Constructability Analysis - Machine Learning Approach

CONSTRUCTABILITY ANALYSIS: MACHINE LEARNING ApPROACH
By Mirostaw Skibniewski, l Tomasz Arciszewski,2 and Kamolwan Lueprasere
ABSTRACT: Computerized constructability analysis, using knowledge-based tools, is the key to effective con-
struction process automation. However, the development of such tools requires a prior acquisition of construct-
ability knowledge. In this paper, results of a feasibility study of automated constructability knowledge acquisition
are reported. In the conducted research, constructability of a beam in a reinforced-concrete frame has been
investigated. For this problem, a knowledge representation space has been developed, relevant constructability
data has been acquired from industry, and a collection of examples has been prepared. In this collection, each
example represents a structural design concept evaluated from the point of view of its constructability. The
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/23/18. Copyright ASCE. For personal use only; all rights reserved.
collection of examples has been used by a learning system to acquire from them constructability knowledge in
the form of decision rules. The experience gained while conducting the automated knowledge acquisition process
has been used to develop initial methodological conclusions regarding the use of learning systems in construct-
ability analysis and to determine the future research directions.
INTRODUCTION result is consistent with the observed knowledge acquisition

bottleneck in civil engineering (Modesitt 1992), which signif-
Constructability is an important feature of a structural de- icantly delayed the development and practical use of knowl-
sign and the construction project site conditions. It detennmes edge-based systems in this domain. For these reasons, we be-
the level of complexity of executing the associated structural lieve that the crucial issue in constructability analysis is the
assembly tasks. Also, it characterizes the feasibility and sim- acquisition of formal knowledge. Since manual acquisition is
plicity of construction of a given structural design concept. usually insufficient for both research and practical purposes
Constructability is determined by all design decisions with im- (Arciszewski and Ziarko 1992), a novel approach to construct-
pact on the construction process. These decisions are made ability knowledge acquisition is needed. Considering the im-
considering construction process-related constraints, which are portance of cost, time, and resource management in construc-
called constructability constraints here. tion engineering, any improvement in constructability analysis
Computerized constructability analysis is a key to effective and structural design incorporating this analysis will have an
construction process automation. However, it is impossible to impact on the productivity and competitiveness of the con-
develop knowledge-based tools for such analysis without ac- struction industry.
quiring formal constructability knowledge for various struc- The main objective of this paper is to describe a novel ap-
tural systems. This knowledge cannot be manually acquired proach to constructability knowledge acquisition based on the
because of the complexity of problems involved. Also, the use of machine learning and to demonstrate its feasibility. It
domain experts find it extremely difficult, if not impossible, to has been accomplished describing the entire knowledge ac-
articulate complicated relationships among many design de- quisition process dealing with constructability of a beam in a
cisions regarding a given structural system and its construct- reinforced-concrete frame, in which a learning system was
ability. However, construction experts can easily provide a used. In this process, a knowledge representation space has
large number of various examples of different structural design been prepared for the problem considered, constructability
concepts that were evaluated considering their constructability. data has been obtained from industry, design examples have
These examples can be used to acquire constructability knowl- been prepared, and constructability knowledge has been
edge through machine learning. learned from these examples.
A prerequisite for the development and the successive use
of knowledge-based tools in constructability analysis is the CONSTRUCTABILITY RESEARCH
acquisition of formal knowledge about the constructability of
a given structural system, or a class of systems, considered in Software developed for design purposes in the 1980s led to
a given design case. By formal knowledge, we understand a early attempts to use heuristics derived from explicit human
collection of decision rules which can be incoporated in a experience in a limited compilation of constructability knowl-
knowledge-based system. Unfortunately, constructability edge (Kirby and Cannalte 1988; Kirby et al. 1991). Some con-
knowledge is interdisciplinary (structural design, construction ceptual work has been accomplished in attempting to improve
process engineering, human factors and ergonomics, construc- the design review process (Kirby et al. 1989). At present, there
tion management, cost engineering, automation technology, are new promising tools, such as learning systems, enabling
etc.), complex, and poorly understood. Therefore, the manual better development of the various tools for constructability
acquisition of this knowledge is very difficult, if not impos- analysis and structural design for constructability. This need
sible, as demonstrated in the study by Kirby et al. (1991). This was recognized by the Japanese construction industry (Lue-
prasert and Skibniewski 1994; Yu and Skibniewski 1995).
'Prof. of Civ. Engrg., Purdue Univ., West Lafayette, IN 47907-1294. During the mid-1970s and early 1980s, several major con-
2Assoc. Prof. of Urban Sys. Engrg., George Mason Univ., Fairfax, VA
struction owners and engineering construction firms recog-
22030-4444.
3Asst. Prof. of Civ. Engrg., King Mongkut lost. of Techno!., Pibul-
nized the inherent shortcomings and limitations of the tradi-
songkram Rd., Bangkok yai, Bangkok, Thailand; fonnerly, Grad. Res. tional owner-designer-contractor interaction. In the 1970s,
Asst., School of Civ. Engrg., Purdue Univ., West Lafayette, IN. research on constructability concepts was conducted in the
Note. Discussion open until June I, 1997. To extend the closing date United Kigdom by the Construction Industry Research and
one month, a written request must be filed with the ASCE Manager of Information Association (CIRIA). This organization defined
Journals. The manuscript for this paper was submited for review and constructability as "the extent to which the design of a build-
possible publication on August 21, 1995. This paper is part of the Joumal
of Computing in Civil Engineering, Vo!. 11, No. I, January, 1997 ing [or other structure] facilitates ease of construction, subject
©ASCE, ISSN 0887-3801/97/0001-0008-0016/$4.00 + $.50 per page. to the overall requirements for the completed [structure]." As
Paper No. 11420. a major constructability contribution, CIRIA considered fac-
8/ JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 1997
J. Comput. Civ. Eng., 1997, 11(1): 8-16

tors that are within the sphere of influence of a project design derings depicting building floor plans, cross-sectional areas of
team (Jergeas et al. 1991). slabs, column sizes, steel and concrete structural connections,
In the mid-1970s, Procter & Gamble, Inc. developed man- etc. On the other hand, construction know-how data, although
uals on constructability (or buildability) based on the results usually poorly structured and/or proprietary, may be provided
of a number of internal corporate studies aimed at document- in the form of specifications, work procedures, task sequences,
ing the benefits of a formal constructability analysis. The man- performance characteristics of suitable construction equip-
uals described constructability objectives, specific procedures ment, physical clearances required to accomplish structural as-
to be adopted for achieving the specific goals, and responsi- sembly tasks, labor and time required to complete the desired
bilities of individual project team members. In the early 1980s, task, estimated cost of task completion, etc.
Texaco, Inc. produced its own constructability reference man- Design constructability and the associated design of the con-
ual, with particular emphasis on modularization and preassem- struction site processes have been an essential part in deter-
bly. The manual included specific project guidelines, an en- mining technical solutions in automated building construction
gineering checklist, and a construction checklist (Russell et al. systems launched by several Japanese construction firms in
1992a). Procter & Gamble and Texaco were among the first early to mid-1990s. This includes both the structural and con-
organizations in the United States to develop formal construct- struction-related concerns essential to the technical feasibility
ability improvement programs (O'Connor et al. 1986), which of a construction project in question. It is widely accepted that
led to further research into the structured approach to con- the essential conditions for a successful dissemination of the
structability improvement methods. The Business Roundtable automation, robotics, and other related emerging technologies
report titled "Integrating Construction Resources and Tech- in the construction industry include the redesign of many types
nology into Engineering" (1982) addressed most of the con- of structures and construction-site structural assembly proce-
structability concerns of the construction industry, based on 35 dures to meet the requirements of the new technology [see
interviews with owners, designers, and contractors in com- e.g., Viscomi et al. (1993)]. Only then the benefits to be
mercial and industrial sectors of the construction industry. Sev- brought about by implementing the emerging technologies
eral recommendations to broaden and strengthen the integra- such as decrease in labor costs, increased productivity, and
tion of design and construction were included. quality of the constructed project can be realized. However,
The term "constructability" became well known in the this can only happen when the issues of constructability of the
United States in the mid-1980s after a study undertaken by newly developed structural designs suitable for innovative site
the Construction Industry Institute (Cll) at the University of assembly are properly addressed through the development of
Texas at Austin defined the needs for a comprehensive rela- a new technology that will enable the acquisition of formal
tionship between owners, architects, designers, construction constructability knowledge.
contractors, and facility users (Tatum 1987). However, rela-
tionships identified in the Cll studies have been limited only MACHINE LEARNING IN DESIGN KNOWLEDGE
to the project-management level and, as such, they were useful ACQUISITION
only for the company and project site management, rather than
for the structural designers. The designers are primarily con- It has been discovered that a human being has an unsur-
cerned with the impact of various structural design decisions passed ability to conduct deduction, i.e., to use general knowl-
on the level of complexity in erecting the given structure. In edge to deal with specific cases. However, induction, i.e., gen-
other words, no analysis of more specific relationships at the eralization of knowledge from individual cases, is difficult for
process and task levels has been provided, particularly in the human beings who usually can handle only a small number of
context of conceptual design. Such relationships require both cases (seven) and attributes (also seven) at the same time. This
design- and construction-specific knowledge, scattered among limitation is caused by the small capacity of the human short-
all parties involved in the project and available mainly in an term memory and is sometimes referred to as the 7n human
implicit rather than explicit form. This makes these relation- memory limit (Arciszewski and Rossman 1992). This limit
ships difficult to extract and expensive to compile using tra- means that human beings are not able to effectively acquire
ditional data collection and analysis means. Cll established knowledge in the form of decision rules in complex domains
three task force committees addressing the issues of construct- that are described by many attributes and/or from a large num-
ability: constructability, constructability implementation, and ber of examples on the order of hundreds or thousands. For-
modularization. Several source documents were produced as a tunately, induction can be relatively easily conducted by a
result of research by those committees which consisted of sen- computer program called a "learning system," which uses
ior representatives of owner firms, construction contractors, machine learning methods to produce decision rules from ex-
and academics. In these documents, various constructability amples of previous decisions. In this case, the number of at-
concepts were presented for each of the three major stages of tributes considered and the number of examples to be used are
facility delivery: conceptual planning, design and construction. mostly limited by the available computer memory and there-
As explained earlier, planning, design, and construction fore this number can be increased when necessary. For this
have been historically fragmented processes involving multi- reason, the use of a learning program may overcome the 7n
ple parties: owners, architects and engineers, contractors and human memory limit and may allow humans to expand their
subcontractors, materials and equipment suppliers, and others. inductive capabilities through the use of machine learning. Re-
As a result, a significant amount of information related to de- search on machine learning in computer science has led to the
sign, construction and anticipated maintenance procedures is development of a number of both experimental and commer-
regularly being lost. Numerous studies in the United States cial learning systems. Such systems are particularly appropri-
engineering construction industry and academia have noted ate for constructability knowledge acquisition because of the
this problem (Tatum et al. 1985; O'Connor et al. 1986; "Con- complexity of the domain and the availability of a large num-
structability" 1987; Fischer 1989; Tatum 1990; Jergeas et al. ber of evaluated examples of previous designs. The use of
1991; Russell et al. 1992). However, no formal method exists decision support systems based on rules is also much more
which would enable a direct inference of an optimal construc- convenient and convincing for engineers than the use of sys-
tion technique to execute a specific project, based on current tems based on subsymbolic learning, such as genetic algo-
structural design. Specific structural design may be provided, rithms (Grierson and Pak 1992; Haleja 1989) or neural net-
for example, in the form of technical drawings or other ren- works (Ivezic and Garrett 1993), particularly that knowledge
JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 1997/9
J. Comput. Civ. Eng., 1997, 11(1): 8-16

in the form of rules is "transparent," and can be easily com- ski (1983). The algorithm was developed to learn classification
prehended by the system's users. (decision) rules from a collection of examples. In this collec-
An extensive review of the state of the art of machine learn- tion, each example is described by a number of independent
ing in engineering knowledge acquisition has been provided attributes and their values and by a single dependent (decision)
in Arciszewski et al. (1994). An updated version of this review attribute and one of its values, usually called "categories of
is included in this paper for completeness. the decision attribute." Produced decision rules are the rela-
There have been several previous studies on the application tionships between various groups of independent attributes and
of machine learning to knowledge acquisition. Gero et al. their values and the dependent (decision) attribute. These rules
(1992), Mackenzie and Gero (1987), and McLaughlin and can be used to classify unknown examples to one of the cat-
Gero (1987) used an experimental learning system based on a egories of the decision attribute. In the learning process, the
decision tree learning algorithm (ID3) to acquire architectural most general rules are sought, called "discrimination rules,"
design knowledge. EXTASY, an expert system for analysis and for the individual categories of the decision attribute. Each
design of microwave towers, has been developed with a learn- such rule is supposed to discriminate examples belonging to a
ing component and used in experiments demonstrating the ac- given category of the decision attribute from all examples be-
quisition and use of design knowledge in conceptual design longing to the remaining categories of this attribute.
(Murlidharan et al. 1991). Reich (1992) and Reich and Fenves The AQ15 algorithm uses the concept of STAR (Michalski
(1992) implemented BRIDGER, an application of the CLUS- 1983). A STAR of the event e (a given example belonging to
TER clustering algorithm initially developed by Fisher (1987). the decision attribute category c) against the event set E (a
BRIDGER was used to acquire bridge design knowledge. collection of all examples which do not belong to the decision
Reich and Fenves (1988) also studied knowledge acquisition attribute category c) is defined as a set of all maximally gen-
for floor system design in buildings using an experimental sys- eral conjunctive expressions that cover event e and that do not
tem based on the SOAR learning system (Laird et al. 1986). cover any of the examples in set E. In this case, e (an event)
Maher (1992) and Maher and Li (1992,1993) have been work- is a positive example of a concept to be learned and E is a set
ing on automated acquisition of preliminary design knowledge of some counter examples of this concept, called negative ex-
in which conceptual clustering is augmented by numerical amples. In practical problems, a STAR of an event may con-
methods of linear regression analysis and probabilistic ap- tain a very large number of descriptions. Consequently, in the
proaches to pattern identification. A genetic algorithm was learning process such a theoretical star is replaced by a
used by Maher and Kundu (1993) to develop a system for bounded star that contains no more than a fixed number of
adaptive design. The same algorithm was also used by Grier- descriptions. These descriptions are selected as the most pref-
son and Pak (1992) in a system for the optimization of con- erable descriptions, according to the preference criterion de-
figurations in the conceptual design of skeletal building struc- fined in the problem background knowledge. A general algo-
tures and by Hajela (1989) for structural synthesis. ROUGH, rithm using STAR can be described as follows (Michalski
a learning system based on the theory of rough sets (Pawlak 1983; Michalski et al. 1986):
1982; Ziarko 1989), was used by Arciszewski et al. (1987)
1. Randomly select a positive example.
and Mustafa and Arciszewski (1992) for design knowledge
2. Generate a bounded star of that example against the set
acquisition in the area of structural design of wind bracings in
of negative examples. In the process of STAR generation
tall buildings. Ivezic and Garrett (1994) described an experi-
apply generalization rules, task-specific rules, heuristic
mental learning system, NETSYN, based on a connectionist
for generating new descriptors supplied by problem
learning approach, for the acquisition of conceptual design background knowledge, and definitions of previously
knowledge. Also, Adeli and Yeah (1989) used a neural net- learned concepts.
work to learn engineering design principles. Milzner and Har- 3. In the obtained star, find a description with the highest
becke (1992) were involved in the development of an experi- preference according to the assumed preference criterion.
mental learning system, LEAR, based on constructive 4. If the found description covers a set of positive examples
induction, for learning design knowledge. This system, how- completely, go to step 6.
ever, has not been used yet for experiments with engineering 5. Otherwise, reduce the set of positive examples to contain
design examples. Whitehall et al. (1990) and Lu and Chen only events not covered by the learned description and
(1987) applied machine learning to design knowledge acqui- repeat the whole process from step 1.
sition using several experimental learning systems. The AQl5 6. The disjunction of all generated descriptions is a com-
and AQl7 learning systems, employing the STAR methodol- plete and consistent concept description. As a final step
ogy (Michalski 1983), were applied by Arciszewski and Dy- apply various reformulation rules defined in background
bala (1992) for learning design rules in the area of conceptual knowledge in order to obtain simpler expression.
design of wind bracings in steel skeleton structures of tall
buildings. The same collection of examples was used by Ar- The central step in the foregoing algorithm is the generation
ciszewski et al. (1994) for various design knowledge acqui- of a bounded star. This can be done using a variety of methods.
sition experiments with constructive induction. Thus, the algorithm can be viewed as a general schema for
A growing interest in the applications of machine learning implementing various learning methods and strategies. The
to design knowledge acquisition can be observed. For exam- AQ15 algorithm, which directly utilizes the concepts of STAR,
ple, during the last four ASCE conferences on computing in can be tuned using many input parameters that control exe-
civil engineering (1993, 1994, 1995, and 1996) several ses- cution of the program. For example, two possible modes of
sions on machine learning in engineering were held. Also, dur- operation, generalization and specialization, can be used. In
ing the 1995 Joint International Conference on Artificial In- the generalization mode, the learning system induces rules as
telligence, in Montreal, Canada, a similar session took place. general as possible, i.e., they involve the minimum number of
In the research reported in this paper, INLEN (inference and attributes, each with the maximum number of attribute values.
learning) experimental learning system was used. INLEN was In the specialization mode, the learning system generates rules
developed in the George Mason University's Laboratory for as specific as possible, i.e., with the maximum number of at-
Machine Learning for the PC platform and is available for tributes and the minimum number of attribute values. Then,
research and engineering purposes. The program is based on redundant values are removed from extended selectors in the
the learning algorithm AQl5, originally proposed by Michal- rule.
101 JOURNAL OF COMPUTING IN CIVIL ENGINEERING 1 JANUARY 1997
J. Comput. Civ. Eng., 1997, 11(1): 8-16

TABLE 1 Dependent and Independent Attributes for Beam Structural Analysis and Their Values
Attributes Attribute values
(1 ) (2)
Reinforcement ratio (ReRa) Low Average High
Reinforcement ratio of the first beam-to-column connection (CoBeRal) Low Average High
Reinforcement ratio of the second beam-to-column connection (CoBeRa2) Low Average High
Number of attaching slabs (NoSla) None One SameTwob DiffTwod
Number of attaching walls (NoWall) None One SameTwo DiffTwo
Changes of beam on the first side or left side (BeChal) None SlichgeReinfa AlIChgeReinf' WDChangee AlIChange'
Changes of beam on the second side or right side (BeChal) None SlichgeReinf AlIChgeReinf WDChange AIIChange
Constructability evaluation (ConEva) Poor Good Excellent
aSlichgeReinf = Slight changes In reinforcement.
bSameTwo = Same slabs or walls on both sides.
eAllChgeReinf = All reinforcements were changed.

"DiffTwo = Different slabs or walls on each side.
ewoChange = Width and depth were changed.
'AllChange - Both reinforcement and size were changed.
FEASIBILITY STUDY 1. ReRa represents reinforcement ratio of the beam, and

three ratios are considered: low, average, high.
A study has been conducted to demonstrate the feasibility 2. CoBeRal represents the reinforcement ratio of the first
of machine learning in constructability analysis. It is briefly beam-ta-column connection with three values: low, av-
described and its selected results are reported next. erage, high.
The subject of the study was the constructability analysis of 3. CoBeRa2 represents the reinforcement ratio of the sec-
prefabricated beams in a 12-story building comprising a ond beam-to-column connection with three values as for
square-shaped concrete frame. This building had a bisym- attribute 2.
metrical square, 7 X 7 m, floor plan. For simplicity, the study 4. NoSla represents the number of slabs attached to the
was conducted assuming a construction process of a building beam. The maximum number of slabs is two. Therefore,
frame in which a conventional reinforced-concrete construc- four attribute values are: none, one, SameTwo, DifITwo.
tion system is used. Value one means there is one slab attached to the beam.
Value SameTwo means that two identical slabs are at-
Building a Knowledge Representation Space tached to the beam. Value DifITwo means that there are
two different slabs.
To conduct machine learning of constructability decision 5. NoWall represents the number of walls attached to the
rules, a knowledge representation space had to be developed beam. The choices are the same as in the NoSla attribute:
for the structural system considered. A knowledge represen- none, one, SameTwo, and DiffTwo. Values SameTwo
tation space is a collection of attributes and their values that and DifITwo represent the two walls, one above the
are used to characterize decisions regarding the problem under beam and one below. However, DiffTwo also represents
consideration. The knowledge representation space contains a the case in which the two walls are not identical.
number of attributes describing the building and its construc- 6. BeChal describes the changes in steel reinforcement and
tion, which are called independent attributes, and one char- size of the beam on the left or first side of the considered
acterizing the constructability of a given building that is de- beam. Five values were used: none, SliChgeReinf,
pendent on the independent attributes and is therefore called AllChgeReinf, WDChange, AllChange. Value none
dependent or decision attribute. In a properly developed means that the beam on the left side of the column has
knowledge representation space, the identified attributes exactly the same shape and reinforcement as the one on
should be sufficient to distinguish between any two different the right side. Value SliChgeReinf denotes that the team
solutions to the problem being considered. In our study, the on the left side has the same size but slightly different
problem is described by the symbolic, or nominal, attributes reinforcement. Value AllChgeReinf denotes the beam on
and their feasible values. For example, when a nominal attri- the left side with the same size but different reinforce-
bute reinforcement ratio is considered, its nominal values may ment. Value WDChange represents that the width or
be low, average, or high. These nominal values represent spe- depth of the beam on the left side has been changed but
cific ranges of variation of the reinforcement ratio. the reinforcement has not. Value AllChange represents
In our research, the constructability of a beam has been that the two beams are entirely different in size and re-
considered. This beam constructability can be designed with inforcement.
slabs and/or a secondary beams on the left and/or right side 7. BeCha2 represents the changes to the beam on the right
with respect to the columns and with the walls above and or second side as in the previous attribute regarding the
below the beam. In this case eight nominal independent at- beam on the left.
tributes describing the beam design were used, and they are
discussed in the following section. For the simplicity of our Individual attributes and their feasible values are shown in
study, only nominal attributes have been used. However, there Table 1.
are machine learning systems available that allow knowledge
acquisition using both nominal and numerical attributes. Preparation of Examples
The dependent attribute, or decision attribute, is construct-
ability evaluation (ConEva). It is a measure of the construct- A collection of 31 examples (see Table 2) was manually
ability of a given design and is assumed to have three values: prepared for this study. The examples represented various
poor, good, and excellent. The nominal values of this attribute beam structural design cases that were evaluated considering
are determined by experts on the basis of their experience and their constructability. All examples were developed for the
formal analysis of individual design cases. same building in terms of dimensions, loading, and other es-
The independent or condition attributes are as follows: sential design assumptions.
J. Comput. Civ. Eng., 1997, 11(1): 8-16

TABLE 2. Examples of Beam Design
Number ConEva ReRa CoBeRa1 CoBeRa2 NoSla NoWall BeCha1 BeCha2
(1 ) (2) (3) (4) (5) (6) (7) (8) (9)
1 Poor High High High DiffTwo SameTwo WDChange AIlChgeReinf
2 Poor Average High Average DiffTwo One SlichgeReinf WDChange
3 Poor Average Average High DiffTwo DiffTwo AIlChgeReinf None
4 Poor High Average High SameTwo One AIlChange SlichgeReinf
5 Poor Average High Average SameTwo DiffTwo AIlChgeReinf AlIChange
6 Poor Average Average High SameTwo One WDChange AllChange
7 Poor High Average Average DiffTwo SameTwo SlichgeReinf WDChange
8 Poor Average Average High Difftwo One WDChange SlichgeReinf
9 Poor Average High High DiffTwo DiffTwo WDChange AIlChgeReinf
10 Poor Average High Average DiffTwo SameTwo SlichgeReinf AIlChgeReinf
11 Poor Average High High DiffTwo None SlichgeReinf WDChange
12 Good Average Average Low None One SlichgeReinf SlichgeReinf

13 Good Low Average Low One SameTwo AIlChgeReinf SlichgeReinf
14 Good Average Average Average One One AIlChgeReinf SlichgeReinf
15 Good Low Average High DiffTwo None None SlichgeReinf
16 Good Low High Low One None SlichgeReinf None
17 Good Average Low High None SameTwo None AlIChgeReinf
18 Good Average High High SameTwo None SlichgeReinf SlichgeReinf
19 Good Low High High SameTwo None SlichgeReinf SlichgeReinf
20 Good Average Average Low None None AlIChgeReinf SlichgeReinf
21 Good Average Low Low None SameTwo AlIChgeReinf AIlChgeReinf
22 Good Average Low Low SameTwo One SlichgeReinf AIlChgeReinf
23 Good Average Average Low SameTwo One SlichgeReinf None
24 Good Average Low Average DiffTwo SameTwo None SlichgeReinf
25 Good Average Low Low One One SlichgeReinf SlichgeReinf
26 Excellent Low Low Low One One SlichgeReinf None
27 Excellent Average Low Average One None None SlichgeReinf
28 Excellent Low Average Low None One SlichgeReinf None
29 Excellent Low Low Low None One None SlichgeReinf
30 Excellent Low Average Low None One SlichgeReinf None
31 Excellent Low Low Average One None None SlichgeReinf
Note: SameTwo = Same slabs or walls on both sides; SlichgeReinf = slight changes in reinforcement; WDChange = width and depth are changed;
DiffTwo = different slabs or walls on each side; AIlChgeReinf = all reinforcements were changed; and AIlChange = both reinforcement and beam size
were changed.
TABLE 3. Decision Rules, Generalization Learning Mode

Decisions Conditions
CoBeRa1 CoBeRa2 NoSla NoWall
ConEva ReRa (column-to-beam (column-to-beam (number of slabs (number of walls
Rule (construction (reinforcement reinforcement reinforcement attached to attached to
number evaluation) ratio) ratio 1) ratio 2) beam) beam)
(1 ) (2) (3) (4) (5) (6) (7)
1 Poor High Average or high DiffTwo
2 Poor Average or high Average High
3 Poor High Average
4 Poor High Average
5 Good Average Low
6 Good Low SameTwo
7 Good Low Average or high Low None or SameTwo
8 Good High High SameTwo
9 Good Average Average Average
10 Good Low Average High
11 Excellent Low Low One
12 Excellent Low Average None
Note: SameTwo = Same slabs or walls on both sides; SlichgeReinf = slight changes in reinforcement; WDChange = width and depth are changed;
DiffTwo = different slabs or walls on each side; AIIChgeReinf = all reinforcements were changed; and AIIChange = both reinforcement and size were
changed.
The examples prepared for our study were balanced (Arci- minutes. The learning process produced three classes of de-
szewski and Mustafa 1989), which means they illustrated var- cision rules for the individual categories of the decision attri-
ious typical cases where the degree of difficulty of construc- bute constructability evaluation (ConEva). The decision rules
tion was known and the examples covered the entire spectrum were learned for examples using both modes of operation, de-
of possibilities in terms of the level of constructability. scribed earlier.
In total, 12 and 10 decision rules were produced, respec-
LEARNING RESULTS tively (see Tables 3 and 4). In Table 3, there are four, six, and
two rules created in the poor, good, and excellent class, re-
All examples were analyzed using the learning system IN- spectively. In Table 4, there are two, five, and three rules cre-
LEN based on the AQ15 algorithm described earlier in the ated in the three classes, respectively. The decision rules in
paper. The machine learning time was short, on the order of Table 4 include most of the independent attributes. However,
J. Comput. Civ. Eng., 1997, 11(1): 8-16

TABLE 4. Decision Rules, Specialization Learning Mode
Decision Conditions
CoBeRa1 CoBeRa2 BeCha1 BeCha2
(column-to- (column-to- NoSla NoWall (changes in (changes in
ConEva ReRa beam beam (number of (number of reinforcement reinforcement
Rule (construction (reinforcement reinforcement reinforcement slabs attached walls attached and size of and size of
number evaluation) ratios) ratio 1) ratio 2) to beam) to beam) left beam) right beam)
(1 ) (2) (3) (4) (5) (6) (7) (8) (9)
13 Poor Average or high Average or high Average or high SameTwo or Not none Not none
Difftwo
14 Poor Average High High Difftwo None SliChgeReinf WDChange
15 Good Average Low One or None or SliChgeReinf or
SameTwo SldichgeReinf AIlChgeReinf
or
AIlChgeReinf
16 Good Low or average Average or high Low or average One Not DiffTwo SlichgeReinf or None or Sli-
AIlChgeReinf ChgeReinf
17 Good Low or average Average or high High SameTwo or None None or Sli- SliChgeReinf
Difftwo ChgeReinf
18 Good Average Average Low None None or one SliChgeReinf or SliChgeReinf
AIlChgeReinf
19 Good Average Average Low SameTwo One SliChgeReinf None
20 Excellent Low Low or average Low None One None or Sli- None or Sli-
ChgeReinf ChgeReinf
21 Excellent Low Low Low or average One None or one None or Sli- None or Sli-
ChgeReinf ChgeReinf
22 Excellent Average Low Average One None None SliChgeReinf
the BeChal and BeCha2 attributes were not included in the All produced decision rules were found by practicing en-
Table 3 rules. Fewer atributes were used in the rules in the gineers as acceptable in the context of examples used to pro-
case when the generalization learning mode was used. duce them. Several of the rules obtained can be considered as
From Table 3, rule 11 can be interpreted as follows: constructability design principles, as general rules, which are
well known to designers familiar with the construction tech-
If the following conditions are met: nologies, equipment, and methods relevant to the erection of
the portions of structures outlined in our examples.
1. Reinforcement ratio of the beam is low.
2. The second column-to-beam reinforcement ratio is low. CONCLUSIONS
3. Number of walls attached to the slab is one.
The conducted research demonstrated the feasibility of ma-
Then constructability evaluation of a beam is expected to chine learning in constructability analysis. We have learned
be excellent. that it is possible to build a knowledge representation space
for a problem of constructability analysis of a structural design
However simplistic at first glance, this result is consistent concept and that this space can be subsequently used for au-
with common understanding by experienced builders of com- tomated constructability knowledge acquisition. The per-
plexities associated with concrete placement using either the formed learning experiments produced knowledge in the form
conventional crane and bucket method or the even more ver- of decision rules that are relatively simple and easy to inter-
satile concrete pumps, with the structural design characteristics pret. Their consistence with the present understanding of con-
as outlined in this rule. structability in the context of conceptual design decisions was
More complex rules were created using the specification surprising to us, particularly considering the limited and pre-
learning parameter. For example, rule 13 in Table 4 should be liminary nature of our study and the small number of examples
read as follows: used for learning.
The research produced also some initial methodological ex-
If the following conditions are met: perience in using machine learning for constructability knowl-
edge acquisition. In particular, we have learned how to co-
1. Reinforcement ratio of the beam is average or high. operate with construction project engineers in identifying
2. The first column-to-beam reinforcement ratio is average attributes and their values and how to prepare an appropriate
or high. collection of examples for automated knowledge acquisition.
3. The second column-to-beam reinforcement ratio is av- Also, we have learned about the utmost importance of the
erage or high. preparation of the knowledge representation space and of ex-
4. Number of attached slabs is "two different slabs." amples. These two stages in the knowledge acquisition process
5. Number of attached walls is "one wall," "same two are particularly difficult, both in conceptual and logistic terms,
walls," or "two different walls." and we initially underestimated the extent of work required to
6. There are changes in the reinforcement and size of beam complete them. The additional difficulty was the interdiscipli-
on the left. nary nature of the research, which required input from the
areas of structural and construction engineering and from com-
then the "constructability evaluation" of a beam is ex- puter science.
pected to be "poor." The research reported here was conducted using only a
small number of examples, and therefore a simple one-stage
The rule is fairly complex due to the assumed learning learning process was performed. In this process, all available
mode. examples were used for learning at the same time and no for-
J. Comput. Civ. Eng., 1997, 11(1): 8-16

mal verification of the acquired knowledge was conducted. struction automation and was focused on using machine learn-
The performed research resulted, however, in experience that ing tools for constructability analysis. This is an area where
allowed scaling up of the automated constructability knowl- progress will have a significant impact on the competitive ad-
edge acquisition process for the case in which a large number vantage in constructability knowledge acquisition capabilities
of examples is used for leaining. As a continuation of the work by individual construction companies. This may lead to better
reported here, a much larger collection of examples was used use of the largely untapped sources of constructability knowl-
at Purdue University (Lueprasert 1996) to acquire construct- edge resident in voluminous, uncomplied, and frequently pro-
ability constraints for the same problem of constructability prietary construction company records of projects performed
analysis of a beam in a reinforced-concrete frame. In this case, in the past.
a multistage learning process was conducted and its results
were monitored for the individual learning stages, and this was ACKNOWLEDGMENTS
done using several empirical error rates. In this way, knowl-
Financial support for this research at Purdue University has been pro-
edge being gradually acquired in a multistage learning process
vided by Obayashi Corporation. The first writer wishes to acknowledge

was immediately verified and the progress of learning deter- the contribution of Akiyoshi Sakaguchi, Manager of Obayashi Corpora-
mined. tion Systems Development and Computing Center, and his colleagues at
The continuation of research is planned. It will most likely the Obayashi Corporation Technical Research Institute for their assistance
include various learning experiments with several learning sys- in compiling construction project data for this research.
tems based on diferent learning algorithms, the use of con- The writers also wish to acknowledge the contributions of Dr. Janusz
Wnyk of the Laboratory for Machine Learning and Inference at George
structability examples provided by several industrial partners, Mason University (GMU) to the generation of constructability rules pre-
and the development of a methodology for using learning sys- sented herein. Part of this research was conducted in the Laboratory for
tems in constructability knowledge acquisition. Machine Learning and Inference at GMU and it has been in part indirectly
The interdisciplinary nature of this research means that its supported by that Laboratory.
continuation and extension may impact both construction and
structural engineering. In construction engineering, the impact APPENDIX. REFERENCES
envisioned will be to lower the cost of data collection and
Adeli, H., and Yeah, C. (1989). "Perceptron learning in engineering de-
compilation during the analysis of past construction projects. sign." J. Microcomp. in Civ. Engrg., 4, 247-256.
More importantly, it may allow for automated compilation of Arciszewski, T., and Dybala, T. (1992). "Evaluation of learning systems:
constructability knowledge without the total reliance on prej- a method and experimental results." Rep., Sys. Engrg. Dept., George
udiced reasoning of human experts with a system of traditional Mason Univ., Fairfax, Va.
notions of what is actually feasible on construction sites. Most Arciszewski, T., and Mustafa, M. (1989). "Inductive learning process:
the user's perspective." Machine learning, R. Forsyth, ed., Chapman
importantly, the constructability rules obtained using a learn-
and Hall, New York, N.Y., 39-61.
ing system could become new knowledge about the feasibility Arciszewski, T., and Rossman, L. (eds.) (1992). Knowledge acquisition
of unknown, yet unproven, innovative structural design con- in civil engineering, ASCE, New York, N.Y.
cepts being scrutinized for their applicability to specific con- Arciszewski, T., and Ziarko, W. (1992). "Machine learning in knowledge
struction projects. Thus, the continuation of this research could acquisition." Knowledge acquisition in civil engineering, T. Arci-
benefit the streamlining of communications between design szewski and L. Rossman, eds., ASCE, New York, N.Y.
Arciszewski, T., Mustafa, M., and Ziarko, W. (1987). "A methodology
experts and construction engineers and the future development of design knowledge acquisition for use in learning expert systems."
of innovative construction technologies. In structural engi- Int. J. Man-Machine Studies, 27, 23-32.
neering, learning constructability design rules should improve Arciszewski, T., Bloedorn, E., Michalski, R., Mustafa, M., and Wnyk, J.
the understanding of machine learning as applied to structural (1994). "Learning design rules for wind bracings in tall buildings." J.
design in the context of constructability analysis. That should Compu. in Civ. Engrg., ASCE, 8(2), 286-308.
produce methodological experience sufficient for the devel- Barrie, D. S., and Paulson, B. C. Jr. (1992). Professional construction
management, 2nd Ed., McGraw-Hill Book Co., Inc., New York, N.Y.
opment of a practical methodology for using learning systems Blaschke, B. C. (1989). "Specifications and constructability." Con-
in structural design for constructability. structed project excellence, Proc., Constr. Congr., I, ASCE, New York,
From the practical point of view, the automated knowledge N.Y., 97-92.
acquisition could remove the existing knowledge acquisition "Can your design be built?" (1987). Civ. Engrg., ASCE, 56(1), 49-51.
bottleneck in construction automation, and in constructability Clark, P., and Niblett, T. (1989). "The CN2 induction algorithm." J.
Machine Learning, 3, 261-284.
analysis in particular. Consequently, that might lead to the de- Cohenca, D., Laufer, A., and Ledbetter, W. B. (1989). "Factors affecting
velopment of a class of knowledge-based computer tools for construction planning efforts." J. Constr. Engrg. and Mgmt., 115(1),
constructability analysis which would be available to structural 70-89.
designers. These tools could be used in structural conceptual Cross, D. M. (1989). "Project constructability-a contractor's view."
design to evaluate individual design concepts in order to se- Constructed Project Excellence, Proc., Constr. Congr. I, ASCE, New
lect, for the final consideration only, those with the highest York, N.Y., 76-80.
Eldin, N. N. (1988). "Constructability improvement of project designs."
constructability. Considering the cost of structural design mis- J. Constr. Engrg. and Mgmt., ASCE, 114(4),631-640.
takes in the construction process, even a small reduction of Fisher, D. (1987). "Knowledge acquisition via incremental conceptual
such mistakes would have a significant impact on costs and clustering." J. Machine Learning, 2, 130-172.
project completion time of complex construction projects. Fischer, M. (1989). "A constructability expert system for the preliminary
The results of the study should be also considered in the design of reinforced concrete structures." Compo in Engrg. Pract.;
Proc., ASCE 5th Conf on Compo in Civ. Engrg., ASCE, New York,
context of automated construction systems and of robotics in N.Y., 60-66.
particular. In this case, a structural design must be prepared Fisher, D. J., and O'Connor, J. T. (1991). "Constructability for piping
satisfying all constructability constraints related to the use of automation: field operations." J. Constr. Engrg. and Mgmt., ASCE,
construction robots. Ail such constraints must be known in the 117(3), 468-485.
explicit form before the design process starts. In this case, the Gero, J. S., Mackenzie, C. A., and McLaughlin, S. (1992). "Learning
constructability analysis is extremely important, because the from optimal solutions to design problems." Optimization and artificial
intelligence in civil and structural engineering, Vol. II, B. H. v., ed.,
automatic construction systems allow only very limited, if any, Kluwer Academic Press, Boston, Mass., 217-250.
changes in the construction process due to structural design Grierson, D. E., and Pak, W. H. (1993). "Optimal sizing, geometrical
mistakes. and topological design using a genetic algorithm." J. Struct. Optimi-
The research reported in this paper is in a new area of con- zation, 6, 151-159.
J. Comput. Civ. Eng., 1997, 11(1): 8-16

Hajela, P. (1989). "Genetic algorithms in automated structural synthe- 2nd Int. Conf. on Artificial Intelligence in Design 1992, I. S. Gero, ed.,
sis." Proc., NATO ASI Conf. on Optimization and Decision Support K1uwer Academic Press, Dordrecht, The Netherlands.
Sys. in Civ. Engrg. Modesitt, K. (1992). "Basic principles and techniques in knowledge ac-
Hendrickson, C. T., and Maher, M. L. (1989). "Issues in computer based quisition." Knowledge acquisition in civil engineering, T. Arciszewski
design/construction integration." Constructed Proj. Excellence; Proc., and L. Rossman, OOs., ASCE, New York, N.Y.
ASCE Constr. Congo I, ASCE, New York, N.Y., 129-136. Morad, A. A., and Cleveland, A. B. Ir. (1989). "Path optimization al-
Hirabayashi, Y. (1993). "Analysis of the development and application of gorithm for constructability improvement." J. Compo in Civ. Engrg.,
automation technology in building production." Proc., 10th Int. Symp. ASCE,498-506.
on Automation and Robotics in Constr., Elsevier Science Publishers Muggleton, S. (1987). "Duce, an oracle-based approach to constructive
BV, Amsterdam, The Netherlands. induction." Proc. IJCAI-87, 287-292.
Hohns, M. (1985). "Learning from failures: procedural changes in the Murlidharan, T. L., Aravind, H. B., Suryakumar, G. V., and Raman, N.
design and construction process to reduce failures." Reducing Failures N. (1991). "Expert tower analysis and design system 11: search strat-
of Engineered Fac.; Proc., NSFIASCE Workshop, ASCE, New York, egies and learning." J. Compo in Civ. Engrg., ASCE, 5(2), 193-210.
N.Y., 75-83. Murtaza, M., Fisher, D., and Skibniewski, M. J. (1993). "Issues in mod-
Hong, I., Mozetic, I., and Michalski, R. S. (1986). "AQI5: incremental ular construction decision support." J. Constr. Engrg. and Mgmt.,
learning of attribute-based descriptions from examples, the method and ASCE, 119(1), 115-130.
user's guide." Rep. Isa 86-5, UlUCDCS-F-86-949, Compo Sci. Dept., Mustafa, M., and Arciszewski, T. (1992). "Inductive learning of wind
Univ. of Illinois at Urbana Champaign, III. bracing design for tall buildings." Knowledge acquisition in civil en-
"Integrating construction resources and technology into engineering." gineering, T. Arciszewski and L. Rossman, eds., ASCE, New York,
(1982). Rep. B-1, The Business Roundtable, New York, N.Y. N.Y.
Ivezic, N., and Garret Ir., I. H. (1993). "NETSYN: neural network-based O'Connor, I. T., and Tucker, R. L. (1986). "Industrial project construct-
support for synthesis." Proc., 5th Int. Con/. on Compo in Civ. Engrg., ability improvement." J. Constr. Engrg. and Mgmt., ASCE, 112(1),
L. Cohen, ed., ASCE, New York, N.Y., 1440-1448. 69-82.
Iergeas, G. E, Tyler, A. H., and McCaffer, R. (1991). "Constructability O'Conner, J. T., Larimore, M. A., and Tucker, R. L. (1986a). "Collecting
and detailed design." Proc., 4th Int. Conf. on Civ. and Struct. Engrg. contructability improvement ideas." J. Contr. Engrg. and Mgmt.,
Comp., Elsevier, London, England, 23-27. ASCE, 112(4), 463-475.
Kirby, I. G., and Cannalte, R. P. (1988). "Automated design review man- O'Connor, J. T., Rusch, S. E., and Schultz, M. J. (1986b). "Constructa-
agement." Camp. in Civ. Engrg.: Microcomputers to Supercomputers; bility improvement during engineering and procurement." Source Doc-
Proc., ASCE 5th Conf., ASCE, New York, N.Y., 769-778. ument 5, Constr. Industry Inst., Austin, Tex.
Kirby, I. G., Cannalte, R. P., Hicks, D. K., and lapel, E. I. (1989). "Con- Patty, R. (1993). "Construction engineering platform for integration of
structability and design reviews: analysis and recommendations for im- constructability concepts and lessons learned at the point of design,"
provement." Tech. Rep. P-89115, U.S. Army Corps of Engrs., Wash- PhD thesis, School of Civ. Engrg., Purdue Univ., W. Lafayette, Ind.
ington, D.C. Pawlak, Z. (1982). "Rough sets." Int. J. Camp. and Information Sci.,
Kirby, I. G., Tupas, M. I., Robinson, P. C., and Bridgestock, G. (1991). 5(11), 341-356.
"Concept development of an automated construction design review Quality in the constructed project: guidelines for owners, designers and
advisor." Interim Rep. P-911 13, U.S. Army Corps of Engrs., Wash- constructors. (1988). ASCE, New York, N.Y., 1.
ington, D.C. Quinlan, I. R. (1986). "Induction of decision trees." J. Machine learning,
Laird, I. E., Rosenbloom, P. S., and Newell, A. (1986). "Chunking in 1(1),81-106.
SOAR: the anatomy of a general learning mechanism." J. Machine Reich, Y. (1992). "Generation of examples for training a learning design
Learning, I, 11-46. system." Proc., ASCE Can! for Camp. in Civ. Engrg., ASCE, New
Lu, S. C.-Y., and Chen, K. (1987). "A machine learning approach to the York, N.Y.
automatic synthesis of mechanistic knowledge for engineering decision Reich, Y., and Fenves, S. I. (1988). "Floor-system design in SOAR: a
making." Artificial Intelligence in Designing and Manufacturing, 1(2), case study of learning to learn." Rep., Engrg. Des. Res. Ctr., Carnegie
109-118. Mellon Univ., Pittsburgh, Pa.
Lueprasert, K. (1996). "Constructability knowledge acquisition: a ma- Reich, Y., and Fenves, S. I. (1989a). "The potential of machine learning
chine learning approach," PhD dissertation, School of Civ. Engrg., techniques in civil engineering." Artificial Intelligence in Engrg. Des.
Purdue Univ., West Lafayette, Ind. and Manufacturing, 3(3).
Lutz, I. D., Hancher, D. E., and East, E. W. (1990). "Framework for Reich, Y., and Fenves, S. J. (1989b). "Inductive learning of synthesis
design quality review database system." J. Mgmt. in Engrg., ASCE, knowledge." Int. J. Expert Sys.: Res. and Applications, 5(4), 275-297.
6(3),296-312. Reich, Y., and Fenves, S. I. (1992). "Automated acquisition of design
Mackenzie, C. A., and Gero, I. (1987). "Learning design rules from knowledge by concept formation." Knowledge acquisition in civil en-
decisions and performances." Artificial Intelligence in Engrg., 2(1), gineering, T. Arciszewski, and L. Rossman, eds., Monograph on
2-10. Knowledge Acquisition in Civ. Engrg., ASCE, New York, N.Y.
Maher, M. L. (1992). "Automated knowledge acquisition of preliminary Russell, J. S., Radtke, M. w., and Gugel, J. G. (1992a). "Project-level
design concepts." Proc., ASCE Conf. for Compo in Civ. Engrg., ASCE, model and approaches to implement constructability." Source Docu-
New York, N.Y. ment 82, Constr. Industry Inst., Austin, Tex.
Maher, M. L., and Li, H. (1992). "Automatically learning preliminary Russell, J. S., Gugel, I. G., and Radtke, M. W. (1992b). "Benefits and
design knowledge from design examples." J. Microcomputers in Civ. costs of constructability: four case studies." Rep. 109, Dept. of Civ.
Engrg., 7(1), 73-80. and Envir. Engrg., Univ. of Wisconsin-Madison, Wis.
Maher, M. L., and Kundu, S. (1993). "Adaptive design using a genetic Skibniewski, M. J., and Chao, L.-C. (1992). "Evaluation of advanced
algorithm." Preprints formal design methods for CAD, I. S. Gero and construction technologies with analytical hierarchy process." J. Constr.
E Sudweeks, eds., Key Ctr. of Des. Comp., Univ. of Sydney, Australia, Engrg. and Mgmt., ASCE, 118(3),577-593.
211-228. Tatum, C. B. (1987). "Improving constructability during conceptual plan-
Matheus, C. (1989). "Feature construction: an analytic framework and ning." J. Constr. Engrg. and Mgmt., ASCE, 113(2), 191-207.
application to decision trees," PhD dissertation, Univ. of Illinois at Tatum, C. B. (1989a). "Management challenges of integrating construc-
Urbana-Champaign, III. tion methods and design approaches." l- Mgmt. in Engrg., 5(2), 139-
McLaughlin, S., and Gero, I. S. (1987). "Acquiring expert knowledge 154.
from characterized designs." Artificial Intelligence in Engrg. Designing Tatum, C. B. (1989b). "Managing for increased design and construction
and Manufacturing, 1(2), 73-87. innovation." J. Mgmt. in Engrg., 5(4), 385-399.
"Measuring the cost of quality in design and construction.• , (1989). Publ. Tatum, C. B. (1990). "Integrating design and construction to improve
10-2, Constr. Industry Inst., Austin, Tex. project performance." Proj. Mgmt. J., 21(2), 35-42.
Michalski, R. S. (1983). "A theory and methodology of inductive learn- Tatum, C. B., Vanegas, I. A., and Williams, J. M. (1985). "Constructa-
ing." Machine learning: an artificial intelligence approach, R. S. bility improvement during conceptual planning." Tech. Rep. 290, Dept.
Michalski, I. G. Carbonell, and T. Mitchell, eds., Morgan Kaufmann, of Civ. Engrg., Stanford Univ., Stanford, Calif.
Los Altos, Calif., 1. Tatum, C. B., Vanegas, J. A., and Williams, J. M. (1986). "Constructa-
Michalski, R. S., Mozetic, I., Hong, I., and Lavrac, N. (1986). "The bility improvement using prefabrication, preassembly and modulari-
multi-purpose incremental learning system AQ15 and its testing appli- zation." Tech. Rep. 297, Dept. of Civ. Engrg., Stanford Univ., Stanford,
cation to three medical domains." Proc., AAAI-86, 1041-1045. Calif.
Milzner, K., and Harbecke, A. (1992). "Incremental learning for im- Tatum, C. et aI. (1987). "Constructability concepts file." Publ. 3-3,
proved decision support in knowledge based design systems." Proc., Constr. Industry Inst., Austin, Tex.
J. Comput. Civ. Eng., 1997, 11(1): 8-16

Thcker, R. L., Peterson, C., Meyer, J., and Simonson, T. (1990). "A based learning: using domain theory to guide induction." Annu. Rep.,
methodology for identifying automation opportunities in industrial con- Knowledge-Based Engrg. Sys. Res. Lab., Univ. of Illinois at Urbana-
struction." Source Document 56, Constr. Industry Inst., Austin, Tex. Champaign, Urbana, Ill.
Viscomi, B. V., Michalerya, W. D., and Lu, L.-W. (1993). "Automated Ziarko, W. (1989). "ROUGH data analysis and case-based expert system
construction in the ATLSS integrated building systems." Proc., 10th development tool." Proc., Case-Based Reasoning Workshop, Morgan
Int. Symp. on Automation and Robotics in Constr., Elsevier Science Kaufmann, Pensacola, Fla.
Publishers, Amsterdam, The Netherlands. Zweig, A. (1984). "Combining design and construction: spark to design
Whitehall, B. L., Stepp, R. E., and Lu, S. C.-Y. (1990). "Knowledge- excellence?" Civ. Engrg., ASCE, 54(2), 6.
J. Comput. Civ. Eng., 1997, 11(1): 8-16

Constructability Analysis - Machine Learning Approach

Uploaded by

Constructability Analysis - Machine Learning Approach

Uploaded by

CONSTRUCTABILITY ANALYSIS: MACHINE LEARNING ApPROACH

By Mirostaw Skibniewski, l Tomasz Arciszewski,2 and Kamolwan Lueprasere

INTRODUCTION result is consistent with the observed knowledge acquisition

J. Comput. Civ. Eng., 1997, 11(1): 8-16

J. Comput. Civ. Eng., 1997, 11(1): 8-16

J. Comput. Civ. Eng., 1997, 11(1): 8-16

eAllChgeReinf = All reinforcements were changed.

FEASIBILITY STUDY 1. ReRa represents reinforcement ratio of the beam, and

J. Comput. Civ. Eng., 1997, 11(1): 8-16

12 Good Average Average Low None One SlichgeReinf SlichgeReinf

TABLE 3. Decision Rules, Generalization Learning Mode

J. Comput. Civ. Eng., 1997, 11(1): 8-16

J. Comput. Civ. Eng., 1997, 11(1): 8-16

vided by Obayashi Corporation. The first writer wishes to acknowledge

14/ JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 1997

J. Comput. Civ. Eng., 1997, 11(1): 8-16

JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 1997/15

J. Comput. Civ. Eng., 1997, 11(1): 8-16

16/ JOURNAL OF COMPUTING IN CIVIL ENGINEERING / JANUARY 1997

J. Comput. Civ. Eng., 1997, 11(1): 8-16

You might also like