- 2 days ago
An early look at artificial Intelligence. Guests includes Edward Feigenbaum of Stanford University, Nils Nilsson of the AI Center at SRI International, Tom Kehler of Intellegenetics, Herb Lechner of SRI, and John McCarthy of Stanford. Featured demonstrations include Inferential Knowledge Engineering and the programming language LISP. Originally broadcast in 1984. Copyright 1984 Stewart Cheifet Productions.
Category
🤖
TechTranscript
00:00Micro Focus, creators of visual programming tools for software development,
00:27is pleased to provide major funding for the computer chronicles the story of this continuing evolution welcome to the computer chronicles i'm steward sheffae sitting in for gary kildall this week is herb leckner of sri
00:56herb we think of chess as the ultimate game of skill a game that requires mental agility intelligence if you will yet here i am playing a game of chess with a computer which is analyzing board positions and applying a certain kind of intelligence to figure out what its next move should be that's the subject of our program today artificial intelligence and in some people's minds ai suggests attempting to duplicate the way a human brain works is that what ai is in fact
01:24not in most modern ai research stuart early research in ai looked at duplicating human thought processes but current ai research is more concerned with duplicating the end result of intelligence and computers that act as experts in expert systems and computers that can communicate with us in our language understanding some of the context of that language are two areas that are receiving a lot of attention in artificial intelligence research today
01:52okay we're going to meet two of the world's foremost experts on artificial intelligence and we'll take a look at two fascinating examples of expert systems one of the leading experts on expert systems is dr edward feigenbaum of stanford university we asked professor feigenbaum about the evolution of artificial intelligence the computers as you know are general symbol processing devices capable of manipulating any kinds of symbols of which
02:22the numbers constitute one important class but computers are much more general than that we've known about the generality of computation since at least the time of touring in the 1930s and actually i've tracked it back to intuitions that babbage
02:39had that were reported by aida lovelace after whom the aida programming language is named in 1842 aida lovelace wrote that the analytical engine of babbage's constituted the
02:52the link between the mechanical world and the world of the most abstract concepts.
02:58That currently, in the modern terminology, is called the physical symbol system hypothesis
03:03and is the basis for artificial intelligence work.
03:07In artificial intelligence as a science, we talk about the use of computers to process
03:13symbolic knowledge using logical inference methods, symbolic inference methods.
03:19In other words, we're talking about inference and not calculation in the traditional sense.
03:25We're talking about knowledge and not numbers in the traditional sense.
03:30A current application of inferential knowledge engineering is in the field of expert systems,
03:35programs like this oil exploration advisor.
03:39Developed to assist drilling rigs in remote areas, the program behaves much like a qualified specialist.
03:45It asks questions of its user and then gives advice on how to avoid or correct accidents
03:50so common in drilling operations.
03:53If at some point in the line of questioning the user becomes confused by the query, he
03:57can ask the program to explain itself simply by entering why.
04:02The advisor responds with a specific source behind the question and explains its reasoning
04:08up to that point.
04:10The symbolic processing behind the human-like talents of this program have found applications
04:15in a broad range of fields from medicine to robotics.
04:19Perhaps the most difficult area for AI to master has been natural language, a talent that results
04:24in very friendly computers but that requires enormous processing capacity.
04:29The ambiguities of syntax and context have restricted present systems to very limited applications.
04:35A simple geographic question, depending on the phrasing, can lead to multiple interpretations.
04:41The program will determine the question's most likely meaning only after an exhaustive deconstruction
04:46of the sentence and might even reject an unusual phrasing of the same query.
04:53In a parallel development, investing AI programs with linguistic ability has led to an interactive
04:59study of graphic communication.
05:01In this experiment, the visual dynamics of communication form the foundation for a linguistic expert advisor
05:08capable not only of discerning visual patterns but also acting as a kind of tutor.
05:13One interactive application has given persons without normal speech the ability to adapt or
05:18construct alternative symbolic languages, a remarkable instance of using a computer program's intelligence
05:25to help interpret human intelligence.
05:40Joining us now is Jeffrey Perrone.
05:42Jeffrey is a management consultant involved in expert systems for micros and Niels Nielsen,
05:47director of the Artificial Intelligence Center at SRI International.
05:51Niels Nielsen, could you kind of help us scope the field of artificial intelligence as it
05:56exists today?
05:57Well, I'll try, Herb.
05:58It's a broad field.
05:59If you ask many of different people what constitutes artificial intelligence, I think you'd
06:04get a lot of different answers.
06:05For me, I think it concerns mainly the putting of knowledge into computer programs so that the
06:11programs then can solve certain problems which humans find somewhat easy or perhaps intellectual
06:16challenges sometimes.
06:18And the knowledge that one puts in is the knowledge that's represented in a particular sort of
06:23way.
06:24So the idea is that it's just not smarter programs that do artificial intelligence programming.
06:29It is some difference in techniques that they use relative to what ordinary programming
06:34is.
06:35Well, there's a part of computer science that artificial intelligence is concerned with that
06:38does in fact involve a certain body of techniques that are a little different perhaps
06:41than what goes on in the rest of computer science.
06:45Jeffrey, Niels is involved in the research end of AI and you're involved in some of the
06:50applications.
06:51Are we at the point today where we can apply AI?
06:54Absolutely, Stuart.
06:55In what ways?
06:56Well, I feel that artificial intelligence and specifically expert systems or knowledge-based
07:02systems now are applicable any place that specialized knowledge is used routinely to reach decisions,
07:08troubleshooting strategies, diagnoses, those areas.
07:13So I kind of think of it as something where we've reached a watershed where it is no longer
07:19very expensive or very difficult for individuals with no technical background to build systems
07:24and apply them usefully.
07:26And you distribute a system which is an expert system and also helps one build expert systems.
07:32Is that correct?
07:33Right.
07:34Well, actually, it's a tool for generating particular expert systems applications in virtually
07:39any area where there is that routine use of special knowledge.
07:42And it runs on microcomputers?
07:44Right.
07:45It runs on the IBM personal computer and a number of other compatible microcomputers.
07:49Do you use the same techniques in this system that Nils was referring to or used in systems
07:56that might run on larger equipment?
07:58It uses some of the same ideas and it has its own unique approach as well.
08:04One of the specific areas where this is a different sort of program is that instead
08:08of requiring explicit statement of rules to build a system, it only requires examples
08:15of previous decisions, previous tasks, previous diagnoses to build the systems.
08:20And that's going to break through, I believe, the knowledge engineering bottleneck mentioned
08:25by people like Edward Feigenbaum in building these systems.
08:28Jeffrey, you have expert ease your system up here.
08:32And show us the demonstration.
08:33Sure.
08:34Okay.
08:35Well, this is a particular application generated with this system.
08:38And it was generated by an anesthesiologist, Hilary Don and myself.
08:42And what it does is it makes diagnoses for breathing or airway problems.
08:47The first question here that it asks the naive user is, when was the onset of the problem?
08:52And I'll answer days.
08:54Then we'll ask, what's the quality of the stridor, which is a rasping sound made when
08:58one's breathing, kind of a sort of sound.
09:00And let's say that the quality is moderate.
09:03Who would use this system?
09:04A doctor or...?
09:05This might be used by a physician's assistant to screen patients for perhaps further fine-tuning
09:11by someone with more expertise.
09:13Is the patient drowsy?
09:14Let's say no.
09:15Are there any predisposing factors to developing stridor, prior events that might lead to this
09:21condition?
09:22Let's say yes.
09:23And it comes up with, consider the diagnosis of intrathoracic tracheal stenosis.
09:27Now, I'm not a physician, so I don't know exactly what that means.
09:30But that gives you an idea of how it might lead to a particular diagnosis.
09:35Now, what I'd like to do is I'd like to show you a trivial example called Sunday, which
09:41actually comes with the manual for the program as a tutorial, essentially.
09:46Now, Sunday is a model of deciding what to do on a Sunday afternoon.
09:50How are you going to spend your time?
09:52And it consists of a couple of questions, multiple-choice questions like you've just
09:57seen, and those would be answered.
10:00So, now what I'm going to do is run the query system.
10:03And it asks, do you have your family with you?
10:05Let's say yes.
10:06What is the weather like?
10:07Is it raining or sunny?
10:09Let's say sunny.
10:10So, it says, why don't you take the family to the beach?
10:12Now, the way that you would build one of these models is by starting on something called
10:17the attribute screen.
10:19The attribute screen is where you sketch out the dimensions of the particular problem or situation.
10:25Now, here, the attributes are weather and family.
10:28And there's the advice column, which would be the result of a particular decision.
10:32What size matrix can you do on that?
10:34Okay.
10:35You can actually do, in any particular problem, 31 attributes and up to 255 values per attribute.
10:42And those can then be chained together.
10:44So, you can do very large models, just limited by the size of your disk storage, essentially.
10:50The values, such as raining and sunny, could be thought of as answers to the question represented
10:55by the attribute.
10:57Once you've done a preliminary sketch of this sort, you then go to something called the
11:01example screen.
11:02And on the example screen, you enter in examples of previous decisions.
11:07Now, this was developed by Dr. Donald Mickey in Edinburgh, based on the idea that information
11:14is communicated by experts or masters to their apprentices in the form of examples or cases.
11:20So, here you see, horizontally running across the screen, answers to those questions.
11:26Raining, yes, we're with the family.
11:28The advice would be, go to the museum.
11:30And the next case would be, if it's sunny and I was with the family, then I would go to the beach.
11:35And the third case there consists of, I don't care what the weather is like.
11:40If there's no family with me, then I'm going to go fishing.
11:44Jeffrey, you described this, obviously, as a trivial example, this Sunday one.
11:48But in general, do you have the capacity within a PC to seriously approach this kind of problem solving?
11:54Yes, you can do quite significant things.
11:56As a matter of fact, I've recently been speaking with the Whole Earth Software Review people
12:00about doing some systems to recommend software, such as word processing programs
12:05or communications hardware modems, that sort of thing.
12:08Niels, I want to ask you what I hope doesn't sound like a dumb question, but kind of what's the point?
12:13Why do we develop something like this kind of system?
12:16Is it to replace the doctor, say, in the diagnosis situation?
12:20What is the end result of this?
12:22Well, I think there would be a lot of uses of systems like this.
12:25I think the present ones actually do have a lot of brittle features that perhaps might limit their utility at the present time.
12:32But in the long run, that is 10 years, 20 years, something like that,
12:36these kinds of systems, I think, will be quite generally useful in a wide range of settings.
12:42First of all, the knowledge that these systems contain, at least we hope to put in the knowledge of world-class experts,
12:49people who know so much about the field that there might only be five or six of them in the whole world who know that much.
12:56Now, there can be some pretty good practicing people who, nevertheless, aren't quite as good as the world experts in a particular subject.
13:03So it helps us spread the knowledge of an expert around a good deal in a way which can be copied quite easily.
13:10Okay, in just a moment, we're going to meet the man who invented the term artificial intelligence,
13:14and we'll see a demonstration of knowledge engineering. That's coming up next.
13:31Joining us now is Tom Kaler. Tom is Director of Applied Artificial Intelligence at Intelligenetics.
13:37Tom, you've got a complex system here. First of all, tell me how this would be different from the system we saw in the first part of the program.
13:44Well, one way that it's different is that we can develop a graphics environment which connects to the underlying knowledge bases
13:52so that the user really is just fooling around with meters and valves and objects that they're used to.
13:59Here, for example, if we look at this levelometer, we can just point to it.
14:03Before you get to that, what's the environment we're looking at here?
14:05Oh, this is the key system, the knowledge engineering environment.
14:08And we have knowledge bases hooked up to a control panel right here at the moment.
14:14And it's just a little example to show you how we can hear...
14:17A control panel of what?
14:19This could be a control panel, say, to a nuclear reactor or to an instrument
14:23or any other kind of operational device that you would want to work with.
14:28And here, for example, we're just changing the value of a meter,
14:32and notice that it's generating a strip chart showing the time course behavior.
14:37Another thing we can do with these kinds of panels is we can go in and look at valves,
14:42and we can affect the valve whether it's open or shut.
14:45Now, what's important about this is this is affecting an underlying knowledge base,
14:50which will then apply heuristics and decision-making procedures to the system.
14:55Can you show us that, how you can go to the part? There you go.
14:58Sure.
14:59We solve a problem.
15:00Well, what we see underneath here now is a knowledge base for a nuclear reactor.
15:04And we can have a representation of the components of the nuclear reactor
15:08as well as a representation of the decision-making behavior
15:12that would be carried out by the expert.
15:14And what you will see here is that it's both possible to use rule-based reasoning
15:19as well as representation of the underlying knowledge.
15:22One of the first things we'll do here is just take a look at how we could test the hypotheses
15:28of a particular accident situation.
15:31And we just point to test hypotheses and activate it here.
15:34And what we see the system doing is going through a decision-making process
15:38where it's accessing the underlying knowledge base,
15:41looking at the various states of the components.
15:44And by the way, the states of these particular components can be accessed directly into the instrument
15:49or into the nuclear reactor.
15:50Now, you can actually interrogate the system and ask it what course of logic is it following.
15:55Is that correct?
15:56That's right.
15:57Right now, it's come up and asked the operator a question.
15:59Is it true that the steam generator level is decreasing?
16:02So we can say, why is this being asked?
16:04And it will explain to us that it's doing it in order to prove that the steam generator inventory is inadequate.
16:10We will then give it an answer, and it goes on to reach a conclusion.
16:13Okay, Tom, this system runs in Lisp.
16:16Why is Lisp the right language for an AI application?
16:20Well, one of the reasons that Lisp is so important is that you need to have very powerful symbol-manipulating capabilities.
16:27Decision-making is primarily a symbolic activity.
16:31Knowledge representation is primarily symbolic, and you need that kind of capability.
16:35Okay, thank you. Herb.
16:37John McCarthy has joined us.
16:39John is a professor of computer science at Stanford University and one of the pioneers in the field of artificial intelligence.
16:46Among his other accomplishments, he invented Lisp, the language that Stuart just referred to.
16:52John, why did you invent Lisp?
16:54Well, for artificial intelligence work, the kind of thing that they are doing here.
17:00What are its characteristics that make it different than some other language as applied to artificial intelligence?
17:07Besides the ones that Tom mentioned, one of its important characteristics is that its programs are in the same format as its data,
17:15so that it is easy to make programs that produce programs and look at programs.
17:22So it allows us to deal with facts and logic more than numbers.
17:26That's correct.
17:27I guess it's on everyone's mind when we talk about smart machines.
17:32How smart can machines become?
17:34What are the limits of artificial intelligence?
17:37Well, I see no limit short of human intelligence,
17:41and then with faster machines one could do the equivalent that a human could do in a short time in a very long time.
17:51Neils, what's your thought on that? How far can we go with this?
17:54Well, I'd separate some of the problems that we're facing in order to make machines more intelligent into about three different varieties.
18:00We've heard a lot about how important knowledge is, and one of the important things about knowledge is what knowledge should be represented in a program.
18:07And some of the difficulties that we're having in making programs smarter is that we're not sure exactly what it is we want to tell those programs about certain subjects.
18:15Another category is once we figure that out, how do we represent the knowledge in the computer itself?
18:20And certain kinds of knowledge proving a little difficult for us to represent it.
18:24The third category has to do with how that knowledge is used, and certainly there's various activities there that have been rather successful,
18:31but other things that really have to be done yet.
18:34Is there a frustration level in this field in that there is a lot of hope of what you can do with something like AI,
18:41but yet, as you've pointed out, Neils, it seems harder perhaps than one thought to really put this into a practical application?
18:48John?
18:49Well, there is a certain level of technology in artificial intelligence today, and many people are making applications based on this technology.
19:04On the other hand, something that Neils said earlier rouses a thought in me.
19:12He said that we can put in the knowledge of a world-class expert, and that's indeed true.
19:17We can put in the knowledge of a world-class expert, provided it fits into the format that the present programs allow,
19:24namely of the kind, if this is true, then do that.
19:29But more general kinds of knowledge that are used in a sort of vaguer way are sometimes harder to put in.
19:36Say, in the medical diagnosis area, at the beginning of the program we had a simple example of medical diagnosis.
19:42There were several systems developed, and yet there were problems, in fact, in using that.
19:47What were those problems?
19:49Well, I'm not sure what were the problems of those specific systems.
19:53However, to take a medical example, one can make a definition that a container is sterile if all the bacteria in it are dead.
20:03However, this piece of knowledge is not used in a special way to determine whether a container is sterile by examining each of the bacteria.
20:12It's used as part of more general facts, like if you heat the container enough, then the bacteria will be all killed,
20:19and the fact that if the container is not sterile and you empty it out onto a culture medium, then you'll get colonies of bacteria.
20:26And the still further fact is important, that if you leave the bacteria in food, then they will spoil it.
20:34Niels, where do you see the future practically going in AI?
20:37What's the biggest potential market, if you will, for these kinds of applications?
20:41Well, it seems to me that expert systems, as they have been going along, probably represents a rather large market.
20:46And we'll continue to develop systems that are less brittle over the next 5, 10, 15 years.
20:51Another very important application is computer programs that are able to converse with humans in English, everyday, ordinary language.
21:00And that's going to make computers accessible to a much wider variety of people.
21:05We have just about a minute. I like your phrase, brittle. Explain more what you mean by a brittle system.
21:10Well, it has to do with being able to reason about the context, I think, in which a particular discussion or conversation with the expert system is taking place.
21:20If the information that's needed to reach a certain conclusion is there and is there explicitly, then the system is usually able to come up with some reasonable answer.
21:28If it has to have what you might call common sense knowledge, knowledge that all of us learn by the time we're 10 or 15, then it has a good deal more difficulty and, as a matter of fact, fails in many of those cases.
21:38So, ironically, common sense is the most difficult thing.
21:41There are very few people who are willing to pay for putting common sense knowledge into a computer, whereas it is interesting and commercially important to put knowledge about a nuclear reactor into a computer.
21:51Okay, gentlemen, we're out of time. Thanks very much for being here. Thank you for joining us on this edition of the Computer Chronicles.
21:57The Computer Chronicles.
21:58The Computer Chronicles.
22:26Micro Focus, creators of visual programming tools for software development,
22:41is pleased to provide major funding for the Computer Chronicles,
22:45the story of this continuing evolution.
22:56Random Access is made possible by a grant from Popular Computing, a McGraw-Hill magazine.
23:22I'm Susan Bimba, sitting in for Stuart Chaffee,
23:25and the Random Access file this week, the Commerce Department has made a move to stop the export of technology
23:30to countries allied with the Soviet Union,
23:32hitting the Digital Equipment Corporation with tough new restrictions on its exporting practices.
23:37Now DEC has to get individual export licenses for each sophisticated computer.
23:42It ships to West Germany, Austria, or Norway.
23:44Government officials say in these countries, computers can be easily diverted to the Eastern Bloc countries.
23:50Several months ago, one of DEC's computers was intercepted en route to the Soviet Union.
23:55But the company was found to be blameless in the incident.
23:58Despite a strong lobby by the high-tech industry,
24:01President Reagan has given the Pentagon power to advise the Commerce Department
24:04on computer and high-tech exports to all foreign countries.
24:08The Pentagon already had the right to advise about applications for exports to communist countries.
24:13The Pentagon feels its input would help cut down on the number of computers
24:17that get diverted from non-communist countries to communist ones.
24:20The National Semiconductor Corporation could be banned from making chips for government projects.
24:26Early this month, the company admitted that over a three-year period,
24:30it inadequately tested chips that were used on many military projects.
24:34Now the Defense Logistics Agency is investigating National Semi,
24:38trying to find out exactly who's responsible for lying to the government about the testing.
24:42Despite these problems, National Semiconductor is reporting a profit of $15.4 million
24:48for the quarter ending March 4th.
24:52IBM is reporting big demand for its little computers.
24:55The company says this year its shipments of the PC will be triple what it was last year.
25:00It's estimated that IBM will ship more than 2 million of its PCs, PC Juniors,
25:05and other desktop IBM computers combined.
25:07And now that IBM is making its own portable computer,
25:11it looks as if Compaq, a maker of IBM-compatible portables, will be lowering its prices.
25:16So far, no date has been given for the price cut,
25:19but industry insiders say it will probably be around the beginning of April.
25:23Right now, the Compaq portable sells for $2,995.
25:28That's $200 more than the IBM portable.
25:31With a projected $500 cut, the Compaq will sell for $300 less than the IBM portable.
25:36Atari Incorporated is making cuts, but it will be employee cuts.
25:41The video game maker says it plans to lay off about 300 workers.
25:45At the same time, an Atari spokesman says there are plans to expand the sales and engineering staffs by about 100.
25:51Last year, the company lost over $500 million,
25:55laid off more than 2,000 of its Bay Area workers,
25:58and moved part of its operation to Taiwan and Hong Kong.
26:01Well, last week, we told you the Japanese Trade Ministry may introduce a bill
26:06that would reduce protection of U.S. software in Japan.
26:09This week, the U.S. government is warning that if Japan's parliament ratifies the bill,
26:14the U.S. will retaliate in kind.
26:16If approved, the bill would limit copyright protection to U.S.-made software,
26:20making it necessary for U.S. software owners to license their programs to Japanese makers.
26:25Well, a little closer to home, our software reviewer, Paul Schindler,
26:30has some information for blackjack buffs who don't have money to burn.
26:33If you gamble in Nevada or Atlantic City, that's a familiar sight, the shuffling of blackjack cards,
26:44usually accompanied by the sight of your money being swept in by the dealer at the end of the round.
26:50Well, if you're tired of losing, there's a computer program for you.
26:53It's Ken Houston's Professional Blackjack.
26:56Now, it's pretty rare when the writer gets top billing when they name a computer program,
27:01but Houston deserves it.
27:03He's a former official of the Pacific Stock Exchange, and now he's a full-time gambler.
27:08Of course, some people would say that isn't much of a switch.
27:11But anyway, on a recent trip to Las Vegas, Houston was kicked out of five casinos
27:15after winning more than $3,000.
27:18Now, most computer blackjack games will just play blackjack with you,
27:23but the Whole Earth Software Catalog and Review says Houston's Professional Blackjack
27:28will teach you how to play the game and win using various point-counting methods.
27:33The graphics are crisp.
27:35The table is green.
27:36The card's realistic-looking.
27:38And the sound is good, too.
27:40You can hear the cards hitting the table.
27:42Another use of sound, the program advises you on betting and card play,
27:47and it beeps at you when you goof.
27:48Ken Houston's Professional Blackjack is $70 from Screenplay in Chapel Hill, North Carolina.
27:55For Random Access, I'm Paul Schindler.
27:58Next week, Paul reviews Volkswriter Deluxe.
28:00And that's it for this week's Random Access.
28:02I'm Susan Bimba.
28:04Random Access is made possible by a grant from Popular Computing,
28:09a McGraw-Hill magazine.
Recommended
41:16
|
Up next
8:09
20:32
4:46
22:13
5:06
8:20
32:10
3:21
0:17
4:12
8:30
19:42
2:02
32:53
20:10
2:11
Be the first to comment