that application functions. Further, the system’s knowledge of its domain must be in a form that is readily applicable to the full range of situations that the system is likely to encounter. Thus, one major concern of A.I. is the identification of different types of knowledge and the development of techniques for formally representing and effectively utilizing knowledge of these different types.
A second general requirement of intelligent behavior is the ability to infer conclusions from one’s knowledge. Conclusions can be necessary consequences of a given body of knowledge or they can be plausible inferences that are suggested by the body of knowledge. A. I. is concerned with mechanisms for representing the knowledge that tells us which conclusions to draw, and with mechanisms for drawing these conclusions in specific situations. The general importance of this inferential activity is highlighted by its occurrence in one important definition of meaning: we can say that system understands a given fact about the world if it understands the consequences that this fact entails in various situations. For example, to possess an understanding of the fact that the Earth exerts a significant gravitational force, a system should understand that an unsupported object will fall, that energy is required to raise the elevation of an object, etc. Knowledge that expresses classes of valid inferences tends to be powerful because of its applicability in a wide range of situations.
Most areas of significant human concern are fraught with uncertainty. This uncertainty can take many forms, such as uncertainty regarding which of many plausible conclusions are true, uncertainty regarding the likely success, the likely cost, and/or the likely benefits of alternative possible actions, and uncertainty regarding the possible relevance of different conclusions or different problem-solving approaches to the system’s current goals. A.I. is concerned with techniques for representing these uncertainties, and with techniques for intelligently evaluating and choosing among alternatives. Some example techniques that have been used in A. I. to cope with uncertainty are probability’, possibility theory12, and the use of explicit construction of hypothetical worlds”’.
Within the framework discussed so far, a typical A. I. application can be described as consisting of a body of knowledge about its domain, and an inference technique for drawing conclusions from the combination of this body of domain knowledge and knowledge about a specific situation or problem. In general, there will be many possible conclusions that the system might draw. Most of these conclusions will be either false or irrelevant to the system’s goals. For example, consider the problem of planning a trip from an office in Los Angeles to the top of the World Trade Center in New York City. An intelligent system might be expected to know a great deal about the different travel vehicles (automobiles, airplanes, walking, elevators, etc.) and their possible uses. There are many possible solutions to the stated problem, but most of them are poor in normal circumstances. For example, driving to San Francisco, flying to Boston, and then driving to New York is probably not a good idea. However, the system should know that this is a possibility, since the problem situation could be constructed to make it a good solution (e.g., a sell-out on all suitable flights out of Los Angeles to New York, etc.) In the general case, a system will be capable of recognizing many conclusions or possible actions that resulting inferior solutions, false conclusions, or irrelevant issues to its current goals. The difficult task is the one-off finding the practical, true, and relevant actions and conclusions among the large space of possibilities.
This general problem can be characterized as one of Search. In this way, search occupies a central role in intelligent behavior. A. I. is concerned with techniques for representing search spaces (the set of possibilities that need to be searched) and techniques for conducting searches effectively and efficiently. Issues connected with search include techniques for explicitly representing a system’s goals, techniques for evaluating the plausibility and efficacy of alternative possible solutions, general strategies for controlling a search and for focusing on the most fruitful possibilities, and the representation and use of specific knowledge that can be used to guide search.
A lower level view of the A.I. concepts discussed in this section can be obtained through the notion of a symbol’. A symbolism a system’s internal representation for something in the real world. For example, symbols can be used to represent abstract concepts, to represent specific physical objects, to represent relationships between objects, etc. An A.I. system’s model of the domain in which it operates is composed of a set of symbols that are structured in some appropriate manner. (One of the most important structuring paradigms for symbols is their grouping by relationships; e.g., the fact that John is a man might be represented by establishing a relationship between the symbol that represents John, the symbol that represents the set of all men, and a relationship symbol that captures the notion of membership). Operations can then be defined on these symbolic models in order to ascribes meaning to the symbols and their structuring. These operations work by modifying or extending the set of symbols and the symbol structure. Such operations can implement inferences, problem-solving strategies, general transformations on the state of the model, etc. At this low level, the fundamental issues of A.I. can be characterized as the identification of those types of structures on symbols and of operations on symbol structures that combine to yield intelligent behavior.
Types of Knowledge
The representation and application of knowledge is a central theme in all A.I. In this section, we briefly introduce some of the more important types of knowledge that have been identified together with some general remarks about their formal representation within a computer. One of the important attributes of any domain is the set of different kinds of objects and relationships that exist within that domain * For example, in the domain of ocean transportation, the object types include the notions of surface ships, submarines, hydroplanes, motor-powered vessels, sail powered vessels, and nuclear-powered vessels. An intelligent system that is reasoning within this domain must have knowledge about these different types of objects. One conclusion of A.I. is that much of the necessary knowledge can be represented in the forms of inter-relation of concepts by abstraction and of intra-description of concepts through knowledge of their attributes. For example, a sail-powered vessel is a kind of surface ship and a nuclear-powered vessel is a kind of motor-powered vessel. Thus, all properties of surface ships in general are possessed by the more restricted class of sail-powered vessels and similarly, the properties of nuclear-powered vessels include those of the more general motor-powered vessels. In A.I. parlance, we say that the concept of sail-powered vessel inherits from the concept of surface ship. The set of concepts in a domain can, in general, be organized into a conceptual taxonomy wh. Jhereeach specific concept is described in terms of its specialization with respect to more general concepts. For example, in this domain, we can introduce the concept of nuclear-powered submarine by describing it as a specialization of both submarines and of nuclear-powered vessels. The description of nuclear-powered submarine then inherits all the properties of both more abstract concepts and it need only be explicitly concerned with properties that are unique to the case of a submarine that uses nuclear power (i.e., properties not shared either by other submarines or by other nuclear-powered vessels). In general, concepts can be described in tens of their similarities and differences to other related concepts.
Abstraction organizes the set of concepts into a taxonomy. The detailed properties of the specific concepts can, to a great degree, be specified with the simple technique of identifying attributes lots of the concepts. For example, every ocean vessel has attributes that include some locomotion device, some storage capacity, some cruising range, some cruising medium, etc., combined with more general attributes that are inherited from the more abstract concept of physical object such as weight, color, shape, etc. The concept of a submarine can be specified by restricting the range of some of these attributes (e.g., the locomotion device must be an engine, the cruising medium is water to some maximum depth) and by adding some new attributes (e.g., ballast tanks). This representation of conceptual knowledge can be summarized as constructing description (called frame in A. I. jargon) that represents a protocypical instance of each concept, representing this frame by specifying its associated attributes and the ranges over which these attributes can vary, organizing these frames into a taxonomy by the abstraction relation, and finally, obtaining succinctness in the representation using inheritence on the representations of the frames derived from the abstraction relations. Conceptual taxonomies provide a good representation for the general classes of objects in a domain.
In reasoning about a problem in the domain, it is also important to represent and manipulate facts about specific objects that occur in the problem situation. To represent these facts A. I. has drawn upon techniques developed in formal logic”. Specifically, relationships are used to express facts about specific objects that are represented as symbols, propositional connectives (and, or, not, etc.) are used to build compound logical assertions, and quantifiers (for all, there exists, etc.) are used to express general properties of classes of objects. For example, the assertion “The Santa Maria is a sailing ship” could be represented by expressing a membership relation between a symbol that represented the Santa Maria and another symbol that represented the concept of “sail powered vessel”. The assertion “Hydroplanes are generally faster than surface ships” might be represented as “For a l hydroplanes h and for all surface ships s, h is probably faster than.” A. I. has frequently found the need to extend the traditional notions of formal logic to deal with issues like the representation of time and states of a process, and the fundamental notions of uncertainty (such as the “generally” in the last example).
To represent knowledge about the necessary or likely consequences of a given problem situation, one of the most m o n techniques employed in A. I. is that of “if-then” For example, the rule “IF x is a motor-powered ship that has no fuel Then is immobile” can be used in many different situations to discover properties of a particular ship. This rule could be used to detect a problem in a plan for a given ship to transport some cargo between two points in the case that the ship would run out of fuel in transit (thereupon leading to the requirement of a fueling stop). Alternatively, the rule could be used to devise a technique for immobilizing a ship (by cutting off its fuel source). The first use of the rule is an instance of forward chaining, i.e., deriving and acting upon the consequent (THEN part) of a rule because of the occurrence of a situation among those described in the rule’s antecedent (IF part). The second use of the rule is an instance of backward chaining, i.e., using a rule as one possible technique for establishing the situation described in its consequent by attempting to establish the situation described in its antecedent. The importance of if then rules in A.I. systems derives from the fundamental flexibility of these rules. Each rule states some independently valid aspect of the domain in which the system operates. The interpreter of the rules is then free to organize sequences of rule applications (inference by forward or backward chaining) into whatever structure best fits the requirements of each problem situation. Thus, the actual procedural structure (i.e., the sequence of actions taken) of an A. I. system composed of if-then rules can be radically different in situations that have different processing requirements. This leads to the fundamental property of adaptability in A.I. systems.
If-then rules that manipulate logical assertions (e.g., the last example) are sometimes called inference rules. A slight generalization of this notion leads to production rules. In this case the rules are of the same form, “IF antecedent THEN consequent”, but the antecedent and consequent can be arbitrary symbolic expressions that are not necessarily logical assertions. Production rules are frequently used to transform a symbolic model of something, while inference rules are generally used to extend a logical model. A s an example of production rules, consider the task of automatically configuring a computer to operate with a specific set of peripheral devices. At some point during this configuration task, the relationship between a collection of devices and the central computer might simply be represented with an abstract “interface” symbol. A production rule in this domain might say, “IF x is an interface whose attached devices numberless than four and are all serial, THEN consider implementing x as anRS-232 card.” Production rules can, for example, be used to take an abstract symbolic model of the desired result and successively transform it into a concrete full operational specification.
Rules are a powerful form of “active” knowledge because of their fundamental flexibility; the system is free to organize them in specific ways in response to each problem situation. However, this flexibility is also their primary drawback. Rules provide the A. I. system with the knowledge of the entire set of possible interpretations, possible conclusions, and possible actions in its problem situation. If the system is to be an effective problem solver in other than the most trivial domains, it must also have knowledge about& to go about its task of searching the information that it can derive from its rules. One technique used in A. I. for representing this type of knowledge, viz., knowledge about how to use more basic knowledge in an effective way, is the procedural problem-solving strategy’. For example, in the computer configuration domain described above, one possible problem-solving strategy might be, “To find an implementation for an interface, group its attached devices into serial devices and parallel devices, order the devices based on their data transfer rates and the priorities of the functions they provide, and then find a separate implementation for each related group of devices.” This strategy differs from the more straightforward rules in its use of procedural concepts like “perform these actions in this sequence”, or “iterate through all of the items in this set and perform some common action on them”. Traditional software systems can be viewed as implementing a single strategy that solves the entire problem; i.e., they are a single interconnected procedural solution. A.I. systems maintain their flexibility by separating their strategies into independently valid pieces of advice, as in the case of if-then rules (which can be viewed as particularly simple strategies). The more complex strategies can then be pieced together in order to respond to each problem situation in a special way, just as rules can be pieced together in this manner. There is a tradeoff between strong procedural knowledge (large strategies) that tends to be “brittle” and limited in its scope of applicability, and very general rules that tend to be difficult to effectively organize in specific situations.