An undergraduate curriculum in expert systems design or knowledge engineering

• Authors: Mary Micco
• Venue: CSC '87 Proceedings of the 15th annual conference on Computer Science
• Year: 1987

By way of introduction, Chatham College is a small liberal arts college that has obtained a grant to implement a curriculum in artificial intelligence even with the minimal equipment and resources that we do possess.We had been involved with preparing and teaching courses to familiarize undergraduate students with the field of artificial intelligence for several years. Last year we decided to develop a cluster of advanced courses that would turn out 'knowledge engineers'. There is a serious shortage of such skilled people in the Pittsburgh area. We planned for them to go immediately to Industry and begin working on expert systems, with a wide range of applications. These included but were not limited to diagnostic systems, financial modelling applications, and shop floor management. Under the terms of our funding we also encouraged students to become entrepreneurs by first developing and then marketing their own expert systems targeted for specialized markets.Students selecting this upper level cluster have already completed our core computer science requirements and have a good math background. One of the key factors in this project has been that we worked closely with the business community seeking their input into the curriculum planning process. They also provided us with the internship opportunities we requested and gave the students exposure to a much wider range of equipment than we could afford to provide in house.The program has been designed to run on IBM PC's with 512K. We purchased copies of Golden Common Lisp and Turbo Prolog as well as using an Altos with Xenix and C to get us started. We are currently negotiating for a grant for 2 Hewlett Packard artificial intelligence workstations with four terminals each. These cost about $50,000 each and are far beyond our price range. Two things are encouraging. Prices are falling very rapidly and the other is that NSF and other foundations are beginning to make instrumentation grants available to smaller institutions for the purchase of sophisticated equipment of this type. Our most difficult task has proven to be finding faculty to help teach these courses. We have had to draw on local industry to help with part time instructorsThe whole cluster consists of 6 courses and a 6 credit internship. The first introductory course. Artificial Intelligence, involves an overview of the field. The history of the movement from its roots to its current popular status are explored. We based the course on Patrick Winston's text Artificial Intelligence. Each of the major research areas is explored. The approach taken was one of studying the classic toy problems. This made it possible to show the difficulties that have been encountered at a practical level and then to review progress in the field, while permitting students to gain actual programming experience. For example, in Natural Language we introduce the Eliza program as an early example of an attempt to use the keywords approach. Students are required to produce a modified conversation. We also look at the Turing test. Drawing from Rodger Schank's book The Cognitive Computer we go on to talk about scripts, using the Restaurant Script as an example. Students are asked to put together a travel script in teams of four. This section is wrapped up with a discussion of natural language programs today.In the area of vision research, so critical in Robotics, we used the classic toy blocks problem, encouraging the students to key it in and modify it. We found that actually working with the code helped the students to gain a much more vivid appreciation of the difficulties involved. Winston has an excellent presentation on the subject. It was difficult to improve on it.Having introduced two specialized problems in the field, we move the students into the challenge of finding a general problem solving technique. At this stage they have no difficulty in accepting the fact that no such algorithm can be developed and that it is rather a matter of determining the rules of thumb or heuristics governing the specific problem domain with a sufficiently large knowledge base to cover all likely contingencies and options.To help to clarify this point, we delve into game playing strategies. Students are asked to write a tic tac toe game wherein the computer learns from its own mistakes. The first version they are asked to produce simply stores patterns of play in an array. The program searches for a pattern match. If one is found it will play the next move in the pattern string. If none is found a move is picked at random from those available. As it plays it keeps adding to the number of pattern strings it recognizes. The key element in this program is that the last byte in the pattern string contains a score for that particular pattern. The score for each winning pattern is incremented by one each time it is selected. Losing patterns have one subtracted from their score. Eventually when the score reaches zero that particular pattern will not be played again. The principle involved is that of reward and punishment. The patterns that win are reinforced by means of higher scores, while losing patterns are punished by having points taken away.The next refinement students are asked to add in, is an algorithm that will assist the computer to pick a better move not at random but with a set of clearly specified rules. This has proven to be an interesting and challenging exercise for our students. It has provided them with a thorough backdrop for an in-depth discussion of game-playing strategies…brute force versus attempting to minimize the search path.To help students to gain a better appreciation of the work being done, we also made use of many examples of industrial applications captured on videotape from the Texas Instruments satellite broadcasts done last year. The lectures by various prominent figures were used to stimulate discussion of the relative benefits of artificial intelligence applications and the new technologies available.This leads very naturally to the final topic, the Japanese Challenge and 'our response to it, particularly in the development of new hardware. We look at the progress in VLSI chips and parallel processing that is making possible much larger and more sophisticated applications.Obviously it is also important to give the students training in the languages of the field. We have developed a course that concentrates on Lisp and then follow it with a course in Prolog. I would like to have added several other of the artificial intelligence languages including Ops5, and Flavors but decided to wait until we obtained the artificial intelligence workstations in order to keep our software costs down to manageable levels.The Lisp course, based on Golden Common Lisp, available for micros with 512k minimum, focuses on the structure of list processing languages in general and the use of primitive functions to build ever more complex functions to perform all sorts of information processing tasks on lists. It took a considerable amount of time to select the version of Lisp that would provide all the functions that we required for the course. While Golden Common Lisp was not as comprehensive as the version of Franz Lisp we had been running on a VAX mainframe, it did have enough features to support a first course.During the course students have a number of small assignments and are required to write a variety of functions. Here the emphasis is on the mechanics of list processing languages. We usually finish up with two larger assignments. The first is an inheritance problem where students explore the use of properties and a search problem where they are asked to get the computer to determine the best route from one city to any other when several options are available. A variety of search strategies are proposed and then compared…breadth first, depth first, hill climbing. The text we used for this course was Robert Wilensky's Lispcraft.Meanwhile, running concurrently, students have also been taking a course in Learning, Memory and Cognition, where they are studying in some depth human information processing and problem solving behavior. In our case, because we do not have the expertise in the Computer Science Department to handle this course it is being taught by the Psychology Department. The course is important because it provides a strong theoretical base upon which to build. If, in fact, we can develop computers that can think like humans and exercise judgment, then we must have a good working knowledge of how we do in fact perform these tasks ourselves.A constant theme running through these courses is the concept that there is no one problem solving technique but rather that in each type of problem specialised knowledge is needed and in addition we need to have rules of thumb or heuristics to guide us when more than one option is available.In the Prolog course students begin working with developing several small expert systems. We build a diagnostic system using the bicycle repair one as a model. Students are free to choose their own topic. They have ranged from diagnosing physical symptoms of illness to lawnmover repair. Here they are working with a very limited domain. As part of this course we discuss numerous examples of expert systems currently being used. Another assignment that students find quite challenging is the Monkey and Bananas toy problem, taken from Lee Brownston's book “Programming Expert Systems in Ops5: an introduction to rule based programming”. Here the problem domain is limited to a room with a number of objects in it. The monkey is presented with the problem of reaching the bananas no matter where they may be positioned in the room, even if they appear on the ceiling. Several other objects are in the room, including a ladder, a sofa. Students explore in some detail how to handle all the various sets of rules that are needed depending on the room layout and the actual location of the bananas.At the same time students are being given a course in C running in Unix to give them exposure to a portable language that could be called from the Lisp programs and that would help to make their applications run more efficiently. It was clear from the industry experts that we talked to that in many cases the expert systems are rewritten in C after first being developed in either Lisp or Prolog.Having built a reasonably firm foundation, the students are then moved into the Expert systems design course where they undertake more ambitious projects of their own under directed supervision. The first time this course was offered we did not yet have any industry contacts and were limited to what was available on campus. We selected as our project the design of an expert system that would provide the students some first hand experience at knowledge representation. They were to develop a system that would substitute for the reference librarian in the library. The great advantage of this project was that the expert whose knowledge base was to be tapped was readily available to the students.In the first few weeks of the course the students discussed the selection criteria for good projects. Then the class as a whole interviewed one of our reference librarians, Irma Smith who had agreed to participate as our resident expert. The class explained what they were trying to accomplish and decided with her input to use the case study approach. Each group of students then met with her individually to evaluate how she deals with 3 particular reference requests. They were asked to develop sets of rules for the three case studies assigned to them, and develop a prototype system. This took a couple of weeks. In the meantime they were getting lectures on managing costs and risks in project development.The next stage in the process involved meeting as a whole class and combining the best features of each of the different sets of rules. The variety was surprising, and promoted a great deal of heated discussion as the different teams defended their approach to the problem. By the time the students had merged all the rules they had come up with, it was clear that the rules could not simply be added at random. We discussed breaking the process down into clearly defined subsets of rules that would be accessed only when needed for specific applications. Modular design was the best available solution to the problem that presented itself and helped to ensure that the correct sets of rules would be triggered.The need for testing was obvious and each team was asked to test the merged system with the specific cases that they had started with. This necessitated tracing through the firing of the rules and tinkering with the firing order. This proved to be much more time consuming than we had anticipated.At this point we introduce the problem of integrating their little expert system into the existing environment of the library. Here we used the existing and proposed library automation projects and explored how to integrate our expert system as a front end to direct the patron to the appropriate on line databases. Students quickly realized that there was a great deal more involved than had appeared at first. They explored the interface with the on-line card catalog, discovered that they would also have to connect with the circulation system and the interlibrary loan system.The problem of technology transfer was discussed at length...why it occurs, how to recognize problems before they get out of hand, how to ease the introduction of new technology. Several of the library staff felt very threatened by the project and were very critical of the students' efforts.Following this step, students were asked to explore cost estimating techniques for sizeable systems. The next step involved a critical analysis of the user interface. The importance of considering the humans who must use the systems and different techniques for building friendlier user interfaces were explored. Particular emphasis was placed on the use of graphics in the front end design. Students were shown a number of interfaces, both good and bad then they were asked to develop a proposal for their own system. We did not require them to develop any code because of the time constraints.Although originally we intended to get the students to work on an industrial application in the Expert Systems Design course, the library project proved to be very successful particularly because the problems were reasonably familiar to the students and they have ready access to a real expert who can be interviewed at some length. Another great advantage from our point of view is that the project could be repeated each time the course is offered.Because of the need to integrate their systems with databases of various types, we strongly encourage students to also take a course in Data Base Management Systems. We also hope to add in a required course in Graphics or Human Factors in Systems with a strong emphasis on designing better front-ends.By the time students have completed this sequence of six courses in a minimum of two semesters, they are ready for an internship in industry helping to develop real, working expert systems. Our limited experience in finding internships suggests that many manufacturers are interested in diagnostic systems of one sort or another, and financial institutions are also beginning to be interested in possible applications to the financial management area. However it is still difficult to find people who are willing to help the students to have a good experience and there are not many good field work supervisors available. The “toy problem” approach to teaching has worked well for us. We start the students out with examples from Ph.D. dissertations then gradually build them up to designing substantial expert systems of their own, thereby providing them with practical experience with many different approaches and techniques.In closing I should mention that this 'expert systems design' track as we have chosen to call it is still very new and experimental. We are still learning from our mistakes.