HFS89: An Expert "Critiquer" for Propulsion Gear Design
Appeared in the the Proceedings of the 1989 Conference of the Human Factors Society, October, 1989.
An Expert "Critiquer" For Propulsion Gear Design:
A Case Study In Intelligent Decision Support
Ellen K. McKinley
Westinghouse Science & Technology Center
Pittsburgh, PA
Michael L. Mauldin
Carnegie Mellon University
Pittsburgh, PA
Emilie M. Roth
Westinghouse Science & Technology Center
Pittsburgh, PA
ABSTRACT
This paper describes an "intelligent" designer's aid that was developed to
support the design of marine propulsion gears. Key elements of the system
include: conversion of gear design formulas from a procedural to a
declarative form to facilitate inspection: a direct-manipulation interface;
and encoding of "expert" design constraint knowledge. The case study
demonstrates that delivering "expert knowledge" is often only a small element
of an "intelligent" support system, and provides a concrete illustration of
the importance of a cognitive task analysis in defining the elements of an
effective support system. This design solution should have applicability to
other engineering design tasks.
INTRODUCTION
Advances in computer science and artificial intelligence (AI) are providing
powerful new tools that expand the potential to support complex cognitive
tasks: diagnosis, troubleshooting, planning, design. While these tools
provide new opportunity for support, they also create new challenges: how to
deploy the machine power afforded by these new technologies to support human
activity (Roth, Bennett & Woods, 1987; Woods & Roth 1988). In this paper we
describe a case study in the design of intelligent decision support systems:
an intelligent system to support the design of marine propulsion gears. The
case study demonstrates how encoding and delivering "expert knowledge" is
often only a small element of a total support system, and provides a concrete
illustration of the importance of a cognitive task analysis in defining the
elements of an effective intelligent decision support systems.
Propulsion gear design encompasses many features of engineering design
problems in general. While there is a large experiential base on how to
design gears, it remains a creative task with no closed form solution. In
this respect, it differs greatly from the types of procedural tasks (e.g.,
maintenance diagnosis) that have received the most attention in the applied
expert system literature (e.g., Richardson, 1985). The gear design process
involves consideration and balancing of multiple constraints based on multiple
points of view on the problem (e.g., stress, noise, manufacturability,
maintenance). The relevant knowledge of constraints resides in different
individuals or groups. For example, gear designers do not have ready
information on manufacturing capabilities and constraints that impact on the
economic viability of alternative gear designs. As a result, the design
process often involves multiple iterations across groups.
A need was felt to build a designer's aid that could encode and deliver expert
knowledge of the multiple design constraints and design strategies for meeting
them in order to reduce the need for across group iterations. The need was
perceived to be all the more pressing because several individuals with expert
knowledge of the manufacturing process and its impact on gear design were near
retirement age. The perceived goal was to build an "expert system" that would
capture their unique manufacturing knowledge for delivery to the primary gear
designers. This challenge was how to utilize advanced computational
techniques to support the creative design process rather than build la system
that would automate the design task (Woods & Roth, 1988).
COGNITIVE TASK ANALYSIS
A thorough cognitive task analysis of the cognitive activities involved in the
gear design process was conducted (Woods & Roth, 1988). This included
analysis of the kinds of manufacturing knowledge that impacts gear design,
analysis of kinds of contributions made by the experts in gear manufacture to
gear design, and analysis of the sources problems and bottlenecks in the gear
design process in general. The analysis revealed that experts ingear
manufacture contributed significantly to gear design. There were three
sources of contribution, two of which could be captured in a design aid.
First, gear manufacturers possessed knowledge of manufacturing constraints
(e.g., capabilities and limits of available tooling machinery) that impacted
on the feasibility and economic viability of gear designs. Second these
experts had developed computer programs (written in BASIC and FORTRAN) that
defined the exact parameters of the gears for manufacturing purposes.
Translation of this software into a non-opaque format for use by the primary
gear designers turned out to be a key element for effective support of the
gear designers. Third, gear manufacturing experts contributed to creative
design solutions for unique design problems. This third, creative
contribution, is what marks true human expertise and falls beyond the bounds
of what can be captured by state-of-the-art Al systems.
Analysis of the gear design process also revealed that one of the major
sources of problem was the opacity of the existing codes for computing gear
parameters. Because gear design is not a closed form problem, there are often
several alternative formulas which may be used to compute particular gear
parameters (different ones may apply under different circumstances, and/or may
be advocated by different engineering organizations). The opacity of the
software made it difficult to identify what formula was being used in any
particular case. This made it difficult to trust any given program, reconcile
inconsistency in answers among programs, and incorporate new formulas into the
programs. In addition, the design software they were using was run in batch
mode, thus delaying feedback, and inhibiting rapid iteration in design.
A generic software system was developed to meet these requirements. Key
elements in the design solution were:
- A generic engine for manipulating and computing algebraic formulas was
developed. Formulas are encoded in a declarative rather than procedural form.
In setting up and modifying the program, engineers type in what the formulas
are and the system determines how to compute them and in what order. The
system also provides the capability to encode multiple redundant formulas for
interdependent variables (e.g., the ability to define variable A as a function
of variable B and B as a function of A). This is a feature not normally
available in conventional programs and allows the engineers flexibility in
choosing which variables to provide as input and which to have the system
compute in any given instance.
- A direct manipulation interface (Schneiderman, 1982) was developed for
rapid design iteration. See Figure 1. A spreadsheet like format is employed
that allows the engineer to enter and change data while getting immediate
feedback on the effect of each change on the overall design. Solid lines
indicate fields where the engineer may input a value. Dotted lines indicate
fields of variables that are computed. As the engineer moves from one cell to
the next, the formula encoded for the current cell along with the variables'
names that depend upon that formula are displayed at the bottom of the screen
in the message area, making the formulas used in the computation and
interdependencies among variables transparent.
- A capability to encode expert rules was provided that allows
constraints on gear designs to be encoded (e.g., manufacturing constraints
knowledge). The systems alerts the engineer if constraints are violated
and recommends corrective actions. The out of bounds parameter appears in
reverse video. The constraints being violated and any recommendations
appear at the bottom of the screen. See Figure 2.
- The system also includes a graphic display the gear tooth profile.
This gives the engineer the ability to see his gear design dynamically
change as he modifies gear parameters. See Figure 3.
This system has been delivered to the gear design engineers and has been well
received. The intelligent design environment allows the engineer to rapidly
try out alternative designs, see (and be able to modify) what formulas are
employed, and get immediate feedback on the implications of a design change
(both in terms of a graphic display and a check of whether any design
constraints are violated). The system reduces the time spent in the mundane
part of the design process and allows the engineer more time for developing
creative design solutions.
Figure 1: Basic design screen partially completed
Figure 2: Basic design screen showing constraint violation
Figure 3: Basic design screen and graphical display
CONCLUSIONS
Several conclusions with applicability beyond gear design may be drawn from
this study. First, the study provides a concrete illustration that
intelligent decision support need not be synonymous with "delivering expertise
in a box". In this case, encoding of expert design rules was only a small
element of the total intelligent decision support system. The primary key to
performance enhancement was the conversion of engineering formulas from a
procedural format to a declarative format that allowed them to be more readily
inspected and manipulated. Second, the direct manipulation system that was
built demonstrates unique application of advanced computational techniques
with application beyond the initial gear design problem.
References
Richardson, J. (Ed.) ``Artificial Intelligence in Maintenance.''
Proceedings of the Joint Services Workshop, New Jersey, Noyes
Publications, 1985.
Roth, E.M., Bennett, K., Woods, D. D., ``Human Interaction with an
`Intelligent Machine,''' International Journal of Man Machine
Studies, 1987, Vol. 27 , pp. 479-525.
Roth, E.M. Woods, D. D., ``Aiding Human Performance: I. Cognitive
Analysis,'' Le Travail Humain, 1988, Vol. 51, pp. 39-64.
Schneiderman, B. ``The Future of Interactive Systems and the Emergence of
Direct Manipulation,'' Behavior and Information Technology, 1,
1982, 237-256.
Woods, D.D., Roth E.M., ``Cognitive Engineering: Human Problem Solving with Tools,'' Human Factors, Vol. 30, No. 4, 1988, pp. 415-430.
Last updated 14-Sep-94