Machine Learning in Engineering Automation - The Present and The Future (Computers in Industry, Vol. 17, Issue 2-3) (1991)

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

Computers in Industry 17 (1991) 91-100 91

Elsevier

IMS '91--Learning in IMS

Machine learning in engineering automation


The present and the future
Bradley L. Whitehall a n d Stephen C-Y. Lu
Knowledge-Based Engineering Systems Research Laboratory, Department of Mechanical and Industrial Engineering, and Beckman
Institute for Advanced Science and Technolog3', Universitv of Illinois at Urbana-Champaign, Urbana. IL 61801, USA

This paper describes how engineers can use machine learning ships and trends in the plethora of data available
techniques to improve their productivity. The current use of to the engineer. In the future, machine learning
machine learning in engineering automation is described, and
programs will interact with existing computer tools
areas of future research are discussed. Machine learning pro-
grams are able to identify useful concepts from data. Recogniz- to enhance their productivity and also act as an
ing relationships from the plethora of data enhances the associate to the engineer by suggesting ideas
knowledge an engineer can use for problem solving. In the learned from past problem solving sessions. Ma-
future, machine learning programs will interact with existing chine learning cannot only increase the power of
computer tools to enhance productivity and to assist the en-
gineer by suggesting ideas remembered from past problem
software programs for engineering, but can also be
solving sessions. Machine learning can increase the power of used to help the engineer learn about problems
programs for engineering and can help the engineer learn and domains. For example, constructing an expert
about problems and domains. system often helps an organization to better un-
derstand the problem the system is solving. This
Keywords: Machine learning, Machine learning requirements,
new understanding goes beyond what is captured
Engineering automation, Intelligent engineering. in the knowledge base, because the organization
was forced to look at the problem in a method-
ological manner. In conjunction with expert sys-
1. Introduction tems, machine learning can be used to help an
organization understand aspects of problems in
In the past decade the computer has become an greater detail than might have occurred otherwise.
indispensable tool to the engineer. This trend will With machine learning capabilities, a program can
continue with more innovative software applica- identify concepts and important features of a
tions being developed to use the ever increasing problem without the biases humans impose upon
power of new processors. To continue the large solutions.
gains in engineering productivity (even with in- This paper consists of three major sections. The
creased computational speed) software programs next section describes machine learning tech-
will need to be more intelligent. One method of niques from a user's viewpoint, i.e., what is re-
making programs that can increase productivity in quired to use these systems and what is the output
engineering automation is to have them learn while obtained. Section 3 describes how machine learn-
they perform a task. ing can be used in engineering automation today.
In this paper, machine learning in engineering Section 4 explains where machine learning re-
automation is discussed. Current machine learning search is heading and what it implies for intelli-
programs learn concepts that can help to increase gent engineering software in the future. This is
engineering productivity by recognizing relation- followed by closing remarks.

0166-3615/91/$03.50 ~3 1991 - Elsevier Science Publishers B.V. All rights reserved


92 IMS ' 9 1 - Learning in IMS '( ,.,mpuWr~ :n induste~

2. Machine learning: An engineering perspective ML

Machine learning algorithms are often cate-


gorized by the type of knowledge provided to the Data DMI_ IML lnf~l_
system and the transformations performed by the !
learner in order to apply the knowledge. Figure 1 • /// '

is such a hierarchical classification of learning ,., ~. ~ ' ,,,


Rote I ~ !
systems. In the rest of this section, four different
types of learning systems are presented: database,
Fig 1, Hierarchical classificalfim ,,i machine Icaruing
deductive machine learning (DML), inductive
machine learning (IML), and integrated machine
2.1. Database learning
learning (Integ-ML). Database learning methods
are used extensively by engineers today. Deductive
Systems that learn hy met];orizing prevtousl?
machine learning and inductive machine learning
seen results and facts perform database learning
systems have been researched for a number of
Database learning systems make no inductive gen-
years and can now be used to help with specific
eralizations from the input data provided. They
engineering problems. Research with integrated
remember the facts that are entered and can pro-
machine learning systems is just beginning; these
vide summaries from subsets c,f the facts based
learning approaches will be useful in the future.
upon provided instructions.
Two types of database learning are role learn-
StephenC-Y. Lu received his MS and
mg and learning by instruction [1]. Rote learning
PhD degrees from Mechanical En- stores the initially given facts and recalls them
gineering and Robotics Institute of
Carnegie-Mellon University, Pins-
when the correct key is provided. Learning b~'
burgh, PA. He is currently a tenured instruction occurs when a system remembers given
Associate Professor and the founding facts, but recalls information m a form different
Director of the Knowledge-Based En-
gineering Systems Research Labora- than initially provided. An example for an en-
tory in the Mechanical and Industrial
Engineering Department at the Uni-
versity of Illinois at U r b a n a -
Champaign (UIUC). He is also a Re- Bradley L. Whitehall received his MS
search Associate Professor of the and P h D from the Department of
Computer Science Department and the Beckman Institute for Computer Science at the University ol
Advanced Science and Technology at UIUC. His research Illinois at U r b a n a - C h a m p a i g n in the
interests are in the development of artificial intelligence based area of artificial intelligence. Dr
techniques for advanced engineering automation, and in the Whitehall is currentl,/ with United
integration of these techniques with traditional engineering Technologies Research Center, East
methods. Currently, he is developing the knowledge processing Hartford. CT, The work described m
technology to support various concurrent engineering and sys- this paper was completed while he was
tem management tasks• He has published over 100 technical a Postdoctoral Research Associate
papers, reports, and book chapters in this area, and served as a with the Knowledge-Based Engineer-
keynote speaker for several national and international con- ing Systems Research Laboratory in
ferences. Professor Lu is an associate editor of the Journal of the Mechanical and Industrial En-
Engineering for Industry, ASME Transactions, and the Interna- gineering Department at the University of Illinois at Urbana
tional Journal of Computer-Integrated Manufacturing Systems. Champaign. His research interests include applying machine
He also serves on several editorial committees including SME learning techniques to engineering problems and using artifi-
Transactions, Journal of Intelligent Manufacturing, and the In- cial intelligence to improve design processes. He has publica-
ternational Journal of System Automation. He has organized tions at conferences in artificial intelligence both in the United
many technical conferences and workshops, and was the States and Europe, He managed a major section of the Robols
Chairperson of the 1986 A S M E Symposium on Knowledge- and Beyond." The Age of Intelligent .Machines development
Based Expert Systems for Manufacturing and the 1989 A S M E effort at the University of Illinois. (A request from the Boston
Symposium on Concurrent Product and Process Design. He is M u s e u m of Science for major universities to demonstrate
an active member of ASME, SME, IEEE, and AAAI, and has artificial intelligent systems for the general public.) Recently,
served as a technical consultant and expert panelist to various he was the co-developer of the first cognitive science course at
industries and federal agencies. He received the Presidential the University of Illinois. The aim of the course is to give the
Young Investigator (PYI) Award from the National Science student an overview of all the areas of cognitive science (artifi-
Foundation in 1987, the Outstanding Young Manufacturing cial intelligence, psychology, neural biology, anthropology, and
Engineer Award from the Society of Manufacturing Engineers linguistics). Dr. Whitehall has conducted industrial seminars
in 1988, and was selected as an Outstanding Young M a n of on artificial intelligence and machine learning and has been
America in 1988. In 1990, he received the Xerox Senior Faculty involved with companies applying machine learning techniques
Research Award from U I U C for his high quality research to production problems. He a member of the organizing com-
accomplishments over the previous five years. In the same mittee for the session Machine Learning in Engineering Auto-
year, he was also appointed as a University Scholar for his mation at the Eighth International Workshop on Machine
excellent contributions to scholarly activities. Learning.
Computers in Industry B.L. Whitehall, S. C- Y. Lu / Machine learning in engineering automation 93

gineering system is providing a summary of statis- example requires a plan for machining slots, the
tics for a subset of the input data. new rule will improve the speed with which plans
Most engineers use programs with database for slotted parts are produced.
learning capabilities frequently. They do not con- The EBL system ARMS [4] learns plans for a
sider a spread sheet program to be learning. Such robotic arm. The system is provided a plan for
systems are so common place that their usefulness assembling a part. The system uses its background
is taken for granted. In the future, engineers will knowledge to understand the plan and generalize
use these learning systems with other computer it. The generalized plan can then be used by the
tools and more complex learning systems to in- robotic arm to assemble similar parts. The gener-
crease productivity. alized plan helps the robotic arm work in situa-
tions where the provided parts are in different
2.2. Deductive machine learning positions from the training example, making the
assembly process more robust.
Deductive learning [2,3] works with an existing
knowledge base to make it more efficient (faster). 2.3. Inductive machine learning
Deductive learning can improve the performance
of a knowledge-based system by adding rules that Inductive machine learning systems produce a
combine knowledge already existing in the knowl- concept description for a class of data by identify-
edge base. The new rules compile (cache) knowl- ing characteristics that distinguish one class from
edge that is used for many types of problems. The other classes. For a system to do this, a large
new rules allow the inference engine to work ex- number of examples must be compared. In gen-
pediently because only one rule is required where eral, the more examples provided, the more accu-
many rules (along with pattern matching and rate the concept description will be. IML systems
chaining involved in using those rules) were re- generalize from the specific examples provided
quired to solve the problem. when constructing the concept description. The
To learn a new rule, deductive learning systems generalized concept description is useful for iden-
must be provided with a training example, The tifying unseen examples because rules can be con-
learning system assumes that the knowledge base
structed of the form:
is complete and correct. The system constructs a
proof of the solution for the training solution,
recording how variables are instantiated and which concept description ~ data class.
rules are used. Deductive learners generalize the
proof structure to create a new rule which can The generalized rule can provide insight to an
then be used by the inference engine as though it engineer by highlighting aspects of the data that
was an originally provide piece of knowledge. would go unnoticed without this analysis tool.
Next time a problem similar to the training prob- Two types of inductive learning are discussed
lem is encountered, the system can use the learned next. Learning from example systems construct
rule instead of constructing the solution from ini- concept descriptions for preclassified examples.
tial principles. Deductive learning is also called Discovery systems are provided with a set of un-
explanation-based learning (EBL) because the classified data and must classify the examples and
proof or explanation of the solution is constructed learn the concept descriptions for those classes.
from the training example. There are symbolic and subsymbolic (con-
Explanation-based learning is useful when a nectionist or neural net) approaches for both types
class of problems use many of the same rules. A of inductive learning. Both approaches are useful
new rule can be learned to speed up the knowl- for different types of problems. The focus of this
edge-based system's performance on this problem paper is on symbolic approaches, because the rules
class. For example, generative computer-aided produced can provide insight to the user in a more
process planning systems use knowledge to pro- comprehensible representation. Subsymbolic ap-
duce plans. If a number of parts presented to the proaches can learn functionally equivalent concept
system use the same rules in constructing the plan, descriptions, but weights on network links do not
a new rule can be constructed. If the training help in understanding a solution to a problem.
94 1MS '91~-Learning in I M S ( ~mputer.~ m lndu.str'.,

2.3.1. Learning from examples tries the learning techniques on all the subprob-
Learning from example (LFE) systems induce lems. Thus, different parts of the problem can be
general concepts from a set of classified examples solved by different learning tools to increase the
[5 7]. The simplest case is when two classes are overall performance the rules learned [13]. (This
provided, positive and negative. LFE systems learn approach is described in more detail in Section
a concept of the positive class by comparing at- 3.1.)
tribute values for the examples and finding combi- The second approach to integrating machine
nations of the values that describe examples learning systems is to combine the capabilities of
labeled as positive but do not label (or label only a DML and IML systems by using knowledge from
few of) the negative examples. If only a few exam- both examples and background knowledge. These
ples are provided these systems find descriptions types of systems will become more important m
that do not fully described the desired concept. the future because they allow an engineer to pro-
The use of learning from example systems for vide more guidance to the learning ~ystem, even
systhesis tasks are described in Section 3.1. when complete knowledge is lacking, Two forms
of this integrated learning system approach are
2.3.2. Discovery presented: knowledge-based learning (KBL) [14]
Machine learning systems can provide useful and analogy [15].
information even when given unlabeled examples.
discovering important classes of data as well. 2.4.1. Knowledge-based learning
Numerical clustering techniques are used by en- KBL systems allow induction from examples to
gineers and scientists to provide insight to data. be guided by background knowledge. KBL sys-
Discovery systems can provide the same type of tems work with the inference engine of an experl
insight, but can group examples based upon fea- system. When an inference engine cannot solve a
tures other than numerical distance. Machine problem, KBL uses heuristics to determine possi-
learning systems can classify examples based upon ble extensions to the knowledge base that would
the conceptual cohesiveness on the examples [8-10]. allow a solution to be found. For each selected
One such measure is finding simple explanations extension a partial proof structure is constructed
for the groupings of data. (Section 3.1 describes a (similar to EBL). The proof structure is used to
situation where discovery was used in an engineer- collect a subset of the provided examples and
ing domain.) induction is performed over the subset of exam-
Another type of discovery system finds ples to create a rule. The new rule is then used to
mathematical formulas that describe a set of data allow the inference engine to complete the prob-
[11,12]. These systems discover equations covering lem.
sets of the data points provided. Events are classi- The KBL1 system [14] has learned rules for an
fied by the equations, which are the concept de- expert system that determines phase timings for
scriptions. Equations are particularly useful de- intersection stop lights. KBLI was provided with
scriptions because they can be used with other an incomplete set of rules for configuring phase
numerical systems. timings, a set of previously solved examples, and
an intersection for which a timing could not be
2,4. Integrated machine learning found. By using the provided rules to construct a
context for the unsolved problem, Kin,1 selected
There are two ways that machine learning sys- solved examples that match that context and in--
tems can be integrated. One approach is to in- duced a new rule.
tegrate a number of learning systems that use
knowledge in the same form (e.g., classified exam- 2.4.2. Analogy
ples). The new integrated learning system consists
of a tool box of learning techniques. In order to Analogy systems try to solve one problem by
make this integrated tool useful, a program is looking at solutions to other types of problems.
written to determine which of the tools should be These systems must first find a mapping from the
used for a given problem, An even better system new problem to a previously solved problem. The
decomposes the problem into subproblems and relevant aspects of the solved problem are trans-
Computers in Industry B.L. Whitehall, S. C-E Lu / Machine learning in engineering automation 95

formed to apply to the unsolved problem. Per- a set of examples, learning from example systems
forming these mappings is very difficult. The anal- determine which aspects of the data distinguish
ogy system must generalize aspects of the prolems between the classes. The concept description pro-
to find a match and to apply the solution to the duced by LFE captures the important attributes
new problem. Background knowledge about the and values for a particular class. An engineer can
problems and solutions is required to help guide use this description to better understand the data.
this induction and problem selection. (Hall [16] When each event has a large number of attributes,
provides a comparison of a number of analogy the class descriptions produced by LFE remove
systems.) It is difficult to use analogy in engineer- the unimportant attributes and allows a user to
ing systems today because the approach is com- focus on the important attributes. Although learn-
plex and ill-defined. However, research is under- ing systems do have biases, they are logical. (Typi-
way in using analogy to assist in mechanical des- cal biases for machine learning systems are: pro-
ign [17]. duce short descriptions, produce descriptions that
cover the most examples, produce as few disjuncts
as possible.) When humans create knowledge from
3. Machine learning used in engineering today data, ill defined biases often influence the results.
An example of how machine learning al-
Although machine learning systems cannot gorithms can be used to construct a comprehen-
construct large knowledge-based systems for en- sive model from data is presented in [13]. The
gineering at this time, learning systems are useful AIMS systems is used to construct a model of a
to engineers in a number of areas. In this section a simulator for tool cutting. AIMS induces rules defi-
number of general engineering problems are de- ning the outcome of the simulator for different
scribed along with how machine learning systems classes of input examples. The system can produce
can assist in solving those problems. In each area rules that are very accurate but complex, and less
described, at least one machine learning approach accurate rules that are easier to comprehend. Thus,
has been applied and the results of the learning the system can produced a hierarchy of rules from
systems have been used by engineers or in con- least c o m p l e x / a c c u r a t e to most c o m p l e x /
junction with other computer systems. accurate. The rules can be used for different deci-
sion making processes during engineering depend-
3.1. Data and knowledge synthesis ing upon the time available to construct a model,
the accuracy of results currently required, and
The knowledge synthesis task is defined as the desired detail.
transformation process from a set of raw data into Discovery systems also create knowledge about
knowledge that can be used to fulfill a specific data that is useful to an engineer. Conceptual
goal in a particular domain. This synthesis process clustering programs place events into classes that
is commonly seen in engineering decision making have clear concept descriptions. While numerical
activities where, given a set of facts (from experi- taxonomy systems can also produce clusterings,
ments a n d / o r simulations), the engineer is ex- often no comprehensive description of the cluster-
pected to come up with a model (or an interpreta- ings can be found. Conceptual clustering systems
tion) which properly describes the given facts and have been used for group technology [18]. CLUS-
accurately predicts new facts. Traditional tech- TER/2 [19] was given a set of parts described with
niques such as statistical regressions, optimization a set of features. The system constructed a classifi-
methods, etc., have been used to aid such processes. cation of the parts based upon the conceptual
Regardless of their respective strengths and weak- cohesiveness of the parts. The descriptions pro-
nesses, a common limitation of these traditional duced had meaning to the engineers using the
techniques is that they are difficult to comprehend clusterings.
and subject to human bias.
Machine learning systems can be used to aid 3.2. System integration
synthesis processes. Learning from example sys-
tems and conceptual clustering (discovery) sys- After knowledge has been systhesized from
tems create knowledge about data. Provided with data, it can be integrated with the qualitative
96 / M S ' 9 1 - - L e a r n i n g in I M S ~ :~mpute::~ m lndusfr;

Problem

Solution
_j- 1
~Unsolved / Rules
\ // \ ,...Ut
~¢ules

Example ':///
Examples

Fig. 2. Knowledge-based learning approach to knowledge integration.

knowledge acquired from domain experts through expert system cannot solve a problem, the KBL
interviews. Therefore, inductive learning, which is system is called. KBL uses the knowledge provi-
used as a tool for knowledge synthesis, also pro- ded by the engineer (the same knowledge used b~.
vides a base for the knowledge integration task to the expert system) to guide an inductive module
develop more comprehensive domain models- a within KBL. The induced rules '.~re then added t~,
complete collection of all the available knowledge the knowledge base to allow the problem to be
about a specific task of interest. This integration completed.
can occur in two distinct ways. First, the rules
produced by an inductive learning system can be 3. 3= Prediction
presented to an expert during knowledge acquisi-
tion. The expert can then critique the machine Machine learning systems can be used to pre--
generated rules. It is often easier for an expert to dict outcomes of events in a couple of ways. First,
modify rules that are almost correct than to gener- the systems can be used to predict the class of a
ate rules from nothing. The I M L rules can assist given input example. As shown earlier, IML sys-
in getting more data from the expert. A second tems produce concept descriptions by generalizing
form of integration occurs when an expert cannot the input events. If a new example is seen that
provide rules for a detailed phenomenon, but can matches the concept description for a class, we
provide examples of the phenomenon. The expert can predict that the event belongs to that class.
can provide higher level rules for the system, and The concept description does not have to specify a
an expert system can learn rules for the lower level single class. The Probabilistic Learning System
specifics. (PLS) [20] learns rules that classify an event into a
Knowledge-based learning systems can be used category with a probability. For example, if an
to integrate knowledge from the engineer and the event matches one concept description it would
rules produced by an inductive learning system, as have an 80% chance of being a positive event and
shown in Fig. 2. When the inference engine of an a 20% change of being a negative event. Another
Computers in Industry B.L. Whitehall, S. C- Y. Lu / Machine learning in engineering automation 97

description might give the event a 40% chance of the final product. Determining when there exists a
being positive. Subsymbolic systems are also use- shortage of parts of a subcomponent before it
ful for this task. impacts the main line is a difficult task. With
Inductive learning systems have been used for knowledge about part queue lengths and resulting
computer security [21]. These systems learn rules outcomes (gathered from past experiences or a
that characterize the behavior of a computer user. simulation of the assembly process), machine
The rules predict the sequences of commands nor- learning systems can induce rules that can indicate
mally typed in by a user. When the predicted where future shortages may occur. These rules can
pattern is not matched, the system notes the ex- be given to a human production manager or ad-
ception. When enough exceptions are identified. ded to knowledge systems that control production.
the system indicates that a user is not following
their normal patterns, and the account security
might be violated. 4. Machine learning use in engineering in the
Learning systems can also be used to predict future
the values of attributes of events based upon the
other known attribute values. Discovery systems Before machine learning becomes a robust tool
such as COBWEB [10] can be used to construct that can be used in engineering for a wide range of
classifications of events. If the system is then tasks, a number of areas need to be researched.
given an event with a known class, but some This section outlines those important issues for
attribute values are missing, those attribute values future research. In many of these areas, research is
can be inferred. The neural net approach of inter- already underway; however, it will be some time
attire activation and competition [22] allows a net- before results from the laboratory can be used to
work to infer missing attribute values of an incom- construct tools for engineering.
plete event.
4.1. Utilizing existing knowledge
3.4. Adaptive monitoring control
A weak point of most current machine learning
Machine learning systems can be used to moni- systems is the lack of consideration for knowledge
tor and control complex processes. If those available from sources in addition to the input
processes can be decomposed into a fixed number examples. Deductive learning does use outside
of s t a t e / a c t i o n pairs then symbolic learning is knowledge, but assumes that knowledge is com-
appropriate, otherwise subsymbolic computing plete and correct. Learning systems need to take
might give better results. To apply learning to advantage of the knowledge that engineers can
such situations, the systems must see examples of provide easily. Many LFE systems allow attributes
the relevant states along with the actions required. to be weighted, so that an engineer could specify
This decomposition makes the control problem the attributes thought to be the most important or
similar to a prediction problem, where the result that he would prefer be used in the synthesized
of matching a concept description (state) is a set rules. As learning systems work on more complex
of actions to be performed. For the system to be problems it is important that they be able to use
truly adaptive, the system needs to learn incre- more complex forms of knowledge, such as rules
mentally. Each new event is stored with the com- and scripts. Rules could help the system to con-
puted action and the quality of the result. The struct new attributes (see below) and realize when
learning system can learn new rules that improve it is suggesting inappropriate answers. Scripts
the performance of the overall system. To guide could help learning systems understand the se-
the learning system, there must be a well defined quence in which certain tasks occur, such as as-
notion of success or improvement so the learning sembling a class of parts.
system can evaluate it previous performance. Knowledge in engineering can come from a
As an example, machine learning can be used variety of sources besides the engineer. Other
to construct rules to assist in managing an assem- computer tools used by the engineer such as com-
bly line. In most complex assembly lines, subcom- puter-aided design (CAD) and finite element anal-
portents must be assembled before being added to ysis (FEA) systems can provide useful knowledge.
98 1 M S '91---Learning in I M S ~ ~mpulerr m lnduslrv

Learning systems need to be able to tap into this propriate for machine learning as well. 'Fhe most
information and use it to construct rules that informative method of distinguishing between a
consider all the complexities of engineering prob- set of events may not be the actual attribute
lems. values provided. Constructl~,e mduc~+ion is the pro°
cess of adding higher-level attributes to the exam-
4.2. Learning in structural domains ples [25]. The higher-level descriptors can then be
used to construct the synthesized rules. The pro-
Engineering problems involve complex struc-
tess is constructive, because it iakes provided at-
tures that require rich representation, such as first
tributes and combines them or applies known
order predicate calculus. Propositional logic is not
operations to them in order to build the new
adequate. Most machine learning systems deal
attributes. Such processes include logical connec-
only, w i t h attribute/ value pairs. An
tives between binary valued attributes and
attribute/value representation cannot describe a
mathematical operations with ,numeric attributes,
situation were one object is on top of, next to, or
Future systems will be able to construct attributes
inside of another object. Objects are treated as
based upon knowledge provided bv engineers,
whole entities with a uniform set of properties. In
order for machine learning to be useful for en-
4.5, Integration o1 machine tearntn~ with existm~
gineering problems such as design and planning,
engineering software
the systems must be able to work with structure.
Some work has been done in this area (e.g., [7,23]).
One of the most challenging aspects of making
machine learning useful to engineers is integrating
4. 3. Learning with continuous tJariables
learning techniques with existing engineering
Engineering problems require the use of con- software. Getting learning systems to work with
tinuous (real) numbers. Machine learning systems database systems, C A D systems and FEA sys,
(typically) deal only with discrete values. The terns is a complex task. Learning systems need lo
world might provide real quantities but some dis- use information from these systems, as described
cretization always occurs over the values that are earlier. However, learning systems need to provide
actually used (measurements do not have infinite knowledge to these systems and assist the engineer
precision). Dealing with large numbers of discrete in making better use of such ~ystems. Learning
numbers is awkward and limiting (e.g., there is not could help control the interface t~ such tools and
a value to indicate the average between two ad- make them more flexible to accommodate individ-
jacent discrete values). Some machine learning ual work habits. Learning systems could learn
systems have been modified to work with continu- procedures for the tools and when they should be
ous values [24], but at this time, there is not a well applied. As an example, a learning system for an
defined approach for learning when some attri- FEA algorithm might learn when different types
butes are real valued and others are nominal. of meshes should be automatically generated for
Engineering problems also deal with mathe- parts.
matical formulas. There exist some discovery sys-
tems that can find mathematical formulas but 4.& Exploratory wstems
they cannot handle all the complexities of experi-
mental data. They do not handle noise and impre- While simple discovery systems currently exis~
cision (inexact values). Learning from example and are useful for some engineering problems,
systems do not work with formulas at all. This future systems will be more powerful. With the
facility is important if they are to distinguish ability to use information provided by the en-
between complete sets of data describable with gineer, such systems could perform experiments
equations. and explorations in the background. The results
can be presented to the engineer for future consid-
4.4. Constructiee / rneta learning eration. Exploratory systems need to be integrated
with databases containing engineering informa-
Considering abstractions of problems often tion. The systems can use the engineering knowl-
helps one to find a solution. This idea is ap- edge to discover important relationships and inter-
Computers in Industry B.L. Whitehall, S. C- E Lu / Machine learning in engineering automation 99

esting aspects of the data. Additionally, the infor- An AI Approach, Morgan Kaufmann. Los Altos, CA,
1983, Vol. 1, pp. 3-24.
mation can be used to guide experiments and to
[2] Thomas M. Mitchell, Richard M. Keller and Smadar T.
complete partially solved problems. Kedar-Cabelli, "Explanation-based generalization: A uni-
fying view", Mach. Learn., Vol. 1, No. 1, January 1986,
pp. 47-80.
[3] Gerald DeJong and Raymond Mooney, "Explanation-
5. Conclusion based learning: An alternative view", Math. Learn., Vol.
1, No. 2, 1986, pp. 145 176.
Machine learning is useful in its current state to [41 A.M. Segre, Explanation-based learning of generalized
help with many engineering problems. Learning robot assembly plans, PhD thesis, Computer Science De-
partment, University of Illinois, Urbana, IL, 1987.
can be used to synthesis knowledge from complete
[5] J. Ross Quinlan, "Induction of decision trees", Math.
models and data. The synthesized knowledge as- Learn., Vol. 1, No. 1, 1986.
sists in the integration of models with other forms [6] T.G. Dietterich and R.S. Michalski, "A comparative re-
of knowledge useful in engineering problem solv- view of selected methods for learning from examples", in:
ing. Learning creates rules that can predict future R.S. Michalski, J.G. Carbonell and T.M. Mitchell (eds.),
Machine Learning: An AI Approach, Morgan Kaufmann,
outcomes based upon past performances, an im-
Los Altos. CA, 1983, Vol. 1, pp. 41 82.
portant type of classification. [7] W.A. Hoff, Ryszard S. Michalski and Robert E. Stepp,
In order for machine learning to become as INDUCE 3: A program for learning structural descriptions
useful to engineers as the pocket calculator, learn- from examples, Tech. Rep. UIUCDCS-F-83-904, Com-
ing systems must be able to handle all the data puter Science Department, University of Illinois, Urbana.
IL, 1983.
types and knowledge engineers manipulate. The
[8] R.E. Stepp and R.S. Michalski, "'Conceptual clustering:
systems must integrate with current systems used Inventing goal-oriented classifications of structured ob-
by engineers for more beneficial results. To con- jects", in: R.S, Michalski, J.G. Carbonell and T.M,
struct competent machine assistants for engineers, Mitchell (eds.), Machine Learning: An At Approach, Los
learning systems must be used to generalize an Altos, CA, 1986, Vol. 2, pp. 471- 498.
[9] R.E. Stepp, Conjunctive Conceptual ('lustering: A Meth-
engineer's actions and interface those actions with
odology and Experimentation, PhD thesis, Computer Sci-
the other engineering tools. Learning systems also ence Department, University of Illimfis, Urbana. IL, 1984.
need to become more robust by determining which [1(1] D.H. Fisher, '" Knowledge acquisition via incremental con-
machine learning technique is most appropriate ceptual clustering", Math. Learn., Vol. 2, No. 2, 1987, pp.
for a given problem and by performing deductive 139 172.
[11] Brian Falkenhainer and Ryszard S. Michalski, "'Integrat-
reasoning when appropriate and inducing knowl-
ing qualitative and quantitative discovery: The ABACUS
edge when background knowledge is incomplete. system", Math. Learn., Vol. 1, No. 4, 1986, pp. 367 401.
[12] Pat Langley, Herbert A. Simon, Gary L. Bradshaw and
Jan M. Zylkow, Scientific Discovery: Computational Ex-
plorations of the Creatir,e Process, MIT Press, Cambridge,
Acknowledgment MA, 1987,
[13] S.C-Y. Lu and D.K. Tcheng, "Building layered models to
Funding for the research projects described in support engineering decision making: A machine learning
this paper are provided by Allied Signal Aero- approach", J. Eng. Ind., ASME Trans., February 1991.
[14] B.L. Whitehall, Knowledge-based learning: Integration of
space Company, Digital Equipment Corporation,
deductive and inductive learning for knowledge base com-
and National Science Foundation. The develop- pletion, PhD thesis, Computer Science Department, Uni-
merit of the inductive/deductive decision-making versity of Illinois, Urbana, IL, 1990: Also published as
framework also benefits from general research technical report no. UIUCDCS-R-90-1637.
grants from Cray Research, Inc., Ford Motor [15] J.G. CarbonelL "Learning by analogy: Formulating and
generalizing plans form past experience", in: R.S. Michal-
Company, and Xerox Corporation to our labora-
ski, J.G. Carbonell and T.M. Mitchell (eds.), Machine
tory. These supports are greatly appreciated. Learning." An ,41 Approach, Morgan Kaufmann, Los Al-
tos, CA, 1983, Vol. l, pp. 137-162.
[16] Rogers P. Hall, "Computational approaches to analogical
reasoning: A comparative analysis". Arti£ lntell., Vol. 39,
References No. 1, May 1989, pp. 39-120.
[17] T. Bardasz and I. Zeid, ~'Proposing analogical problem
[1] J.G. Carbonell, R.S. Michalski and T.M. Mitchell, "'An solving to mechanical design", in: Proc. 1991 NSF Design
overview of machine learning", in: R.S. Michalski, J.G. and Manufacturing Systems Conf., Austin, TX, January
Carbonell and T.M. Mitchell (eds.), Machine Learning: 1991. pp. 389-394.
100 1MS "91--Learning in I M S omputer~ m Indu.~tr;

[18] S.C-Y. Lu, "Machine learning techniques for group tech- competitive learning", m: 1).1:,. Rumeltlart, ,I.L. McClcl-
nology applications", C1RP Ann., Vol. 39, No. 2, 1989. land and the PDP Research Group ieds,L Parallel Distrib-
[19] R.S. Michalski and R.E. Stepp, "Learning form observa- uted Processing: Explorations in th< .,lqtcros:ructure ¢~f Cog-
tion: Conceptual clustering", in: R.S. Michalski, J.G nitum, M1T Press. Cambridge, M A 1086 Vo]. !. Chap 5
Carbonell and T.M. Mitchell (eds.), Machine Learnin~ pp. I5! [93
An ,41 Approach, Morgan Kaufmann, Los Altos, ( A i23] J.R. Quinlan, " [ . e a [ n m g ic,gv, ii defmitiolt> iI~>i~
1983, Vol. 1, pp. 331-364. relations" $1ach. L e a r n Vol ~ ",~, ~ 199IL
[2t)] L. Rendell, "A general framework for induction and a 124] B.L Whitehall. S.(-Y. l u md R . i Stepp, ':('aq: q
study of selective induction". Mach. Learn., Vol, 1. Nt~ 2 machine learning too] for e n g i n e e r i ~ g ' /!~: ~1 4r:t/i l:~,~e/,~
1986, pp. 177-226. Eng.. 1991.
[21] H.S. Teng, K. Chen and S.C-Y. Lu, "Security audit trail [25] l.. Rendell, R. Seshu and D. I'chctig., '~ More iobust c o u
analysis using inductively generated predictive rules", in cept learning using dynamically ,~ariablc bia,,,", m: Pro<
Proc. 6th I E E E Cor~/~ on Artificial Intelligence Applica- 4th Int. Machine Learning Works/m/,, Morgan Kaufmann.
t wns , 1990. l.os Ahos. CA. 1087. pp. 66 ":~
[22] D.E. Rumelhart and D. Zipser, "t=eature discover,, b'~

You might also like