Report on the Fuzzy versus Conventional Control Debate Daniel Y. Abramovitch and Linda G. Bushnell highlight of the 1998 IEEE Conference on Decision and Control was a debate on “Fuzzy versus Conventional Control” held at one of the plenary sessions. Organized under the auspices of the CSS Membership Activities Board History Committee (Dr. Linda Bushnell, Chair), the debate featured Professor Lotfi Zadeh of the University of California at Berkeley taking the side of fuzzy control and Professor Michael Athans, recently retired from MIT, arguing the merits of conventional control. The debate activities consisted of three parts: the dinner beforehand, held Tuesday evening; videotaped interviews with Athans and Zadeh, requested at the dinner and conducted Wednesday morning; and the debate itself, held Thursday morning. Videotapes of both the debate and the interviews are available via the IEEE History Cent er ( UR L : h tt p :// w w w. i e e e . org/ history-center/) for a small fee to cover the cost of reproducing and sending the tapes.
Science Department at the University of Nova Lisboa in Portugal, who does fuzzy linear programming for decision support systems. (At that time she was doing a postdoc with Zadeh.) More relevant to the debate was the posturing of the participants during the dinner, akin to armies maneuvering for position before a battle. (As Sun Tzu says, the battle is won before the armies take the field [ 11.) Thus, the maneuvering took the form of trying to cast the scope of the debate. Athans’s tack was clearly to keep the debate narrowed to one of conventional versus fuzzy control for problems in which either is applicable. Zadeh, on the other hand, was clearly interested in broadening the debate to include task control-areas that tend to require rule-based control. Although the discussion was very courteous, the participants’ intent was ob-
utes per question, and as there were four panelists, each panelist would get at least one but not more than two questions. The panelists had been chosen so as to have equal representation from the conventjonal control side (Professors Karl Astrom and Gene Franklin) and the fuzzy control side (Professor Reza Langari and Dr. Demitre Filev). The participants would then have three minutes each to give closing statements. After dinner, the panelists met to go over their questions and make certain that a fair spread of questions would be asked. We also agreed to start the session 15 minutes early, at 8: 15 a.m., to allow afew more minutes for the debate. The next morning we were able to interview Zadeh and Athans and capture them on videotape. Gene Franklin generously agreed to serve as the interviewer,
vious: to choose the playing field that
and the authors each made videotapes. A
would favor their side of the debate. It was especially interesting to note that Zadeh was already choosing not to contest a key Predebate Activities point that Athans was expected to hamSeveral interesting sidelights sur- mer away at: that for any standard control rounded the dinner and the interviews. problem, one could always outperform First, it became apparent that Athans and fuzzy control using conventional-albeit Zadeh are close friends. Lotfi Zadeh was a quite sophisticated-control. Instead, member of Michael Athans’s thesis com- Zadeh intended to focus the debate in armittee at UC Berkeley and actually helped eas that are hard to model and quantify, get Athans into the optimal control field for which heuristics play a key role. In that by translating Pontryagin’s papers from regard, he clearly was hoping to convince Russian for the Berkeley control crowd. the attendees that fuzzy control provides a Furthermore, it was through Zadeh that means of integrating rule-based control Athans met his wife, Rita Almeida with conventional methods. Ribeiro, now a professor in the Computer The debate format was also finalized that evening. There would be 15-minute D.Abramovitch is with Hewlett-Packard opening statements by each of the particiLaboratories, 1501 Page Mill Road, M/S pants, followed by a question-and-answer 4U-12, Palo Alto, CA 94304-1126, USA, session in which a panel member would
[email protected]. L. Bushnell is with ask a question of one of the participants US Army Research Office, P.O. Box and they would have two minutes to re12211, Research Triangle Park, NC spond. The other participant would then 27709-221 1 USA, bushnell@aro-emhl. have one minute to rebut. As this would army.mil. require a minimum of three or four min-
fascinating aspect of the interviews was the history each of these men represented, and how the criss-crossing paths of careers and lives had had their effect on what we now simply think of as control theory. As mentioned earlier, it was Lotfi Zadeh who translated the Russian optimal control papers for his cohorts at Berkeley, and Michael Athans happened to be a student there. Zadeh spoke Russian because he spent the first 10 years of his life in what is now known as Azerbaijan, at that time
88
Michael Athans and Lotfi Zadeh “wrestle” at the Awards Banquet.
IEEE Control Systems
certainty (which he liked). In fact, he said he often found it hard to find the connection between fuzzy control and the principles of fuzzy logic as developed by Lotfi Zadeh.
Zadeh’s Statements
Michael Athans making his opening remarks.
part of the Soviet Union. The interviews set a good backdrop for the debate and are available from the IEEE History Center on the same tape as the debate.
lo^ Zadeh making his opening remarks.
rules, it was hard to say much about the performance or stability of the system. He also noted that this first set of fuzzy-rules-based controllers implement a human’s control law and therefore have a built-in assumpThe Debate tion that the human is a good controller. If The debate itself opened with a general that were the case, he said, control theory welcome and introductions by Professor would never have been developed. It is preDouglas Birdwell, General Chair of the cisely because the human is a poor control1998 CDC, followed by opening state- ler in many situations that we need control ments from Athans and Zadeh, which are theory. Thus, he claimed that fuzzy control summarized in the following sections. is bad for high-performance feedback control problems. The second generation, said Athans, Athans’s Statements It was obvious that Athans had come was oriented around overcoming such obquite prepared to tackle the specific ques- jections and focused on using fuzzy logic tion of whether fuzzy control could out- to “interpolate”between several good conperform conventional control in well- ventional control designs and then find a known control problems. He stressed of- Lyapunov function for the whole system. ten and strongly that, in the end, most such He claimed that this was ad hoc at best, perproblems could be solved with much formed poorly, and in fact could be outbetter results from a conventional control done by most gain-scheduling methods. perspective. He broke down fuzzy control During his closing remarks, he brought into two categories: home his final point with an analogy, First generation (Mamdani): rule- showing a drawing of a Ferrari and, next to based feedback control with the it, a drawing of a donkey. He asserted that rules being interpolated by fuzzy fuzzy augmentation of conventional consets, and trol was akin to having this Ferrari (conSecond generation (Takagi-Sugeno): ventional control) speeding down the fuzzy augmentation of crisp control highway and claiming that by having the laws. donkey (fuzzy control) get out and push, He assertedthat first-generationsystems the Ferrari would go slightly faster. were only appropriate for toy class probOne of Athans’s main objections to lems. Furthermore, he said that since no in- fuzzy control was that it does not use herent explicit model is used to derive the fuzzy logic’s method of capturing un-
June 1999
In his opening statement, Zadeh said that Athans had painted quite a picture, but one in which he could not recognize fuzzy control, and that the picture was not connected to reality. He immediately noted the large number of books on fuzzy control, including the recent one by Passino and Yurkovich [2], and stated that if the audience members took it upon themselves to read these books, they would see that many of the issues Athans had brought up were resolved by these texts. Stating that he wanted to take a broader view, he asked the question, “What is fuzzy control?” and then produced a slide listing a series of papers he and others had written over the years, starting with his 1965 paper introducing the concept of fuzzy sets [3] and ending with the Mamdani and Assilian paper [4] that is considered the first true fuzzy control paper. He said that the principal concepts in all his early papers were linguistic variables and fuzzy if-then rules, and that the “seminal” paper by Mamdani and Assilian applied these concepts to the control problem. Drawing distinctions between crisp control and fuzzy control, which are reproduced in Table 1, Zadeh emphasized that fuzzy logic compresses data by using linguistic variables, and that this is what humans do so well. He said that fuzzy control accepts the fact that a solution may not always work for every situation, but is happy with a solution that works many, many times. He drew a parallel to the computer science paradigm where it is understood that many practical and useful
Table 1. Zadeh’s delineation of fuzzy versus conventional control.
I
Crisp Control
Differential equation based
Rule based
Start with system equations, goals, and constraints
Start with human solution
Provides certain
Adds to those tools
tools
I I
I 89
solutions do not work all the time in every situation. Showing a sanitized quote from Professor Rudy Kalman that vigorously opposed the concepts of fuzzy logic, he stated that he and Kalman had been riding on the same train for a while and that he (Zadeh) had gotten off, but that Kalman was still on that train. He added that many others were still on that train. Zadeh then spoke about all the applications that were being completed using fuzzy logic and gave a count, based on an Internet search, of the growing number of papers that are based on fuzzy control. He posed the following questions to the audience: “Do you really think that all these people are stupid? Do you think they do not know what they are talking about?’ He then answered, “No, I think not.” He strongly urged that control systems people get away from thinking about what he called “set point control” and get more into the “task control” of replacing human functions. He used driving a car as an example of task control, repeatedly pointing out that humans drive a car “with no measurements and no computation.” He stressed that fuzzy control is not a replacement for but an addition to conventional control.
The Question-and-AnswerPeriod In the question-and-answer period that followed, the fuzzy control panel members asked questions of Athans and the conventional control panel members asked questions of Zadeh. As time remained, each participant fielded one additional question. Below is a synopsis1 paraphrasing of the questions and answers. As this report “compresses data” using “linguistic variables,” there may be some loss of the original meaning; however, the videotape available from the IEEE History Center provides an accurate record. Franklin, addressing Zadeh, stated that in Zadeh’s talk and in many papers it seemed that fuzzy control was a methodology for implementing a control scheme when one already had a scheme in their head. Did this make it something of a programming language? Zadeh replied that there was much more to fuzzy control than that. He said that one started with the fuzzy rules in their head and then used the “calculus of fuzzy if-then rules” to implement them. So he agreed that in that sense it was a bit like a programming language. However,
90
there were many books and papers that show how the field was now moving toward automatic rule generation. Athans said, “I agree with Gene.” He elaborated by saying that he did not see a normative, descriptive process in existence for fuzzy control design. Filev noted that many of the applications of fuzzy control were task control systems rather than set point control problems. He asked what Athans’s feelings were about fuzzy control in this context. Athans stated that, generally, his remarks were limited to fuzzy feedback systems. He seemed to say that as one moved closer to the man-machine interface, he had fewer problems with fuzzy techniques. However, he thought that most of these applications threw away any sense of methodology in favor of a fuzzy interpretation. Zadeh’s response was to urge the audience to look at the statistics on the numbers of papers and books out there. Athans looked at a small set of papers. He said the field was moving toward task-oriented control. Classical feedback theory had little t? say about this. Astrom stated that while he had not read all the papers on fuzzy control, he had read quite a few of them. In these papers, there was something quite mysterious, referring to the whole notion of model-free control system design. Did Zadeh really believe that one could develop effective control systems without any use of differential equations? Zadeh emphasized that he did not regard fuzzy control as a replacement for classictl control, but an addition to it. When Astrom requested a clarification on the model-free assumption, Zadeh said that there was no differential equation model, but rather a fuzzy model in the designer’s head. Athans agreed that there should be a mathematical model. He referred to the thousands of papers and applications as “a waste of human effort.” He had a problem with folks luring students into doing control by showing them some simple fuzzy feedback rules. He did not think that they were getting a good idea of what control was all about. Langari admitted that many of the applications of fuzzy control are in loworder, overdamped process control applications. He asked Athans if he would agree that fuzzy control methods provide a semantically transparent way for people to address control problems.
Athans’s answer was largely that perhaps this was true for some simple problems, but how could it possibly be extended to large-scale or multivariable systems? It all fell apart then. Telling someone that if they know a few simple fuzzy rules they know control design, he said, is analogous to teaching someone how to use a stethoscope and to take blood pressure readings and having them think they have become a brain surgeon. Zadeh again took issue with Athans trivializing all the books and papers on fuzzy control and implying that all those people are “stupid.” He said he believed that in the not-too-distant future, fuzzy control would be a standard part of a controls curriculum. Franklin said that everyone remembers the advent of optimal control in the form of the LQG (Linear Quadratic Gaussian) controller and all the hype that surrounded that. Could Athans compare and contrast the hype surrounding LQG in the 1960s with that surrounding fuzzy control in the 1990s? (At this point, nearly everyone in the room was laughing, since Athans made his name in optimal control in the 1960s.) Athans (after some hemming and hawing) said that he thought the hype for LQG was a necessary step. Although there were some problems with LQG at the beginning, those have been resolved. He showed an LQF (Linear Quadratic Fuzzy) slide on which he said that although one can cast this problem, it is intra c ta ble . T h e diffic ulty was not philosophical; it was mechanical. There was no practical way to do it. Zadeh said that he thought LQG and fuzzy control were two different beasts. He thought that LQG was a useful procedure in a very narrow set of problems, whereas fuzzy control addresses a much wider set of problems. He predicted that fuzzy control would continue to grow and that more books, papers, and applications would show up. (At this point, Athans noted that “it grows like a plague,” prompting a quick laugh and reply from Zadeh.) h t r o m remarked that Zadeh’s paper count slide showed an exponential growth in the area of fuzzy control, and things with an exponential growth tend to flatten out at some point. When did he see the growth in fuzzy control flattening out? Zadeh first expressed his respect for Astrom, noting that he had always been fair. He then said that fuzzy control will IEEE Control Systems
continue to grow in importance, suggesting that its name may disappear, but that it would move into the mainstream. This may take some time. Fo l l o w i s$ i w question-and-answer periodi4he participants made their 3minute closing statements (summarized in the earlier statement sections). The debate then ended with a handshake between the two old friends.
References [l] “Therefore the skilful commander takes up a position in which he cannot be defeated and misses no opportunity to master his enemy. Thus a victorious army wins its victories before seeking battle; an army destined to defeat fights in the hope of winning.” S. Tzu, The Art of War, S.B. Griffith (trans), Oxford University Press, 1971. [2] K.M. Passino and S. Yurkovich, Fuuy Control, Menlo Park, CA: Addison-Wesley Longman, Inc., 1998. [3] L.A. Zadeh, “Fuzzy sets,” Informarion and Control, vol. 8, pp. 338-353, 1965. 141 E. Mamdani and S. Assilian, “An experiment in linguistic synthesis with a fuzzy logic controller,” Int. J. Man-Machine Studies, vol. I , no. 1, pp. 1-13, 1975.
The Global Positioning
(GPS) The PROCEEDINGS OF THE IEEE wil start the year with a very special issue devoted to GPS. Our January 1999 issue is devoted to all aspects of this pivotal technology! Not since the establishment of longitude and latitude conventions has any innovation played such a key role in terrestrial navigation on land, water or in the air!
Assistant Research Engineer UC Berkeley Institute of Transportation Studies $66,400 - $84,200 (depending upon qualifications) The California Partners for Advanced Transit and Highways Program (PATH) has been leading the way in Intelligent Transportation Systems (ITS), Advanced Vehicle Control and Safety Systems (AVCSS), and Automated Highway Systems (AHS) research, since PATH’S founding in 1986. The PATH program is managed by the Institute of Transportation Studies of the University of California at Berkeley, with program headquarters at the University’s Richmond Field Station.
Job # ITS-91 Duties: Conduct research for advanced vehicle control and safety systems’ (AVCSS) design and evaluation projects. This includes applications in the areas of emerging driver-assist systems and Automated Highway Systems (AHS) as well as air or marine vehicle systems. Research will address system-level performance, including safety, productivity, comfort, environmental impact and cost metrics. Will be responsible for applying existing tools and methods in the research program, and advancing the development of the technical framework for tool development and use. Also responsible for developing additional models and interfaces, leading focused research teams, and the preparation of reports and technical papers. Qualifications: Ph.D. in Electrical Engineering or a related field, with a strong research record. Strong technical communication skills. Strong analytical skills and the ability to apply them to the design and evaluation of control/embedded systems. Demonstrated ability to initiate and sustain collaborative research programs. Understanding of AVCSS and AHS application domains is desired Job # ITS-92 Duties: Conduct research on simulation, verification, or automated code generation tools for embedded systems (including Intelligent Transportation Systems). Transition tools into use in existing system design, or prototype development projects. As the leader of a small research team, work on the further development and support of the SHIFT simulation language for dynamic networks of hybrid automata (see http://path.berkeley.edu/SHIFT/index.html),and traffic microsimulators using this language (see http://path.berkeley.edu/ SMART-AHS/index.html). Actively participate in presentations and the preparation of reports and technical papers. Qualifications: Ph.D. in electrical engineering, computer science, or mathematics with a strong research record. Strong technical communication skills and programming skills. Demonstrated ability to develop and apply simulation or other embedded system design tools. Experience in working with transportation simulations is desired. Initial appointment for 1 year. Reappointment subject to availability of extramural funds.
and receive 12 issues for $23.00 IEEE Member Price: $23.00 List Price: $400.00 Subscription No. 500-501 outuk VU, Can& mdndMerK0pkmr IrwS25shqqnngondhlAbng Call 1 800 678-IEEE or 1 732 981 -0060 mail:
[email protected] ‘Emg
h InstiMeof fchicol Md €!ern& fngkeecr 1 ~ . 445 Hoer Lme, W Box 1331, Rwotowm(. NI d855
%%!3
To apply, please send curriculum vitae, the name/address of three references, and a separate sheet of paper (optional) with your name, your gender, your ethnic identity and where you heard about this position, by 8/13/99 to: ITS Job #Search Committee, 109 McLaughlin Hall, UC Berkeley, Berkeley, CA. 94720-1720. The University of Calijornia is an Equal Oppor?unity/AffirmutiveAction Employer
