Measuring Human Skill: An Expert System Approach

by Anne Breakey, Margarete Floeck, Pete Humphrey, and Jeff Skosnik

[This article was originally published in PC AI magazine, Jul/Aug 1997. The magazine can be reached at PC AI, 3310 West Bell Rd., Suite 119, Phoenix AZ, USA 85023 Tel: (602) 971-1869, FAX: (602) 971-2321, E-Mail: [email protected], Web: http://www.pcai.com]

Introduction

The measurement of human skill is important in many settings. It's particularly important at the British Columbia Institute of Technology, which offers industrial training programs in a variety of areas. In these fields, a poorly trained workforce could result in tragic consequences to both life and property.

Motivated by the belief that an effective use of expert systems technology might lead to the development of superior assessment tools, our research group has developed an expert system-based approach to measuring human skill. Using Amzi! Prolog and Borland Delphi, we have created a dynamic assessment system (DAS), in which an expert system presents numerous questions to a user. The system generates questions by considering responses to earlier questions in the session.

The fundamental thought process behind DAS is that an individual's responses to previous questions are often a good indicator of what to ask next. A paper-based examination, with its fixed set of questions, is not capable of dynamic question generation. We believe that this approach to assessment will result in a deeper probing of what an individual does, or does not, know. Our preliminary anecdotal data is consistent with this idea.

The Problem

Our initial problem concerned an unduly high failure rate in trades training programs in the Canadian apprenticeship system. Every individual within the apprenticeship system has already secured a job in his or her trade, and is alternating time on the job with time at a trade school. Failure at the trade school can cause severe problems for the individual and the employer. Individuals who fail at school often repeat their training, which can negatively affect their wages.

Apart from its heavy cost in human terms, failure that results in repeated training increases the costs of delivering the apprenticeship program (and this program is under heavy financial pressures due to budget cutbacks). Another problem associated with failure in apprenticeship training is that the failure rate is higher for women and minorities than for white males, which raises equity issues.

The causes for failure in training are, of course, complex and multiple. Improved training efforts can't overcome all of them. But improved efforts can overcome a significant proportion. In particular, we've had success by providing upgrade training in basic math and science prior to entry into first year apprenticeship training. The failure rate for the several thousand electrical apprentices in British Columbia is around 15%. It would not be appropriate to require every first-year apprentice to take upgrade training in math and science for the sake of the 15% who will probably fail without such training. At present, we use a simple paper-based exam in math and science to try to identify individuals who are at high risk of failure. This exam, however, is widely acknowledged not to do a particularly good job of identifying individuals who may benefit from the upgrade training.

Dynamic Assessment

Recognizing that the computer is capable of dynamic behavior, we thought that the best approach to computer-based examination would be an oral exam. In this type of exam, an individual's prior responses can determine the examiner's future questions.

For example, a skilled teacher might ask a student: "What is negative 2 times negative 3?" If the student responds "Negative 6," the teacher would note that the student got the quantity right but the sign wrong. The teacher might then ask another similar question, and infer from the response that this student is weak in the area of signed numbers. Depending upon the purposes of the oral exam, the teacher might then drop all the other questions he or she might have asked except those relating to signed numbers. In this situation, the teacher is using expert knowledge to guide the questions in order to identify conceptual confusions and ignorance in broad areas.

An Expert Systems Approach

Every experienced teacher of elementary math is aware of many common errors. Quite often, the errors form a pattern, as a confused student does not incorrectly apply a correct principle of mathematics but rather correctly applies an incorrect principle.

For example, to solve for x in the equation 3x = 9, a student might apply the false principle: subtract the numeral value (in this case 3) from both sides of the equation, and then rewrite the equation as x = 6. In fact, the correct principle is: divide both sides of the equation by the numeral value (3), and then (in this case) rewrite the equation as x = 3.

We have found about a hundred principles for which our trades students (in common with many others, of course) have created false alternatives, which they apply consistently in their work. In our assessment system, we have the computer generate questions in which some of the multiple-choice answers reflect these common misconceptions. When a student selects an answer that suggests he or she may be applying a false principle, the computer dynamically generates a similar question. If the student repeats the error, the computer displays the principle that he or she should have used to solve the problem. Later in the assessment, the computer returns to all the questions for which the student received clues about the proper problem-solving techniques for that class of problem.

If the student then correctly answers the question, it's reasonable to conclude that the student was unable to solve a certain type of mathematical problem due to a misconception about how to solve such problems. With minor coaching from the computer, the student was able to solve this type of problem -- i.e., the student can learn (some) math, and, what's more, can learn it from a computer.

In this situation we infer that, although this student has some problems, he or she nevertheless is a good candidate for success in the apprenticeship training program, particularly if first given some remedial training. From this example, it's clear that innovative use of technology can lead to superior assessment in at least two ways:

Thus, the approach of dynamic assessment can not only obtain a better picture of what an individual actually knows, it can (to some degree) also determine the individual's learning potential in the problem domain relevant to his or her training needs.

Limitations of the Approach

The approach is very well-suited to adult learners about to enter training programs for which elementary mathematics is a prerequisite. As students move up to higher levels in their training programs, the possible causes of failure multiply and become more difficult to isolate.

In the case of elementary mathematics, many adults (for example) don't know that the rule for dividing fractions is to invert and multiply. When they were taught this principle, their language skills and general comprehension were probably not at an adult level. Many children simply lack the intellectual maturity even to understand the language in which this basic principle is expressed. Our assessment software will quickly zero in on this sort of problem and notify the individual of the principle. In most cases, the adult learner easily grasps the principle when it is succinctly stated and presented as a clue for the next problem of this type.

Verification

In cooperation with �we have just successfully completed a small pilot project with unemployed youth in which we used the DAS software to determine whether they would meet trades training entrance standards. In cooperation with the British Columbia Ministry of Labour, we will be using the dynamic assessment software on approximately 200 trades students about to enter first-year apprenticeship training. These 200 students are randomly drawn from 20 different trades in the electrical and mechanical occupations. We believe the dynamic assessment software will result in a 5% reduction of the failure rate, as compared with students evaluated via the paper examinations presently in use. Results on the success or failure of this pilot effort will be available in the first half of 1998.

Future Extensions

We've demonstrated the DAS software to local industry and government in British Columbia. We've also shown it to the senior management of Confederation College in Thunder Bay, Ontario, which plans to use the software to support its trades training program. Given the strong interest already shown in the software, its continued use and further development appear certain. Two areas for further technical development are actively underway.

The first area relates to the difficulty of applying the software to non-elementary skills, where the possible causes of failure decompose into a multitude of confusing possibilities. The premise from which we begin our investigation of this problem is that failure at a higher level is often due to the individual having marginal skill at a lower level.

Many individuals for whom algebra appears to present insurmountable obstacles in fact have problems with arithmetic. Similarly, many individuals who fail college-level calculus repeat the subject only to fail again, because their real problem is that they do not have a proper grasp of algebra. This suggests a relationship among skills akin to the relationships expressed in a family tree.

If this is the case, then (at least in the area of mathematics) the very complex relationships among different skills, at different levels, might be deducible in much the same way the family relationships are deducible in classic family tree programs. Because DAS is a Prolog-based system, the analogy between a family tree and the tree of knowledge is an interesting one to pursue.

At present, assessments are composed in an ad hoc manner: each link from one question to the next is hand-coded; DAS does not provide any sort of conceptual framework to classify questions and organize them with respect to one another. We will recast the system so that, for a given domain, a "tree of knowledge" (similar to a family tree) is developed, and failure at one level in the tree will automatically result in the generation of a "nearest ancestor" question.

The second area for further research relates to the use of synthetic speech. Our thought is to use the definite clause grammar, which Prolog supports, to generate verbal clues and help via one of the popular text-to-speech commercial development kits. This line of research stems from our interactions with the technical school in the Arab gulf, which has expressed a strong desire to develop tools for assessing English language skills, both written and oral. The native tongue of most of the students at the College is Arabic, and the language of instruction is English. Given Prolog's roots in natural language processing, this is a very exciting area for further applied research.

Application to Distance Education

Another area of application we intend to explore is the development of a new generation of distance education software. The instruction is "assessment driven" -- individuals will receive an initial assessment and then, if necessary, focused instruction will be provided online, followed by a re-assessment (and repeated training if necessary).

Though a very natural extension to our present work, we will only undertake this effort after recasting DAS into the "tree of knowledge" model we described. Working in partnership with the Open Learning Agency in British Columbia, we will use the next generation of DAS software to develop an "assessment-driven" distance education training package. This package will train students in the field of programmable logic controllers and will teach elementary ladder logic programming.

Project Tools

Everyone on the DAS development team has a strong background in Prolog programming. Nevertheless, we felt that to program in a purely Prolog environment would be inefficient, particularly since we wanted to include multimedia and a sophisticated user interface.

We decided to build the user interface in Borland's Delphi, which provides a visual development environment as well as an excellent database and user interface design tools. This meant that the Prolog portion of the overall program could handle decision-making and intelligent tasks (tasks for which Prolog is ideally suited). We selected Amzi! Prolog to use with Delphi because:

A Software Engineering Perspective

From a software engineering perspective, the Dynamic Assessment System is especially noteworthy because of its unique integration of Amzi! Prolog and Borland's Delphi development environment. Although it's now a full-scale application, it started out as a means for exploring the possibilities for combining these two disparate programming environments. As a result, it uses some innovative techniques to make use of the best features of these two environments.

In the system, Prolog predicates can request GUI services from Delphi by means of "external predicates," a mechanism in Amzi! Prolog which lets Prolog access non-Prolog functions as though they were built-in predicates. Conversely, Delphi functions, written in Pascal, can invoke Prolog predicates, unify arguments, and backtrack to obtain additional solutions via call and redo functions.

Because the Dynamic Assessment System is an MS Windows program and therefore event-driven, it presents some additional challenges. Delphi is designed primarily for the Windows environment, and the main event loop is its responsibility. It deals with the task of handling the Windows system and user input events. When the initial application form is created, the Delphi component initializes the Amzi! Prolog engine and registers the external predicates implemented in Pascal on the Delphi side. From that point on, all activity is scheduled in response to asynchronous events.

In order to process a question, a Delphi function calls a Prolog predicate to present the question to the user. The Prolog code determines the components to display and how to display them, based on the definition of the current question and the user's performance on previous questions. The Prolog predicates, in turn, call external predicates supplied on the Delphi side to perform the actual GUI operations.

When the user performs an action, such as requesting help or a hint, an appropriate Prolog predicate is called to determine the proper course of action. When the user selects an answer, the selection registers in a database and moves on to the Prolog engine, which determines the question to present next. Delphi uses its built-in relational database capabilities to track each student's performance and generates appropriate reports for administrators.

One area in which the two systems are especially tightly coupled is in the equation-drawing facility. A generic text and symbol composition facility is included in the system that allows text and graphics elements to combine to produce equations and figures. These drawing algorithms were most easily expressible in Prolog. Because of the requirement that applications be able to redraw any portion of their window output, however, we designed these routines to compile equation definitions into a series of drawing primitives. Whenever the Delphi event-handling functions are notified to redraw the contents of the equation/figure windows, a Prolog predicate is called which runs down the list of drawing primitives to re-render the output equations and figures.

By using the cross-language integration features of Amzi! Prolog, it was possible to develop the Dynamic Assessment System using Borland's Delphi to handle the composition and control of the user interface, and Amzi! Prolog to handle the question-management and presentation tasks. This allowed us to take advantage of the strengths of each tool and express different aspects of the system in the language best suited to the task.

Figure 1: DAS has excellent plotting capabilities. Function plots are generated dynamically rather than stored. This saves disk space and allows more flexibility.

An AI Perspective

When a DAS session runs, it plays a Prolog script file. A Prolog script file consists of a set of question matrices. A particular question matrix consists of a set of frames with slots to hold initial and transient values. The system can generate slot values dynamically and update them dynamically in response to user behavior. The final frame in a given question is the reaction frame, in which the author of the particular assessment can stipulate a list of actions to be performed, given the response to the question at hand.

This simple mechanism is extremely powerful, particularly given DAS's ability to generate equations and graphs dynamically. New question frames can be asserted into Prolog's database and then asked via next(Q) in the reaction frame of the current question. The default behavior is simply to ask the next question in the script file, unless instructed to do otherwise in the reaction frame. The expert rules accompanying particular script files are captured in the reaction frames. The actions slotted into the reaction frame can include:

Script files can be composed dynamically from a Prolog database of question matrices: Expert principles generate the appropriate values for the various question matrices. Alternatively, one can write a complete script file in a standard program editor and the system can consult it at runtime. We compose simple script files in an MS word macro written for this purpose. If the reaction frame is empty in all questions, the system defaults to ask each question in the script in their listed order. The default behavior delivers a standard exam with a fixed set of questions asked in a fixed order.

Listing 1 presents an excerpt from a Prolog script file, with some comments in italics preceded by %.

Figure 2: The Question Design Form used to generate DAS script files

question(alg2, [                        %�alg2� = question identifier; referenced in "next(X)" reaction frames 
  category:'Algebra',                   % �Algebra� = question topic; used to identify help topics
  prompt:paint(                         % �paint� creates a graphics windows
    [$Solve for x:$,                    % this slot holds specific content of the question
     $ $,                       
     eq(y=m*x+b) ]),                    % eq(y= m*x + b) draws an equation into the graphics windows
  response:choices(                     % defines a multiple choice question; alternative slot value = true/false
    [a: paint([eq(x=(y-b)/m)]),         % choices given here
     b: paint([eq(x=y/m-b)]),
     c: paint([eq(x=m*y+b)]),
     d: paint([eq(x=sqrt(m*(y/b)))]) ]),
  answer:b,                             % indicates the correct answer
     hint: text([$Divide both sides by m$]),  % this slot contains what is displayed when user clicks on hint
     help: true,                        %  true makes the help button visible, giving access to help; alternative slot value = false
     reaction: [                        %  reaction frame
        ans(a): [qMake1, next(aa1)],    %  list of prolog clauses, execution triggered by user answer
        ans(b): [qMake2, next(ab1)],    %  selection.   These clauses are used to capture expert rules
        ans(c): [qMake3, next(ac1)],    %  defining system behavior in response to student 
        ans(c): [qMake4, next(ad1)],    % performance.
        ans(_): [help(linear), next(alg2)] ]   ]).   %  �ans(_)� = "Don�t know"; reaction:  open help file.

Listing 1. An excerpt from a Prolog script file

Figure 3: The Screen resulting from the Prolog Script excerpt.

Biographies

Anne Breakey was Manager of Apprenticeship Standards for the province of British Columbia, Canada; she is now Director of Special Projects for Pacific AI.

Margarete Floeck is the President of the Pacific Artificial Intelligence Systems Corporation (Albion, B.C., Canada). You can reach her via email at [email protected] or at (604) 467-4625.

Pete Humphrey is with Stagecoach Consulting Services (Sandia Park, NM). You can reach him via email at [email protected] or at (505) 286-1323.

Jeff Skosnik is on leave from his position as the Program Head of Industry Services at the British Columbia Institute of Technology (Burnaby, B.C., Canada). You can reach him care of Pacific AI.

Acknowledgments

We gratefully acknowledge the contributions of Mary Kroening, Dennis Merritt, Joe Nodeland, and James Webb. James Webb is a student at the British Columbia Institute of Technology who wrote an MS Word Macro that generates Prolog assessment script files. This makes DAS available to non-prolog programmers and others who would prefer not to struggle with the syntax in the question frames of the Prolog script files for assessments.

Note

Visit http://www.amzi.com for further information on the Delphi/Prolog programming environment.