Computers ind. Engn@Vol. 13, Nos i-4, pp.15-20, 1987 Printed in Great Britain. All rights reserved
THE
ROLE
OF
THE
DEVELOPING
BY
Douglas
0360-8352/87 $3.00+0.00 Copyright C 1987 Pergamon Journals Lid
INDUSTRIAL ENGINEER IN EXPERT SYSTEMS
S. W a t t s
ano
Hamed
Kama!
Eldin
InOustrial Engineecing ~ Management Oklahoma State University
Stillwater,
ABSTRACT The moat popular area of Artificial Intelligence application today is in expert systems. This paper contains a discussion of expert systems, otherwise known as knowledge-based systems and knowledge systems. The principal components of an expert system, and the evolution of expert systems are presented. The suitability of a task to an expert system is proposed. When a task is suitable for an expert system application, the system must be developed by a knowledge engineer. The methodology that the knowledge engineer must go through to develop an expert system is demostrated. Industrial engineers have formal training in many areas which can be useful when assumming the role of knowledge engineer. These areas of industrial engineering and how they are beneficial is discussed. What the future may hold in store is also pondered.
74378
MJ4C~~]I~e~TMA] ARSAY~TI~Im~'T .,
T ~ t e d ~r~ I16], sourceSperryCorp.
Figure I: DESIGN CYCLE FOR SOME EXPERT SYSTEMS system as: "a computer-based system that searches for a solution within a set of statements or a body of knowledge formulated by the experts in a specific area." A very simple definition is presented by Hayes-Roth [2], in which he states that an expert system "simulates expert human performance and presents a human-llke facade to the user." Today, expert systems are a very active development area. For example, Japan has a ten year research project to develop the fifth generation computer as an expert system. {4 - II] The Department of Defense has identified artificial intelligence as one of the ten most critical technologies to pursue in the remainder of this century. [5, 6] The reasons for the fast growth of expert systems are due to several fundamental problems. Tou [3] summarizes these problems well, as follows: * preservation of knowledge * proliferation of knowledge * dissemination of knowledge * application of knowledge
BACKGROUND The term artificial intelligence was first introduced by John McCarthy in 1956. [I] Since then, several areas of application and research have developed. These areas are as follows: * * * * *
OK
knowledge-based systems vision systems speech recognition natural language processing automatic programming
Expert systems are based on domain-specific knowledge. That is, knowledge in a specific field. This knowledge, in the form of specialized know-how, in combination with a fast ~nd consistent symbol processor can perform as #ell as a human expert. In order to get the ~nowledge from the experts and into a form of specialized know-how that the computer could understand, the area of knowledge engineering came to be. Matching the performance of experts is emphasized in knowledge engineering. A few knowledge engineering applications were underway beginning in the mid 60's. By the end of the folowlng decade, several projects have accomplished significant results. [2] Some of the more noteworthy expert systems which have been developed, when they were started, and the manyears required to develop them are shown in Figure 1.
The advantages associated with an expert system are more reason for their development. The advantages of an expert system are as follows: 1. can cope with uncertainty in data and knowledge [12] 2. can explain how the current answer was arrived at [12, 13] 3. can modify the current knowledge base when something new is learned [12] 4. can be made complete [12]
EXPERT SYSTEMS Many authors have defined expert systems ~ffferently. However, Tou [3] defines an expert
15
16
Proceedings of the 9th Annual Conference on Computers & Industrial Engineering 5. can offer multiple conclusions, ranked by some measure of confidence [12] 6. can preserve knowledge forever [5] 7. provide a means to employ know-how where it is needed, when it is needed, and at great speed [5] g. tan be easily distributed [i4] 9. can prevent "pockets of knowledge" or "knowledge bottlenecks", where only one employee may know how to do something [14, 15] i0. can be applied to many areas II. makes timely utilization of new scientific and technological knowledge [3]
Because of the reasons given above, many expert system applications are being developed or are in the process of being developed. The most successful expert systems to date are the ones used for diagnosis. [I] This is because their rules are based on a well-understood diagnostic process. Although the applications and advantages of expert systems are many, there are some drawbacks. These disadvantages are as follows: * *
*
*
*
* *
the expert system may exhibit gaps in knowledge at unexpected times [I] expert systems are unlikely to have complete, clear functional specifications [i] knowledge engineers cannot reliably predict their behavior in situations not tested [I] expert systems cannot recognize when a problem is outside its domain of expertise [6] the expert system has no independent way to check that its conclusions are reasonable [6] expert systems can't learn from experience the cost for a commercially developed expert system can range from $50,000 to SlO0,O00 [16]
Currently, the major bottleneck in the development of expert systems is the knowledge acquisition process. [17] One reason ~or this is due to an acute shortage of knowledge engineers, and another is because the knowledge acquisition process is very expensive. [17] The high cost can be contributed to the fact that, even with the knowledge engineer using special tools, considerable skill in artificial intelligence programming is still necessary to create an expert system which can perform a nontrlvial task. [18] However, with more people claiming the role of knowledge engineer, and with more tools being developed everyday, expert systems are currently becoming more popular everyday. EVOLUTION OF EXPERT SYSTEMS The principle components of an expert system are the knowledge base, and the inference engine. The knowledge base contains the general facts and rules of thumb, or hueristic knowledge. The inference engine interprets the knowledge in the knowledge base and uses it to arrive at a solution. Because the inference engine and knowledge base can be separated, this permits one inference engine to be used with different knowledge domains. When the inference engine is separate, it is known as a shell, or a tool for expert system development. Up until the late ig7Os, all expert systems had to be custom built because there were no
!!Figure 2: EVOLUTION OF EXPERT SYSTEMS
tak,,fr~ I?l
tools available. At that time, however, shells were made available along with other universal tools, and ready-made systems. [7] Custom made expert systems were started in the mid 60s and literally cost millions to develop [7]. They are not as expensive now, but it is still an enormous task because the expert system is developed using no tools. This means that the knowledge engineers must be concerned with two things: knowledge acquisition, and knowledge manipulation. The shells, or shell systems, are just the expert system with only the representation and reasoning components. This allows the knowledge engineer to be concerned with only the acquisition of the knowledge, and perhaps a few modifications to the shell. Using a shell to develop an expert system is roughly analogous to using a computer spreadsheet program to perform complex calculations. The person using the program is just concerned with getting the information into the computer so the computer program can manipulate it. Ready-made expert systems are systems which are aimed at specific knowledge domains. These systems contain some basic natural laws and are equipped to deduce problems from given test results. ~or example, an expert system which performs engine diagnosis for an automobile mechanic could be used by other auto mechanics working on the same type of engine. This possibility paves the way for companies to develop their own expert system and then capitalize on their efforts by marketing their development. It is estimated that this practice will become commonplace within the next ten years, [71 The universal tools, or general purpose construction tools, basically simplify the acquisition of new domain knowledge and allow the expert system to perform complex calculations in addition to doing reasoning. One example would be a tool which could be used for a genetic research expert system as well as an automated production plant expert system. Because the shells and tools are relatively new, the distinction between them is not real clear. In this paper, however, a tool will refer to both shells and universal tools, as in the above example. Table [ gives an example of the many tools which are now available for large computers and personal computers. This number is sure to increase rapidly in the coming years. Cross [181 discusses characteclstics of some recently developed tools. DEVELOPMENT CRITERIA Determining if a expert system is a very this step, management task to be solved 120~,
task is suitable for an critical first step. In identifies the potential or management may iden-
Watts and Eldin: Developing expert systems TABLE i:
AVAILABLE TOOLS FOR COMPUTERS [19]
LARGE COMPUTERS Advisor ART HPRL IKE KEE KES Knowledge Craft Piton Reeveal S.I Verson 2 TIMM
PERSONAL COMPUTERS ~SP Advisor ExperOPS5 Expert-Ease Exsys Insight 2 KDS M.I and M. la MicroExpert Micro KES NaturalLink Personal Consultant RuleMaster TIMM-PC
tify several corporate problems. [21] Then, they must consult with the knowledge engineer, who determines which task (if given more than one problem) is suitable by answering the following questions: 1. Is the task too complex to use traditional programming techniques? [21, 22] 2. Is the task of narrow focus, and primarily cognitive? [5, 12, 17, 21, 23, 24] 3. Is there at least one expert in the task domain available and willing to contribute his time and knowledge? [12, 17, 20 - 22, 24] 4. Can the methodology of the experts expertise be expressed? [22, 24] 5. Can the expert system be built with an existing tool? [20] 6. Does the task have a high payoff? [5, 12, 17, 20, 21] The first question must be asked because if the task Is simple enough to use traditional programming to solve, then this would be the best choice since it is much less costly to develop. The second question is asked to make sure the task is of a narrow domain. This ensures that the task does not require knowledge from a very large number of areas. Question three states that there must be experts who are accessible and willing to spend their time contributing to the project by working with the knowledge engineer. If experts are so few that none can find the time to work with the knowledge engineer, then the expert system cannot be developed. The fourth question states that the knowledge engineer and the expert must be able to encode the knowledge domain into usable form for the computer to understand. In order to do this, there must be some methodology that the expert uses, and the knowledge engineer must extract this. The reason for the fifth question was already mentioned. That is, if there Ls not an exlst[ng tool already in existence that can be used, the knowledge engineer cannot develop the expert system without a progra,nmer, and the price of such a system will be very expensive due to costs and time. The last question requires that the knowledge engineer consider all cost and time involved in the development of the expert system, and then estimate the expected cost for constructlJlg the entire system. This estimation should include the following costs: knowledge engineer's, design, development, computing, overhead, travel, expert consultation, and maintenance. Me should then estimate the
17
expected benefits from the expert system. If the payoff is high, this means the present worth of the benefits exceeds the costs, and the project should be taken on. The task of determining if a task is suitable for an expert system is critical, and should not be taken lightly. The above was a fraction of what can be considered before committing to development. For a detailed discussion, the reader is referred to Prerau [21]. DEVELOPMENT PROCESS There have been several methods of Expert System development presented in the past. [2, 5, 12, 14, 17, 20, 22, 23, 25 27] The most agreed upon steps used for developing an expert system are as follows: i. 2. 3. 4. 5. 6.
identification [2, 17, 22, 23, 25] conceptualization [2, 23, 25] formalization [2, 17, 22, 23] implementation [2, 18, 23] testing and evaluating [2, 17, 22, 23, 25] maintenance planning [12, 20]
Identification is the first step once a problem is determined suitable for an expert system. In this step the the expert or experts are hired, and the problem characteristics are identified after becoming familiar with the problem. In addition, a specific development plan should be prepared. Putting the knowledge on paper in the form of English sentences can be a big help during this step. During conceptualization the knowledge is organized, the concepts which can be used to represent the knowledge are found, and the expert system prototype is developed. This is done by being aware of the different knowledge representation schemes used in the different tools which are available. Another interesting way to approach this problem is suggested by Chandrasekaran [28]. He suggest that tools not only be classified on the means of encoding knowledge and carrying out the task, but also on the generic task that they perform. No matter what method is used to evaluate the tools, the knowledge engineer must be aware of the latest tools available. When in the Formalization step, all of the available tools identified in the previous step are evaluated and one ks picked. In addition, additional knowledge is extracted [rum the expert. During the implementation step the knowledge is formulated into rules and encoded into the chosen tool. This requires many hours of consultation between the experts and the knowledge engineer. Also, the prototype system, including the user interface, is implemented. In the testing step the rules are validated by letting the experts test the system. The experts use sample problems and then make sure the system is solving the problem the way they would. After the testing, the knowledge engineer may return to any of the proceeding steps, based on the evaluation. It is through t!~is testing and evaluating step that the system is perfected. The final test, however, is with the end user. During the maintenance planning step actions are planned in order to keep the knowledge in the knowledge base up-to-date. It may be necessary to modify the user interface, or inference engine, to accommodate the changes.
18
Proceedings of the 9th Annual Conference on Computers & Industrial Engineering
INDUSTRIAL ENGINEERS ROLE When discussing the role of industrial engineers in developing an expert system, it must be emphasized that industrial engineers are users, not inventors of expert system methodology. That is, industrial engineers use what is available to create an expert system, they do not try to develop new languages or knowledge representation schemes. This is best left up to the computer scientist. When developing an expert system, industrial engineers rely on four main areas of their training: engineering economy, time and motion study, simulation, and production control. There are, of course, other areas which the industrial engineer can draw from, but these are the main ones. Engineering economy is important in order to justify the expert system development. It was mentioned in the criteria section that all of the costs for system development must be considered, along with the savings expected from the system. Then the system may be developed only if the benefits exceed the costs. This is known as a benefit/cost analysis. Industrial engineers are trained to conduct benefit/cost analyses. This is not a simple job which can be trusted to just anyone, and when an expert system with a high development cost is at stake, the company should make sure this analysis is conducted correctly. This includes such considerations as the current depreciation rates the company can use, the expense amount the company can take, the company's incremental tax rate, and whether the company will borrow funds for the development or use retained earnings. Eli of the above will have an effect on the after tax cashflows, which are then taken to the same point in time as the benefits using the companies minimum attractive rate of return. In addition, the industrial engineer knows to be consistent when doing the analysis in real dollars or actual dollars. [29] The above discussed the costs portion of the benefit/cost analysis. The dollar amount of the benefits is more difficult to determine. This fs where the industrial engineers training in time and motion study, along with engineering economy, can come in handy. Because industrial engineers a r e trained to time motions, they can time the activities which would be done away with when the expert system is implemented. This time savings can then be implemented into a dollar amount using wage rates and overhead amounts. Then, as done with the costs, the dollar amounts can be taken to the same point in time, and the benefit/cost analysis performed. The time and motion training can als~ hel l) the industrial engineer [n ~)bserving the procedures that the expert goes through when working through a problem. By observing the expert, taking notes, and knowin~ in what order the expert does what he does, the industrial engineer ,:an develop the ,nethodology to be placed into the expert syste~n. Simulation can help the industcial engineer in several ways. For one, working with expert system development tools can be compared to working with different simulation languages. When the industrial engineer has ~orked with one simulation language, it is easier to learn others by understanding the different ways the languages do the same thing. This similar method can be applied to the various expert system development tools. Simulation can a]s5 help the industrial engineer in the development process. For example, the procedures that industrial engineers
use to develop follows: I. 2.
3. 4.
5.
a
simulatation
program
are
as
MODEL DEFINITION - defining the model PROGRAM CONSTRUCTION - flowcharting what the program should do and writing it MODEL VERIFICATION - making sure the program works properly MODEL VALIDATION establishing the degree of comparability between the model and the real system RESULTS - running the program and evaluating the system based on the measures of performance
AS one can see, some of the steps discussed above have a great deal in common with the steps used in the expert system development process presented earlier. The purpose of production control is to effectively utilize resources in the production of goods in order to meet customer demand and mazlmize stockholder wealth. If the expert system can be thought of as the good, and the company as the customer, then this definition can also apply to the development of expert systems. In this case, the industrial engineers training in this area can be the strongest case for the industrial engineer working as a knowledge engineer. One approach that industrial engineers may learn to use in the design of a production control system is the traditional systems approach, or operations research approach. [30] This approach involves the following steps: i. 2. 3.
4.
5. 6. 7.
SYSTEM OBJECTIVES determining the system objectives SYSTEM DEFINITION structuring and setting the boundaries of the system SYSTEM COMPONENTS determining the significant components of the system COMPONENTS DEFINITION - studying the components in detail and defining their relation to the whole system COMPONENTS SYSTHESIS - synthesizing the defined components into the system SYSTEM TESTING testing the system against some performance criteria SYSTEM IMPROVEMENT improving lhe performance of the system by returning to steps 2 through 6, as needed
Once again, one can see that some of the steps discussed ~bove have a great deal in common with the steps used in the expert system development process. This approach can be used in the expert system development process in order to minimize overall project cost. This can be done using several methods, some of which are as follows: Critical Path Method (CPM), Project Evaluation and Review Technique (PERT), PERT with cost control aspects (PERT/COST), Graphical Evaluation and Review Technique (GERT), GERT with Queing system modeling in graphic form (QGERr), GERT with Resource (RGERT), GERT with Precedence (PGERT), and Venture Evaluation and Review Technique (VERT). These methods evolved from the Cantt chart and the Navy's milestone method. [30] The industrial engineer knows the differences in the methods, and which methods are best suited for different situations. This allows the project to be p l a n n e d and completed according to a schedule. There are many other areas of training from which the industrial engineer can draw from when developing an expert system. For example, oper-
Watts and Eldin: Developing oxpert systems atlons research can also be applied to the knowledge acquisition process. There may be variations in the actual response of an expert, and the given response during a hypothetical situation. The variations can be analyzed using gaming techniques of operations research. [31] FUTURE Professor Edward Felgenbaum of Stanford University, one of the leading researchers in expert systems, presented an interesting scenerlo of the future of expert systems during the First Artificial Intelligence Satellite Symposium. [35] He explained that a 1981 Defense Science Board "IMPACTS" Study revealed that artificial intelliKence was rated number 2 (advanced software technology was rated number I) of all technologies which will have an order of magnitude impact on the defense of the United States in the 1990s and beyond. He also presented the results of a similar study conducted by Iron Age, the trade magazine of the manufacturing profession. Fifty technologies were ranked in order of importance to the future of manufacturing with artificial intelligence getting the number I rating. In addition, Feigenbaum presented what the expected dollar sales of artificial intelligence related computer products will be, and that percentage of computer industry sales, according to A.D. Little Decision Resources, 1984. That information is summarized in Table 2 below. TABLE 2: Where the AI Market is Going
YEAR 1990 1995 2000
SALES(IN BILLIONS) 5 - i0 30 - 70 S0 - ii0
% OF COMPUTER INDUSTRY 2 - 4 % 5 - 15 % I0 - 20 %
A different future scenerio exists with computers which are designed differently from those presently in use. The computers of today are modeled after the design laid down by mathematician John yon Neumann in the 1940's. It calls for physical separation of the computer's memory and its processor, with a communications llnk in between. [32] Some researchers, however, are dropping conventional computer design and modeling their systems after what they believe the human brain does when it thinks. By using hundreds (possibly thousands) of individual computing units and linking them with many thousands more connections, these researchers try to simulate the brain's own broad and tangled neutron fields. These new systems are called neural networks. Some networks already show amazing humanllke qualities; they can adapt to change by adjusting their own circuitry, and they can learn from experience. [33] Maybe these new systems could someday replace todays expert
systems. SUMMARY Today it is easy to see how far artificial intelligence has come. There are production plants which turn out products made entirely by robots, computers that can speak clearly and understand most human speech, and expert systems which either assist in decision making, or make decisions. All of these achievements have not come easily; all required much research and hard work. The development of expert systems is no
exception.
19
The methodology presented ia this paper was meant to be simple. It is intended to be used for a wide variety of small problems which would not exceed a knowledge engineers ability. In addition, it is also intended to build on experience. By doing this, it allows the more experienced knowledge engineers to custom design their own methodology to some degree. Also, the main goal of this methodology is to extract the knowledge. This can be the most time consuming task for the knowledge engineer. Finially, this methodology is able to provide milestones. This is important in ~rder to assess the progress of the system and know if the project is ahead or behind schedule. Industrial engineecs are the most qualified to assume the role of knowledge engineers, not just because their educational background prepares them well, but also because [ndustrial engineers are people engineers. That is, they constantly work with people to get thiags done. This is one special attribute which industrial engineers possess, moreso than other engineering dlsplines. Many companies have already recognized this potential and formed knowledge engineering teams made up of industrial engineers. One example is a team made up of [ndustr[al engineers at General Dynamics, Ft. Worth Division, which is developing an expert system to be used in production. [34] As tools become available which make the knowledge engineers job easier, the task of knowledge extraction will still remain. Therefore, there will still be a need for knowledge engineers and a methodology to be used. In other words, there must be a ,~ethod to the ,nadness! REFERENCES [I] Denning, Peter J., "Towards a Science uf Expert Systems", IEEE Expert, Summer, 1986. [2] Hayes-Roth, Fredrick, "The Knowledge-Based Expert System: A Tutor [al", Computer, September, 1984. [3] Tou, Julius T., "Knowledge Engineering Revisited", International Journal of Computer and Information Sciences, March, L985. [4] F i s c h e r , Gerhard, "Symbiotic Knowledge-based Computer Support Syste:ns", Automatlca, November, 1983. [5] Hayes-Roth, Frederick, "KnowleJge-Based Expert Systems", Computer, October, 1984. [6] D'Ambrosio, "Expert Systems - Myth or Reality?", Byte , January, 1985. [7] S c h i n d l e r , Max, "Expert Systems", Electronic Design, January [0, 1985. [8] Treleaven, Philip C. and Isabel Gouveia Lima, "Japan's Fifth-Generation Computer Systems", Computer, August, 1982. [9] Furukawa, Osamu, and Syohei Ishizu, "An Expert System Eor Adaptive Quality Control", Int. J. General Systems, Vol. II,
1985. [I0] Davis, Dwight B., "Artificial Intelligence Enters the Mainstream", High Technology, July, 1986. [II] Connah, D. M., and C. ~. Fishbourne, "Intelligent Knowledge-based Systems", Journal of the Institution of Electronic and Radio Engineers, June, 1985. [12] Fisher, Edward L., "Expert Systems Can Lay Groundwork For Intelligent CIM Decision Making", Industrial Engineering, March,
1985.
[13] H a y e s - R o t h ,
Frederick,
Systems", Communications September, 1985.
"Rule-Based of
the
ACM,
20
Proceedings of the 9th Annual Conference on Computers & Industrial Engineering
[14] Freiling, Mike, Jim Alexander, Steve Messick, Steve Rehfuss, and Sherrl Shulman, "Starting a Knowledge Engineering Project: A Step-by-Step Approach", The AI Magazine, Fall, 1985. [15] King, Micheal S., Steven L. Brooks, and R. Michael Schaefer, "Knowledge-based Systems", Mechanical Engineering, October, 1985. [16] Parker, Richard, and Nicolas Mokhoff, "An Expert for Every Office", Computer Design, Fall, 1983. [17] Banares-Alcantara, R., D. Sriram, V. Venkatasubramanian, A. Westerberg, and M. Rychener, "Knowledge-Based Expert Systems for CAD", Chemical Engineering Process, September, 1985. [18] Cross, George R., "Tools for Constructing KnowledgeBased Systems", Optical En~ineerln~, March, 1986. [19] Myers, Ware, "Introduction to Expert Systems", IEEE Expert, Spring, 1986. [20] Harmon, Paul, and David King, Expert S~stems, John Wiley & Sons, Inc., New York, 1985. [21] Prerau, Davis S., "Selection of an Appropriate Domain for an Expert System", The At Magazine, Summer, 1985. [22] Haley, Paul, and Chuck Williams, "Expert System Development Requires Knowledge Engineering", Computer Design, February 15, 1986. [23] Arora, Jasbir S., and G. Bainziger, "Uses of Artificial Intelligence in Design Optimization', Computer Methods, March, 1986. [24] Gevarter, William B., Intelligent Machines, Prentice Hall, Englewood Cliffs, New Jersey, 1985. [25] Bobrow, Daniel G., Sanjay Mittal, and Mark J. Steflk, "Expert Systems: Perils and Promise", Communications of the ACM, September, 1986. [26] McDermott, John, "The Knowledge Engineering Process", Database Engineering, December, 1983. [27] Maher, M. L., D. Sriram, and S. J. Fenves, "Tools and Techniques for Knowledge Based Expert Systems for Engineering Design", Advances In Engineering Software, October, 1984. [28] Chandrasekaran, B., "Generic Task in Knowledge-Based Reasoning: High-Level Building Blocks for Expert System Design", IEEE Expert, Fall 1986. [29] Sullivan, William G., and James A. Bontadelli, "The Industrial Engineer & Inflation", Industrial Engineering, March, 1980. [30] Bedworth, David D., James E. Bailey, Integrated production Control Systems, John Wiley & Sons, Inc., New York, 1982. [31] Phelps, R. I., "Artificial Intelligence An Overview of Similarities with O.R.", J. Opl. Res. Soc., January, 1986. [32] Port, Otis, "Computers That Come Awfully Close to Thinking", Business Week, June, 1986. [33] Larson, Erik, "Neural Chips", Omni, November, 1986. [34] McCollom, N. N., Supervisor of Advanced Systems at General Dynamics, Fort Worth Division, Personal Interview, October 31, 1986. [35] Feigenbaum, Edward A., Randall Davis, Bruce G. Buchanan, and Mark S. Fox, "Knowledge-based Systems and Their Applications", The First Artificial
Intelligence Satellite Symposium, Sponsored by Texas Instruments, November 13, 1985. [36] Barr, Avron, and Edward A. Feigenbaum, The Handbook of Artificial Intelligence, Vol I & II, Heuristech Press, Stanford, CA, 1981. [37] Cohen, Paul R., and Edward A. Feigenbaum, The Handbook of Artificial Intelligence, Vol III, Heuristech Press, Stanford, CA, 1981.