Expert assistants - productivity to the power of ...!

Expert Assistants, a type of Expert System, can improve productivity where learning time in a job is longer than a day or two. They can profitably be used to provide a new employee with the knowledge accumulated by their predecessor over a period of months or years. Development time using current shells is cost effective. The objectives of this paper are to give the reader (a) an appreciation of the functionality of an expert assistant, (b) a basic understanding of the potential for productivity enhancement from such a system, and (c) an appreciation of the possible extensions to general training systems.

What is an expert system?

Expert Systems are based largely on symbolic reasoning. This is what differentiates them from traditional computer programs. By symbolic reasoning, we mean the pattern is not hard and fast. The conditions are not based on numbers but on a structured group of explicit symbols; words in our everyday speech. The Expert System has the ability to manipulate these to solve new problems based on previous decisions and their outcomes. As illustrated in Figure 1, Expert Assistants are a subgroup of Expert Systems.

Figure 1. Where the Expert Assistant fits in

Let us look a little more closely at the more familiar type of computer program. It uses sequential programming. The knowledge we have is applied repetitively to our data. Numeric or character data input is manipulated in a manner consistent with what is called algorithmic knowledge. An algorithm is a well defined procedure which will solve a problem in a finite number of steps; we know there will be an outcome.

Data processing systems as they currently exist have used this model since the days of the earliest computers. Our Computer Aided Instruction programs, Computer Managed Learning, Accounting Systems and Database Management systems all use this programming model.

The Expert System is also a computer program, but it uses rules of thumb (also called heuristics) applied to narrow, technical knowledge. Input is usually in words (our symbols above) and inference based on logic theory is applied to the set of rules to process that symbolic input. Information comes from many sources as indicated in Figure 2. A readable description of the basic inference mechanisms of these systems is contained in Bahrami (1988) or Walters and Nielsen (l988).

Expert Systems deal particularly well with areas which are not amenable to algorithmic solutions. Those problems are quite common in industry and government today.

There is a third type of system. We'll call it the Creative and Artistic model; it is non-sequential and adaptive. It is based on common sense knowledge which we all develop over a period of years of acculturation. It is not possible to build an Expert System using this knowledge with either the current computers or with those we can foresee. It is estimated that there are 2(10,000,000) items encoded in our basic common sense. The largest and fastest computers available today would require a lifetime or two to process this much knowledge.

Figure 2. Sources of knowledge for an Expert System (Adapted from Rausch-Hinden, 1988)

Where is Artificial Intelligence (AI) Now?

Nothing has been built as yet that even remotely approaches the functioning of the cerebellum of even the lower amphibians and reptiles.

• Speech recognition and production, particularly related to Natural (or human) Language,
• Locomotion which has been applied to the area of Robotics,
• Machine Vision which has been applied to the area of Robotics as well as signature and fingerprint recognition, and
• Machine Reasoning which has led to the emergence of Expert Systems. Expert Assistants form one sub-category of Expert Systems.

How do we recognise artificial intelligence?

This process is illustrated in Figure 3.

Figure 3. Turing test of artificial intelligence

Expert Systems have been able to meet the Turing test in limited areas, but will consistently fail (as will all sections of artificial intelligence) when only common sense is required to arrive at an answer.

Those portions of common sense knowledge which are important to the area of expertise can be explicitly included in the knowledge base, but we would not include items of common sense which did not directly affect the function of the system. For this reason it is of utmost importance that Expert Systems know what they don't know. Most criticisms of expert assistants, expert systems or other areas of artificial intelligence stem from outcomes which seem irrational to the human user.

One of the most famous is the MYCIN expert system, developed to help diagnose diseases of the blood and central nervous system. When presented with a case of a male Caucasian exhibiting high fever and assorted other symptoms, the MYCIN program asked "Is the patient pregnant?" Needless to say, common sense knowledge about the ability of males to be pregnant has since been added to the knowledge base of the MYCIN system.

Why do Expert Systems show promise?

Figure 4. Expert System Architecture

What is knowledge?

The knowledge is not always locked away in a person's mind. Rather it may be available in written form, but so voluminous that most employees cannot find the answer they require, even if they know it is contained in the reference manuals. This is a situation where an Expert Assistant has been gainfully employed by organisations.

Levels of expert systems: Assistant, colleague and expert

The Colleague level is not as common as the Assistant. This type of Expert System is used to collate knowledge in a field to leverage productivity for an experienced employee. It can also provide help for less experienced employees such as those who work the night or weekend shifts. Colleague level systems often represent improved Assistant systems. The knowledge base is expanded to provide the system with more expertise.

The Expert level is the least common type of expert system. At this level, knowledge has been gathered from multiple sources and the system can be used in place of an expert. Where there are few experts available or the site is remote or harsh, the Expert level system can be justified. It, again, often represents an improved Colleague level system.

Expert Assistants easily solve four types of problems: Productivity bottlenecks; diagnosis of problems; distribution of policy, knowledge or information; and loss of expertise due to employee turnover or retirement.

Examples of problems where an Expert Assistant may help

Productivity

Productivity on night or weekend shifts of a process line is often lower and more uneven than in the day time. Due to seniority, those on the night shift are often less experienced. An Expert Assistant can be used to make expertise more widely available. Mt. Isa Mining has a prototype smelting process control system which was developed by Cameron Russel of JK Tech. (Russel, 1990)

Personal Loans Assistant was developed at a New South Wales Bank. These loans account for a high proportion of the Bank's customer borrowings. A few loans officers were identified as experts in the personal loans area, and one served as the Expert for the system. (Zawa, 1988)

American Express Authorizer's Assistant provides an automated authorisation service to merchants who accept American Express charge cards. (Briefly described in Tello, 1987).

Australian Bureau of Statistics Seasonal Trend Analyser was developed as a joint project by the ABS and Fujitsu. It was a staged development with the prototype discussed at the 1988 conference on Application of Expert Systems in Sydney (Cox, et al., 1988). It is now in use in Canberra.

Diagnosis of problems

Distribution of policy, knowledge or information

Manuals of policy, procedures or regulations are often daunting to many employees. As the information becomes more complex, there is an increased probability of incorrect information being given to clients because the information is so difficult or time consuming to locate. These documents are easily represented as a knowledge base. Having the knowledge captured in a central location improves the ease and speed with which these manuals can be updated.

Combining a Hypertext interface with an Expert Assistant has special advantages here. Hypertext allows the non-linear connection of text and ideas, a very intuitive and powerful interface technology. Microsoft has published the manuals for their new operating system, OS/2 on a CD ROM disk to facilitate its use. The PC Assistant (Cifuentes, 1989) demonstrates a smaller scale implementation.

Some systems have many manuals and the overlap between them is also significant. An update to one manual may require changes to a dozen others. McDonnell-Douglas would have needed 3.6 tonnes of documents for their newest fighter aircraft. They chose to supply 10 CD-ROM disks with electronic searching and extensive cross references as a better and more usable alternative.

Loss of expertise from employee turnover or retirement

An inhospitable location makes turnover endemic. Some companies have a policy of regularly cycling staff through remote locations. An Expert System can be used to standardise productivity and knowledge transfer at the site. The QUT Staff Orientation prototype (Birtwell, 1990) is designed to help with turnover.

Effectiveness of expert systems shells compared to AI languages

In the past two to three years, we have seen many Expert Systems Shells brought onto the market. An Expert System includes the inference engine to interpret the rules and facts and the user interface, but the knowledge about your problem will be missing. We will discuss the desirable characteristics of a shell later.

Using a good shell is analogous to the difference between a modern car and a Model T, or a modern automatic washing machine compared to the old "copper". The shells are not equally capable, but realistic and well defined goals will allow you to reject those unsuitable and find a shell which can be used for your initial foray.

Why can't expert systems be bought off the shelf?

Even within Sydney, corporate culture differences make expert systems difficult to transfer. Just think of two banks. They both make personal loans, but the criteria used may vary widely. If relevant form numbers are referenced, the problem is compounded.

In professional areas, however, cooperative efforts on the part of national associations are bringing assistant and colleague systems on line. At a recent conference, Hadgraft and Wigan (1990) described work being done to make the publication, Australian Rainfall and Runoff, more helpful to professional civil engineers.

Picking an Expert Assistant project

This is particularly true with Expert Assistants. We want to eliminate the bottlenecks in the organisation and improve overall productivity. We can enhance the performance of both the expert and their subordinate by making information and decision rules more widely available.

To determine whether you have a winner, look at the resources available. Are there manuals covering at least a portion of the problem? Is there a willing expert? Is she or he articulate? Will you have management support? Who will be your key team members and what experience do they have?

Set measurable goals. If we haven't set the goals, how do we know when we get there? Measurements can be in terms of decreased demands on the expert's time, increased productivity in the work unit, improved knowledge about the job throughout the work unit, timely availability of induction training for new employees rather than having to wait for the next course, and so on.

Start small. Prove it will work on a part of the problem. This has many benefits. Your team develops experience in the techniques, better focus can be maintained and interface techniques can be tested with the prototype. Additionally, visible results are available sooner and organisational support is enhanced. However, be prepared to throw the prototype away. It is common for the first approach to the problem to be sub- optimal. The techniques learned and problems solved can be put to use in the new version, but don't be afraid to start over. Renovating a poor approach will be more costly and less effective in the long run.

Is it Solvable using expert systems technology?

If the Expert acquired their competence by solving many problems in this area and a few people with specialised knowledge spend time helping others solve the problem, it is amenable to this technology.

Will it be Feasible as an expert system?

Are the organisational resources adequate? You will need to develop the system. You will also require access to the expert in reasonable (say 1 hour) time chunks.

Are the ultimate users willing and committed to using the system you develop? They should be consulted early to determine their perceived problem areas and how such a system could help. As the prototype is developed, involving them in the testing and incorporating their constructive criticism will improve their commitment.

Is the knowledge readily available or will it all have to come from the Expert? Here it is important to recognise that the way the job actually gets done and the criteria on which the decisions are actually made may differ from written procedures or from the training the company provides. If the knowledge must all come from the Expert, the demands on their time may obstruct the project and make it infeasible.

Are the facts relatively certain? Are there discrepancies in how they are applied? Can the problem be subdivided into smaller problems to allow you to start small?

Can the Expert solve the problem in 3 minutes to 3 hours? For the Assistant level, the 3 to 10 minute problem is the best. Where the problem requires longer, an Assistant can be useful if a large proportion of the time is spent in gathering data.

Is someone available who has experience with this technology? This may need to be a person from another section of the organisation or an outside consultant. This person can flatten the learning curve for the internal team. For someone with reasonable skills in computing, there are good Shells on the market which can allow the prototype to be developed in less than a month, even without external resources.

Will an Expert System be beneficial?

Would the lack of an Expert System expose the organisation to significant problems or threats? This is not usually appropriate for a simple Expert Assistant. However, experience gained in the development of the Assistant can lead to better definition of a system to address this problem.

Selecting an appropriate shell

• Does the Shell run on hardware already in the organisation? Previously, Expert Systems were only available on purpose built machines. This is no longer the case. Hardware platforms suitable for Expert Assistants include the IBM PC (R) and PS/2 (R) and compatibles, the Macintosh(R) and the Unix operating system. Some are also available for mainframes.

• Can a user learn the Shell in less than one week? Many of the Shells which have come on the market in the past two years provide on-line tutorials, on-line help and improved manuals to shorten the learning time.

• Does the Shell's user interface have graphics presentation capabilities? Can the interface be customised to your needs? Does the interface support Hypertext linkages? This is an important consideration for the Assistant. The final system needs to provide an intuitive interface. Hypertext linkages are particularly effective for knowledge distribution and embedded training systems.

• Will the Shell design and vendor-provided examples allow development of demonstration systems in less than one month? Novices have been able to make use of an analogous model provided by the vendor as a basis for quickly developing a prototype system.

• What additional new technology is required to gain the functionality of the Expert Assistant?

• Can the Shell connect to databases and spreadsheets and other software currently in use in the organisation? This facility broadens the scope of applications and can dramatically decrease the number of rules needed in the knowledge base.

• How effective are the knowledge base editors and knowledge acquisition tools? Productivity will be enhanced if these two elements have been designed to facilitate the knowledge gathering process.

• Does the Shell integrate multiple, proven technology paradigms such as alternative search strategies, confidence factors, frames, multimedia input?

• Can software written in programming languages already used in the organisation be linked with the Shell? This is particularly useful when extending the Assistant. Where special requirements are identified which are not catered for in the Shell, these may be handled by routines written in the programming languages already used by the organisation.

Conclusion

Demonstrations of Expert Assistants

KnowledgePro (R) is now available for Microsoft (R) Windows(TM) version 3 and supports laserdisc video in addition to the capabilities demonstrated at the Conference. KnowledgePro is a product of Knowledge Garden in the US. For further information, or to arrange to see the demonstrations, please contact the author.

PC Assistant (Cifuentes, 1989)
This product was developed for use by persons new to personal computing. It allows them the opportunity to explore concepts in a non-linear manner and guides them to the desired information. It makes use of the graphic capabilities as well as the hypertext features.

QUT Staff Orientation Assistant (Birtwell 1990)
This prototype was developed for the Personnel section at QUT. It is a prototype, but has demonstrated the feasibility of all the features which are planned for the working Staff Orientation Assistant. Three additional campuses were added to the QUT just as the prototype was completed. The final product will include all campuses.

Trademarks

Bibliography and further reading

1987 report to the Prime Minister on current/future prospects for applying knowledge-based systems in Australia. Suitable contexts, international experience, current domestic use and research, political and control issues.

Although this book is essentially a practical programmer s introduction to the Al language LISP, it also gives a readable account of the basic mechanisms of expert systems - backward chaining, forward chaining, heuristic search, uncertain rules etc.

Report and prototype of a system to be considered for Staff orientation training at QUT.

Report and system for inexperienced users of IBM personal computer systems at QUT.

Preliminary report of development the Seasonal Time Series Analysis system. Useful description of the selection criteria and organisation of the development project.

Discusses the conversion of Australian Rainfall and Runoff, a major resource document in water engineering to a CD-ROM format, the enhancements to its useability for engineers and the potential for combining self education and operational tools in one location.

Another "Handbook of Products" publication; useful classification of software tools under headings of size' capability, usability, cost, etc. Also covers expert systems concepts.

Useful book. Anna Hart's chapter on Rule Induction, "Role of Induction in Knowledge Elicitation" provides insight into this method of developing the required policy rules for a system.

Tongue in cheek treatment of some of the problems with artificial intelligence.

Nisenfeld, A. E. (Ed) (1989). Artificial Intelligence Handbook. New York Instrument Society of America.

Two volumes, Principles and Applications. Engineering and process control bias. The chapter, "Get Ready, Set ..." by James Davis is a useful guide to selecting and managing an Expert System project.

Recommended for its more recent look at Expert System shells and the factors which should be considered when appraising the ones on offer.

Includes many examples from the development of a large industrial expert system for GTE Laboratories. Provides step by-step guidance for the inexperienced as weld as some practical advice. Also useful for those who have previous experience with the technology.

Recommended for its extensive coverage of the range of commercially available expert systems shells (as of 1988) and for the analysis of the factors involved in selecting a software tool for expert systems development.

Russel described the development of the prototype for Mt Isa Mines.

A collection of readings with contributors from business. Includes examples of expert systems in resource scheduling, procurement, retrieval, project support, worksheet enhancement; and a selection of views on expert systems management. The chapter, "A Survey of Issues in Expert Systems for Management," by Robert Blanning is particularly relevant.

One of a series by Tello in 1987. The style is readable and the series covers shells, LISP and Prolog as well as extension to the languages.

Turing wrote this paper on the possibilities of intelligent machines nearly 10 years before computers were in general use in industry, and nearly 40 years before very large scale integrated circuits gave us the ability to put the power of knowledge based systems into general use.

Easy-to-read account of the basic concepts of expert systems, with particular reference to the practical problems of converting human expertise into automated form.

Appendix: Some useful guidelines for selecting a first project

1. Conditions which indicate a problem AMENABLE to an Expert System solution:

   1. Human experts are being lost through retirement or turnover. The competence of relevant experts was developed by solving many problems in this area.

   2. There is a shortage of human experts. Few experts are available and have highly sought after skills or the Experts spend a lot of time helping others with problem solutions.

   3. Experts are needed in many locations. (Branches, plants, etc.)

   4. Experts are working in hostile environments. (eg. oil rigs, desert pipelines, etc.)

   5. Significant operational or strategic benefits would accrue to the organisation such as a competitive edge, better customer service, increased product quality and so forth.

   Remember, the Expert may not necessarily be highly placed but will represent a bottleneck situation.

2. Conditions necessary for selecting an appropriate problem domain:

   1. Practical values are expected. Common sense or Blue sky problems are both inappropriate.

   2. Heuristics are required for good solutions Good may be measured by:

      time it takes to solve the problem
      quality of solution
      cost of solution
      feasibility of solution
      

      Algorithmic solutions can use the normal programming route. Training is usually better unless turnovers make training uneconomic or appropriately qualified persons are scarce. Problems solved by senses are not appropriate for knowledge based solutions.

   3. High return on investment is expected. Expert System will require significant investment. The first system chosen should be a clear financial winner. Expert Assistants are a good first step due to their much lower cost and potential for high visibility.

   4. The Expert can solve the problem alone. If many experts are required, is there general agreement among them or one who has final authority? Too many experts spoil the system when they cannot agree.

3. Conditions necessary for possible problem domains:

   1. Human experts exist in the domain area. If not, where will you get the knowledge for your knowledge base? That Expert must
      • have time to invest in the project;
      • be willing to cooperate in developing the Knowledge base;
      • be able to articulate his knowledge.

      la. The basic knowledge is available: eg. in procedural manuals, text books, documentation of solved problems. This part is particularly important for the Expert Assistant project as much of the knowledge can be collected from these sources, requiring much less of the Expert's time.

   2. Problem solution does not rely on a wide range of common sense. Remember 2(10,000,000) items are estimated to be captured in our common sense.

   3. The solution can be explained verbally. You need to explain them to the users.

   4. The expert can solve problems of this type within a specified time limit. If it is about 3 minutes it is probably better to train your employees. If it is more than 3 hours, it will be a very large project and definitely not just an Assistant. Be careful that the project is not oversold vis a vis its time and cost.

   5. The organisation is willing to commit adequate resources to the project. Management support is the most important. It will also require people, money, time, hardware and software.

   6. The end users are willing and committed to using it. Problems to avoid include:
      • fear of redundancy
      • fear of technology
      • office politics
      • union conflict

   7. The facts to be dealt with are certain. Expert Systems can deal with uncertainty, but the system will be more difficult to build. This complication is best left to at least the Colleague level system. At the Assistant level, deal mainly with identifiable facts.

   8. The problem can be sub-divided into smaller problems. This is very useful as the development can be staged. One small section or a broad overview of capabilities can be developed to improve estimates of the final time and cost of the system. This also allows the eventual users to try out the prototype and become involved in improvements to the design.