Artificial intelligence ( edition) | Open Library – Uploaded by

Looking for:

[Artificial intelligence winston download

Click here to Download


See more about this book on Archive. When you buy books using these links the Internet Archive may earn a small commission. Previews available in: English. This edition doesn’t have a description yet. Can you add one? Showing 7 featured editions. View all 7 editions? Add another edition? Copy and paste this code into your Wikipedia page.

Need help? Artificial intelligence Patrick Henry Winston. An edition of Artificial intelligence Borrow Listen. Want to Read. Delete Note Save Note. Check nearby libraries Library. Last edited by ImportBot. December 7, History. Artificial intelligence Edit. Publish Date. Edition Availability 1. Artificial intelligence , Addison-Wesley. Not in Library. Libraries near you: WorldCat. Artificial intelligence: instructor’s manual , Addison-Wesley Pub.

Artificial intelligence , Addison-Wesley Pub. Artificial intelligence Publisher unknown. Book Details Published in Reading, Mass. Edition Notes Includes bibliographical references p. Classifications Dewey Decimal Class W56 , , Q W56 The Physical Object Pagination xxv, p.

Community Reviews 0 Feedback? Loading Related Books. December 7, Edited by ImportBot. November 2, November 13, October 7, April 1,


Artificial Intelligence Business: Commercial Uses of Artificial Intelligence – PDF Drive


You also learn how heuristic continuation and progressive deepening enable you to use all the time you have effectively, even though game situations vary dramatically In Chapter 7, Rules and Rule Chaining, you learn that simple if- then rules can embody a great deal of commercially useful knowledge. You also learn about using if-then rules in programs that do forward chaining and backward chaining.

By way of illustration, you learn about toy systems that identify animals and bag groceries. You also learn about certain key implementation details, such as variable binding via tree search and the rete procedure In Chapter 8, Rules, Substrates, and Cognitive Modeling, you learn how it is possible to build important capabilities on top of rule-based problem-solving apparatus. You also learn about knowledge engineering and about SOAR, a rule based model of human problem solving.

In Chapter 9, Frames and Inheritance, you learn about fumes, classes, instances, slots, and slot values. You also learn about inheritance, powerful problem-solving method that makes it possible to know a great deal about instances by virtue of knowing about the classes to which the instances belong. You also learn how knowledge can be captured in certain procedures, often called demons, that are attached to classes. In Chapter 10, Frames and Commonsense, you learn how frames can capture knowledge about how actions happen.

In particular, you learn how thematic-role frames describe the action conveyed by verbs and nouns, and you leam about how action frames and state-change frames describe how actions happen on a deeper, syntax-independent level In Chapter 11, Numeric Constraints and Propagation, you learn how good representations often bring out constraints that enabie conelu- sions to propagate like the waves produced by a stone dropped in a quiet pond.

You see how symbolic constraint propagation solves problems in line-drawing analysis and relative time calculation. You also learn about Marr’s methodological principles. In Chapter 13, Logie and Resolution Proof, you learn about logic, fan important addition to your knowledge of problem-solving, paradigms.

After digesting a mountain of notation, you explore the notion of proof, and you learn how to use proof by refutation and resolution theorem proving. In Chapter 14, Backtracking and Truth Maintenance, you learn hhow logic serves as a foundation for other problem-solving methods.

In particular, you learn about proof by constraint propagation and about truth maintenance. By way of preparation, you also learn what dependency- directed backtracking is, and how it differs from chronological backtracking, in the context of numeric constraint propagation. In Chapter 15, Planning, you learn about two distinet approaches to planning a sequence of actions to achieve some goal.

In this chapter, you learn how to define artificial intelligence, and you learn how the book is arranged. You get a feeling for why arti- ficial intelligence is important, both as a branch of engineering and as a kind of science.

You learn about some successful applications of artificial intelligence. Here is one: Artificial intelligence is.. This Book Has Three Parts To make use of artificial intelligence, you need a basic understanding of how knowledge can be represented and what methods can make use of that knowledge, Accordingly, in Part I of this book, you learn about basic representations and methods. Next, because many people consider learning to be the sine qua non of intelligence, you learn, in Part Il, about a rich variety of learning methods.

Some of these methods involve a great deal of reasoning; others just dig regularity out of data, without any analysis of why the regularity is there. Finally, in Part III, you focus directly on visual perception and lan- guage understanding, learning not only about perception and language per se, but also about ideas that have been a major source of inspiration for people working in other subfields of artificial intelligence.

However, in light of the way progress on language, vision, and robotics can and has influenced work on reasoning, any such separation seems misadvised. Artificial Intelligence Sheds New Light on Traditional Questions 7 In schools, computers should understand why their students make mis- takes, not just react to errors.

Computers should act as superbooks, displaying planetary orbits and playing musical scores, thus helping students to understand physics and musi. Moreover, the task of totally replacing a human worker ranges from difficult to impossible because we do not know how to endow computers with all the perception, reasoning, and action abilities that people exhibit. Nevertheless, because intelligent people and intelligent computers have complementary abilities, people and computers can realize opportunities together that neither can realize alone.

Moreover, it is usually simple to deprive a computer rogram of some piece of knowledge to test how important that piece really is.

It is almost always impossible to work with animal brains with the same precision. Artificial intelligence excites people who want to uncover principles that must be exploited by all intelligent information processors, not just by those made of neural tissue instead of electronic circuits. Instead, there is a new point of view that brings along a new methodology and leads to new theories. Just as psychological knowledge about human information processing can help to make computers intelligent, theories derived primarily with com- puters in mind often suggest useful guidance for human thinking.

Through artificial intelligence research, many representations and methods that peo- ple seem to use unconsciously have been crystallized and made easier for people to deploy deliberately. Once you have finished this book, you will be well on your way toward incorporating the ideas of artificial intelligence into your own systems.

Figure shows the resulting pattern of 2-y values. Outside the regular region lies a chain of 5 islands. The outermost re- gion is occupied by chaotic orbits which eventually escape. For example, a program developed by Kar! Ulrich designs simple de- vices and then looks for cost-cutting opportunities to reduce the number of components. Figure 1. In the rest of this book, you learn about one or two powerful ideas per chapter. There is no bothering with the color of the fox or the size of the goose or the quality of the grain; instead, there is an ex- plicit statement about safe arrangements and possible transitions between arrangements, The representation also is good because it exposes the natural con- straints inherent in the problem.

Some transitions are possible; others are impossible. The representation makes it easy to decide which is true for any particular case: a transition is possible if there is link; otherwise, it is impossible. You should always look for such desiderata when you evaluate repre- sentations. They suppress irrelevant detail: You can keep rarely used details out of sight, but still get to them when necessary.

They are fast: You can store and retrieve information rapidly. They are computable: You can create them with an existing procedure. Semantic Nets Convey Meaning 19 A Representation Has Four Fundamental Parts With the farmer, fox, goose, and grain problem as a point of reference, you can now appreciate a more specific definition of what a representation is.

The structural part specifies that links connect node pairs. And, as long as you are to solve the problem using a drawing, the procedural part is left vague because the access procedures are some- where in your brain, which provides constructors that. From the lexical perspective, semantic nets consist of nodes, denoting objects, links, denoting relations between objects, and link labels that denote particular relations. From the structural perspective, nodes are connected to each other by labeled links.

In diagrams, nodes often appear as circles, ellipses, or rectangles, and links appear as arrows pointing from one node, the tail node, to another node, the head node.

Semantic nets use constructors to make nodes and links, readers to answer questions about nodes and links, writers to alter nodes and links, and, occasionally, erasers to delete nodes and links. Nevertheless, the specifications are sufficiently precise to show that many of the key representations in artificial intelligence form famnily groups. Figure 2. Although this semantic-net family is large and is used ubiquitously, you should note that it is but one of many that have been borrowed, invented, or reinvented in the service of artificial intelligence.

There Are Many Schools of Thought About the Meaning of Semantics Arguments about what it means to have a semantics have employed philoso- phers for millennia. Say that meaning is defined by what the programs do. Let there be explanations of what deserip- tions mean in terms we understand intuitively. From the perspective of descriptive semantics, the net on the left side of figure 2. The net on the right side of figure 2. Of course, the objects and relations involved in semantic nets need not be so concrete, The representation used in the farmer illustration is a semantic net because particular arrangements of the farmer and his possessions can be viewed as abstract objects, thereby meriting node status, and allowed river crossings can be viewed as abstract relations, thereby rmeriting link status.

Evidently the unknown is most likely to be a single-hole switch ot late Number ofholes 5 unknown Many good programmers use a notation much like procedural English at the design stage, when they are deciding what a procedure will do.

Much of, the procedural English then survives in the form of illuminating comments. Feature-Based Object Identification Illustrates Describe and Match Feature-based object identification is one of the simplest applications of the deseribe-and-match method. Feature-based object identifiers consist of a feature extractor and a feature evaluator. Values obtained by the feature extractor become the coordinates of a feature point in feature space, a multidimensional space in which there is one dimension for each feature measured.

Generally, speed and discrimination considerations determine which features are used in particular situations. Candidate features for objects. One way to start is to describe rules that explain how A becomes B and how C becomes each of the answer figures.

Then, you can match the rule that explains how A becomes B to each rule that explains how C becomes an answer, The best match between rules identifies the best answer.

Thus, the describe-and-match paradigm can be used to solve analogy problems. Tho key to solving such problems lies in good rule descriptions. One rule part describes how the objects are tarranged in the source and destination figures. One object may be above, to the left of, of inside of another.

The other rule part describes how the objects in the source figure are transformed into objects in the destination figure. Rule descriptions consist of object-elation descriptions and object-transtormation descriptions. Links shown solid describe relations among source objects and among destination objects. Links shown dotted describe how objects are transformed between the source and the destination deleted to some combination of these operations.

Also, an object may be added or deleted. A typical rule can be described using a semantic-net representation, as illustrated in figure 2. Thus, one the other is not. You could write this specification for geometric analogy nets, of course, without any reference to semantic nets, by importing all the descriptive clements from the semantic-net specification. AS you can see, transforming the semantic-net concept into a geometric analogy net requires only the application-specific recitation of which link labels are allowed and what the nodes and links denote.

If the line erosses the second figure an odd number of times, then the second figure surrounds the first. ANALOGY computes the conter of area of each of the two objects, constructs diagonal tines through the center of area of one of them, and notes which region contains the center of area of the other object. Because the relations used are symmetric, it is not necessary to note both left and right relations.

Finally, ANALOGY uses a matching procedure to decide if an object in one figure can be transformed into an object in another figure by a combination of scaling, rotation, and reflection operations. The dotted links in figure 2. Now that you have seen how rules are constructed, you can see that the example in figure 2.

Relations between objects are determined by comparing centers of area. No object transformations can influence the solution, because no objects are transformed in the move from the souree figure to the destination figure. Note, however, that there is no a priori reason to associate! In going from the source figure to the destination figure, you want to be sure that squares go to squares, circles to circles, triangles to triangles, and so on, But this need to match one object to a geometrically similar object does not hold in comparing two rules.

In the example, answer 3 is to be selected even though the objects in A and B are a triangle and a square, whereas in C and in all the answer figures, the objects are a circle and a dot. This one-for-one association of variables implies that the number of objects that: move from the source figure to the destination Figure must be tho same in both of the two rules. The number of additions and deletions rust be the same as well. Any attempt to match two rules for which the numbers are different fails immediately.

If n objects move from the source figure to the destination figure i each of two rules being compared, there will be n! Symmetrically, for the problem in figure 2. ANALOGY concludes problem whose that the C-to-l rule best matches the A-to-B rule, because only answer 1 Solution is determined by transformations only, Because each figure has only one objec, relations between objects are not relevant.

But if an exact match cannot be found, then ANALOGY must rank the inexact matches, One way to do this ranking is to count the number of matching elements in the two rules involved in each match, as shown in figure 2.

Experimentally, the numbers shown in figure 2. Maximize y 18 AtoB is Ctox a different judgment about how the various possibilities should be ordered. A different set might indicate the opposite preference. Of course, itis possible to elaborate the measure of similarity in other directions. Suppose, for example, that Sag is the set of elements in the A-to-B rule, and that Sox is the set of elements in the C-to-X rule. The retaliation caused a loss for Thomas and a positive tradeoff for Al- bert.

The loss reversed Thomas’s previous success, and the positive tradeoff reversed Albert’s previous success. Just before the wedding, John discovered that Mary’s father was se- cretly smuggling stolen art through Venice. In this ilustration, there are two perspectives. This particu- lar combination of perspectives, events, and a mental state is called a retaliation. Of course, more complicated stories will have more complicated top- level abstraction nets.

Della and her husband, Jim, were very poor. Nevertheless, because Christmas was approaching, each wanted to give something special to the other.

Abstraction Units Enable Question Answering Abstraction units allow you to answer certain questions by matching. Answer by naming the central abstraction unit in the top-level abstraction net. In what way is one story like another? Answer by naming the most highly connected abstraction unit that appears in both top-level ab- straction nets.

If hard pressed, enumerate the other abstraction units that appear in both. Abstraction Units Make Patterns Explicit In this section, you have seen how a base-unit semantic net facilitates sim- ilarity analysis and summary by making mental states, events, and links between them explicit. Thus, the first criterion of good representation— that something important is made usefully explicit—is satisfied.

Some people argue, however, that a base-unit semantic net does not yet pass the computability criterion for good representation because there is no fully specified way to translate text into abstraction-unit patterns.

Problem Solving and Understanding Knowledge 43 Figure 2. Perhaps the important, knowledge concerns the description of concrete or abstract objects. Alternatively, perhaps the important knowledge is about a problem-solving method. Some knowledge may, for example, fit nicely within the semantic-net frame- work, Other knowledge is best embedded in a collection of procedures.

After learning what kind of knowledge is involved in a task, this question should be the one you ask. Are there 40 things to know, or , or 4,? Another is that knowing the size of a problem builds courage; even if the size is large, digesting bad news is better than anticipating even worse news unnecessarily, In any event, the tendency is to overestimate grossly; after seeing that a task is reasonably complicated, it is easy to suppose that it is unimag- inably complicated, But many tasks can be performed with human-level competence using only a little knowledge.

What exactly is the knowledge needed? Ultimately, of course, you need the knowledge. To do geometric analozy problems, you need to know what relations are possible between figure parts, and you need to know how parts can change. To recognize abstrac- tions, you need a library of base and composite abstraction units. Much of learning any subject, from electromagnetic theory to genetics, is a matter of collecting such knowledge.

You solve geometric analogy problems by determining which rules are most. Story plots can be viewed as combinations of mental states and events. Hora []. You also learn about two new representations, both of which can be viewed as special cases of the semantic-net representation introduced in Chap- ter 2.

You will also begin to see that you yourself use similar represen- tations and methods daily as you solve problems. In the rest of this section, you learn more about the generate-and-test method, you learn which sort of problems the generate-and-test method solves, and you learn several criteria that good generators always satisfy.

In identification problems, the generator is said to produce hypotheses. The generator is the procedure that the burglar uses to select and dial combinations. The tester is the procedure that the burglar uses to work the handle. Careful safecrackers make sure that they try all possibilties, without any repeats, untl a twist of the handle opens the safe.

To use the generate-and-test paradigm to identify, say, a tree, you can reach for a tree book, then thumb through it page by page, stopping when you find a picture that looks like the tree to be identified. Thumbing through the book is the generation procedure; matching the pictures to the tree is the testing procedure.

At three per minute, figuring that he will have to go through half of the com- binations, on average, to succeed, the job will take about 16 weeks, if he works 24 hours per day. Informability is important, because otherwise there are often too many solutions to go through. Consider the tree-identification example.

Thus, a state space is member of the semantic-net family of representa- tions introduced in Chapter 2. Tn the rest of this section, you learn about means-ends analysis, a standard method for selecting transitions. Thus, the identified procedure reduces the observed difference between the current state and the goal state Consider the states shown in figure 3.

Solid-line nodes identify the current state and the goal state. Dotted-line nodes correspond to states that are not yet known to exist. Descriptions of the current state, or of the goal state, or of the difference between those states, may contribute to the identification of a difference-reducing procedure In figure 3. The notation indicates that each molecule has eight atoms of carbon, 16 of hydrogen, and one of oxygen. When so used, problem reduction is often called, equivalently, goal reduction.

Conveniently, the names of these procedures are mnemonics for the problems that the procedures reduce. Figure 3.

It works by activating other procedures that find a specific place on the top of the target block, grasping the traveling block, moving it, and ungrasping it at the specific place. Plainly, the following sequence suffices: Grasp D. Move D to some location on the table. Ungrasp D. Grasp C. Ungrasp C. Grasp A. Other nodes have exactly one parent. Clearing the top of block A is shown as an immediate subgoal of grasping block A.

Goals that are satisfied only when all of their immediate subgoals are satisfied are called And goals. Most goal trees also contain Or goals; these goals are satisfied when any of their immediate subgoals are satisfied. The corresponding, unmarked nodes are called Or nodes Finally, some goals are satisfied directly, without reference to any sub- goals.

These goals are called leaf goals, and the corresponding nodes are called leaf nodes. Further suppose that someone asks, How did you clear the top of A? Plainly, a reasonable answer is, By getting rid of block C. On the other hhand, suppose the question is, Why did you clear the top of A? Then « reasonable answer is, To grasp block A.

If the goal is fan And goal, report ail of the immediate subgoals. Each time one specialized procedure calls another, it effects a problem-reduetion step. Thus, problem reduction is the problem-solving method that all but the shortest programs exhibit in great quantity. Ac- cordingly, you often see problem-solving methods working together.

Thus, the initial goal, as shown in figure 3. Good generators are complete, nonredundant, and informed The key idea in means-ends analysis is to reduce differences. A goal tee consists of And goals, all of which must be satisfied, and Or goals, one of which must be satisfied. While Mover is at work, it constructs a goal tree that enables it to answer how and why questions. Here, a travel problem is split apart, using problem reduction, into pieces susceptible to solution using means-ends analysis.

Figure 8. The system answers How did you show? In the horse-evaluation system, taken up in Chapter 7, the principal rule says that a horse is valuable if it has parented something fast, which is a way of saying that a horse qualifies as a good stud or brood mare.

Accordingly, you might well elect to assume that all horses are fertile, banishing the z is fertile antecedent to a new part of the rule where providing assumptions are collected: Modified Fertile-Parent Rule Ie Pris-a horse? Here is fanciful example: Live Horse Rule If?

Of course, there are other ways to treat providing assumptions and unless assumptions, in addition to show-me mode and ask-questions-later mode. Essentially, each way specifies whether you work on providing as- sumptions and unless assumptions using assertions or rules or neither ot both.

Here are representative examples: 1 Decision-maker mode: Assume that providing assumptions are true. This mode is reminiscent of the progressive deepening, idea, introduced in Chapter 6, in connection with adversarial search, Probability Modules Help You to Determine Answer Reliability Rule-based deduction systems used for identification usually work in do- mains where conclusions are rarely certain, even when you are careful to incorporate everything you can think of into rule antecedents.

Each probability reflects how certain an asser- tion is, with 0 indicating that an assertion is definitely false and 1 indicating that an assertion is definitely true. Knowledge engineering is the extraction of useful knowledge from domain experts for use by computers. Often, albeit far from always, knowl edge engineers expect to cast the acquired knowledge in the form of rules for rule-based systems. Suppose, for example, that you are a knowledge engineer and that you fare working on new rules to be used by the BAGGER system introduced in Chapter 7.

If you ignore the heuristic of specific situations, you would have just asked a few real grocery-store baggers to describe what they do. But 1no matter how cooperative your domain experts are, they are unlikely to be able to help you much unless you provide more evocative stimulation.

You might well learn nothing about what to do with, for example, eggs. On the other hand, if you adhere to the heuristic of specific situ you would get yourself into a grovery store so as to watch baggers handle specific situations, like the one shown in figure 8.

Not- ing, for example, that « real bagger handles the two items in figure 8. Note that ideas for extracting knowledge for computers also apply when your motive isto extract knowledge for your own use. To add a question-answering superpro- cedure, for example, you need to deal with only rules and rule histories. Further suppose that you have determined that another rule is needed, one that captures the idea that an animal is a carnivore if it is seen stalking another animal.

When compared with the other rules that conclude that an animal is a carnivore, this proposed rule lacks an antecedent requiring that the animal is a mammal. Noting this lack, it makes sense to ask whether the omission is intended. In the example, the carnivore rule group is applicable. In Chapter 17, you learn about another procedure that can do even more by directly producing rulelike knowledge from precedents and exercises.

Rule Interactions Can Be Troublesome It would seem that rule-based systems allow knowledge to be tossed into systems homogeneously and incrementally without concern for relating new knowledge to old. One particu: lar problem is that the advantage of bequeathing control becomes the disadvantage of losing control, as King Lear failed to foresee. In principle, there is nothing to prevent building more humanlike systems, using rules, because rules can be used as a sort of programming language.

The answer is yes, at least to many computationally-oriented psychologists who try to understand ordinary hu- man activity using metaphors shared with researchers who concentrate on making computers smarter. Rule-Based Systems Can Model Some Human Problem Solving In the buman-modeling world, if-then rules are called productions and rule-based systems are called produetion systems, Hard-core rule-base system enthusiasts believe that human thinking involves productions that are triggered by items in short-term memory.

These transcripts are called protocols. Each time the subject acquires knowledge through his senses, makes a deduction, or forgets something, the state of knowledge changes. By analyzing the way one state of knowledge becomes another, you can draw inferences about the productions that cause those state changes. It would seem that assertions accumulate in the following order: hair, mam- mal, pointed teeth, claws, forward-pointing eyes, carnivore, tawny color, dark spots, and cheetah.

SOAR Models Human Problem Solving, Maybe In artificial intelligence, an architecture is an integrated collection of rep- resentations and methods that is said to handle a specified class of problems or to model insightfully a form of natural intelligence. SOAR features a long-term memory for productions and a short-term memory for items that trigger the produc- tions.

SOAR starts from an initial situation, the current state, in the expectation that it will arrive at an identifiable goal state eventually, One such net is shown in figure 8.

Second, SOAR translates the preference labels and preference links into dominance rela- tions among the states. Current state ce Figure 8. Then, Soar repeats the cycle, placing new links, determining new dominance relations, and updating the current state, until Soar reaches the goal state, One example of a preference net is shown in figure 8.

State C is the current state. The links labeled acceptable, worst, and better carry preference information. SOAR Uses an Automatic Preference Analyzer To use preference labels and links, SOAR uses an automatic preference analyzer—one that resolves inconsistencies caused by rules with limited To see how Soan’s automatic preference analyzer works, consider the nodes and links shown in figure 8. Consequently, SOAR selects state D to be the next problem state, leav- ing SOAR ready to start another eycle of preference marking, preference analysis, and state selection.

While discovering new problem states and labeling known problem states with preference information, SOAR uses no conflict-resolution strat- egy to decide which triggered rule should fire. Instead, SOAR fires all trig- gered rules and decides what to do by looking at what they all do, rather than by using some rigidly prescribed, result-ignorant.

SOAR users are expected to anticipate various sorts of problem-specific impasses so as to provide the appropriate productions for setting up subgoal-handling problem spaces, problem-space-dependent problem states, and problem- state-dependent operators. Probability modules help you to determine answer reliability. Acquisition modules assist knowledge engineers in knowledge transfer from « human expert to a collection of rules. One is to work with specific situations; another is to ask about situation pairs that look identical, but are handled differently.

Rule-based systems can behave like idiot savants, They do certain tasks well, but they do not reason on multiple levels, they do not use constraint-exposing models, they do not look at problems from different perspectives, they do not know how and when to break their own rules, and they do not have access to the reasoning behind their rules.

Rule-based systems can model some human problem solving. Michalski and Patrick H. Winston []. Frames and : Inheritance : In this chapter, you learn about frames, slots, and slot values, and you learn about inheritance, a powerful problem-solving method that makes it, possible to know a great deal about the slot values in instances by virtue of knowing about the slot values in the classes to which the instances belong.

With basic frame-representation ideas in hand, you learn that frames can capture a great deal of commonsense knowledge, informing you not only about what assumptions to make, but also about for what information to look and how to look for that information.

You learn that much of this knowledge is often embedded in when-constructed procedures, when- requested procedures, when-read procedures, when-written procedures, and with-respect-to procedures.

By way of illustration, you see how to use frames to capture the general properties of various kinds of dwarfs, and you sce how to use frames to capture the properties of various kinds of newspaper stories. Once you have finished this chapter, you will understand that frames can capture a great deal of commonsense knowledge, including knowledge about various sorts of objects ranging from individuals to events. You will also know how the CLos inheritance procedure determines a precedence ordering among multiple classes.

This capability enables you to make use of the following general knowledge about fairy-tale dwarfs Fairy-tale competitors and gourmands are fairy-tale dwarfs. Most fairy-tale dwarfs are fat. Graphically, frames may be shown in an alternate, rectangle-and-slot notation. Bach frame’s name is the same as the name of the node on which the frame is based. Accordingly, you can talk about a slot, rather than about a link that emanates from a node.

Frames may Describe Instances or Classes Many frames describe individual things, such as Grumpy, an individual wart. These frames are called instance frames or instances. Other frames describe entire classes, such as the dwarf class. These frames are called class frames or classes. In figure 9. The Managers class is a subclass of the Competitors class, for example. The new instance is connected automatically to the class frames via an Is-a slot in the new instance.

Its input is a frame, the name of a slot, and a value to be installed. Finally, a slot reader retrieves slot values. Its input is a frame and the name of a slot; its output is the corresponding slot value.

If fa superclass has a slot, then the instance inherits that slot. Sometimes, slot values are specified after an instance is constructed, Alter Blimpy is constructed, for example, you can indicate that Blimpy is smart by inserting the value Smart in Blimpy’s Intelligence slot.

Alternatively, the slot values of an instance may be specified, somehow, by the classes of which the instance is a member. Tn the simplest class hierarchies, no more than one when-constructed procedure supplies a default for any particular slot. Often, however, several when-constructed procedures, each specialized to a different class, sup- ply default values for the same slot. Whenever an individual is both a Competitor and Dwarf, both procedures compete to supply the default value, Of course, you could specify an inher- itance procedure that allows multiple procedures to supply defaults, but the usual practice is to allow just one procedure.

How can you decide which when-constructed procedure is the winner? First, you learn about the special case in which no individual has more than one Is-a link and no class has more than one Ako link. Once this foundation is in place, you learn about more complicated hierarchies in which individuals and class have multiple inheritance links.

One of the sample procedures, because it deals with new Dwarfs, is attached to the Dwarf class; the other js attached to the Competitors class. That way, you can find both by search up from the new instance through Is-a links and Ako links. This ordered list is called the class-precedence list: Blimpy Managers class Competitors class — procedure stored here Dwarfs class — procedure stored here Everything class [A procedure that is specialized to one of the classes on the class-precedence list is said to be applicable.

You have Blimnpy’s class-precedence list, which supplies two procedures for computing values for the Physique slot. This kind of ambiguity is always resolved in favor of the most specific applicable procedure—the one that is encoun- tered first on the class-precedence list. There is no loss of generality in CLOS, however, because an instance ean be attached to a class that is wholly dedicated to that instance and that has multiple Ako connections to the desired superclass.

Because Blimpy belongs to both the Gourmands class and to the Diarists class, as woll as the Managers class the class hierarchy branches upward. Note, however, that you must modify depth-first search slightly, be- cause you want to include all nodes exactly once on the class-precedence list. To perform exhaustive depth-first search, you explore all paths, depth first, until each path reaches either a leaf node or a previously- encountered node.

This conclusion seems at odds with intuition, however, because the Gourmands class is a subclass Of the Dwarfs class, Surely a class should supply more specific procedures than any of its superclasses. Because there are three paths from Blimpy to the Dwarfs class, the Dwarfs class is ignored the first and second times it is encountered. Consequently, the Gourmands class is the next one added to the class-precedence list, followed by the Diarists class.

Then, the Dwarfs class is encountered for the third and final time, whereupon it is noted for the frst time, enabling it and the Everything class to be added to the class-precedence list.

No class appears after any of its own superclasses. Endomorphs class, But suppose one Isa link and two Ako links are added, as in figure 9. In Doth instances, the order changes are caused by the addition of Ako links connected to other classes. These order changes are bad because left-to- right order, by convention, is supposed to indicate priority.

Thus, you know the order of a class’s direct superclasses oon the class’s class-precedence list as soon as you know how the direct su- perclasses are ordered: You do not need to know the entite structure of the class hierarchy. Before you learn the details of the topological sorting procedure, how- ever, you will find it helpful to see what happens when a path through fa class hierarchy is expressed as a list of adjacent pairs.

For example, the simple, nonbranching class hierarchy in figure 9. Skip carousel. Carousel Previous. Carousel Next. What is Scribd? Explore Ebooks. Bestsellers Editors’ Picks All Ebooks.

Explore Audiobooks. Bestsellers Editors’ Picks All audiobooks. Explore Magazines. Editors’ Picks All magazines. Explore Podcasts All podcasts. Difficulty Beginner Intermediate Advanced. Explore Documents. Uploaded by martin. Did you find this document useful? Is this content inappropriate? Report this Document. Flag for inappropriate content. Download now. For Later. Jump to Page. Search inside document.

Includes bibliographical references p. Tite Q W56 , All rights reserved. Printed in the United States of America 5. The cover design is by Dan Dawson. The interior design is by Marie McAdam. The following people also have made especially valuable suggestions: Johnnie W.

Eric L. Gunther Ascent Technology , James R. To learn how to obtain this software, send a message to ai3Gai. Your message will be answered by an automatic reply program that will tell you what to do next.

If you wish to report a bug or offer a suggestion via ordinary mail, write to the author at the following address: Patrick H. One group—computer scientists and engineers—need to know about artificial intelligence to make computers more useful. Another group—psychologists, biologists, linguists, and philosophers—need to know about artificial intelligence to understand. Ideas from those disciplines are discussed, in a spirit of scientific glasnost, but those discussions are in, optional sections, plainly marked and easily detoured around.

One of these is the change brought about by the incredible progress that has been made in computer hardware. Many simple ideas that seemed silly 10 years ago, on the ground that they would require unthinkable computa- tions, now seem to be valid, because fast—often parallel—computing has become commonplace.

As described in Chapter 19, for example, one good. Accordingly, about one-third of the chapters in this edition are devoted to various approaches to learning, and about one-third of those deal with neuronlike nets, Still another remarkable change is the emergence of breakthroughs. As described in Chapter 26, for example, one good way to identify a three- dimensional object is to construct two-dimensional templates for the given object, in the given view, for each possible object class. Ten years ago, no one suspected that the required templates could be manufactured perfectly, simply, and on demand, no matter how the object is viewed.

Finally, there is the emphasis on scaling up. These days, it is hard to attract attention with an idea that appears suited to toy problems only. This difficulty creates a dilemma for a texthook writers, because textbooks need to discuss toy problems so that the complexities of particular real worlds do not get in the way of understanding the basic ideas.

To deal with this dilemma, I discuss many examples of important applications, but only after I explain the basic ideas in simpler contexts, This Edition Responds to Suggestions of Previous Users Many readers of the first and second editions have offered wonderful sug- gestions. At one meeting in Seattle, on a now-forgotten subject, Peter Andreae and J, Michael Brady remarked, over coffee, that it was hard for students to visualize how the ideas could be incorporated into programs.

Similarly, feedback from my own students at the Massachusetts Insti- tute of Technology indicated need to separate the truly powerful ideas and unifying themes—such as the principle of least commitment and the importance of representation—from nugatory implementation details. Looking at this book, it might seem that I have neglected my own dictum, because this book has grown to be many times pages long. The book is modular.

If you want to develop a general understanding of artificial intelligence, you should read Chapters 2 through 12 from Part I; then you should skim Chapter 16 and Chapter 19 from Part II, and Chapter 26 from Part IIL to get a feel for what is in the rest of the book.

If you are interested primarily in learning, you should read Chapter 2 from Part I to learn about representations; then, you should read all of Part IT. If you are interested primarily in vision or in language, vou can limit yourself to the appropriate chapters in Part I This Edition Is Supported by Optional Software This book discusses ideas on many levels, from the level of issues and alternatives to a level that lies just one step short of implementation in computer programs.

Winston and Berthold K. In Chapter 1, The Intelligent Computer, you learn about what artificial intelligence is, why artificial intelligence is important, and how artificial intelligence is applied. You also learn about criteria for judging In Chapter 2, Semantic Nets and Description Matching, you learn about the importance of good representation and you learn how to test a representation to see whether it is a good one.

Along the way, you learn about semantic nets and about the describe-and-match method. You also learn how heuristic continuation and progressive deepening enable you to use all the time you have effectively, even though game situations vary dramatically In Chapter 7, Rules and Rule Chaining, you learn that simple if- then rules can embody a great deal of commercially useful knowledge.


Download PDF – Artificial Intelligence – Patrick Henry [en5kjwyd2pno].Patrick Henry Winston | The Online Books Page

Views Downloads File size 33MB. Report DMCA / Copyright. DOWNLOAD FILE. Recommend Stories. Artificial Intelligence – Patrick Henry Thank you unquestionably much for downloading artificial intelligence 3rd edition you have knowledge that, people have see. Downloaded from on by guest. Artificial Intelligence Winston Patrick Henry. Right here, we have countless books Artificial Intelligence.


Leave a Comment