This is the Pre-Published Version.

Journal of Structural Engineering, ASCE, Vol. 129, No. 10, 2003, pp. 1312-1321

Knowledge-Based System on Optimum Design of Liquid Retaining Structures with Genetic

Algorithms

K.W. Chau1 & F. Albermani2

1
Department of Civil & Structural Engineering, Hong Kong Polytechnic University,
Hunghom, Kowloon, Hong Kong
2
Department of Civil Engineering, University of Queensland, QLD 4072, Australia

Abstract
This paper delineates the development of a prototype hybrid knowledge-based system for the
optimum design of liquid retaining structures, OPTLIQ, by coupling the blackboard
architecture, an expert system shell VISUAL RULE STUDIO and genetic algorithm (GA).
Through custom-built interactive graphical user interfaces under a user-friendly environment,
the user is directed throughout the design process, which includes preliminary design, load
specification, model generation, finite element analysis, code compliance checking and
member sizing optimization. For structural optimization, GA is applied to the minimum cost
design of structural systems with discrete reinforced concrete sections. The design of a
typical example of the liquid retaining structure is illustrated. The results demonstrate
extraordinarily converging speed as near-optimal solutions are acquired after merely
exploration of a small portion of the search space. This system can act as a consultant to
assist novice designers in the design of liquid retaining structures.

Keywords: genetic algorithm, knowledge-based system, liquid retaining structures, structural

optimization

Introduction
The design of liquid retaining structures is specialized and requires assimilation of knowledge
from heuristics, research findings and standard engineering methodology. Since these types
of structures are exposed in corrosive environment, crack width control is, among others,
very crucial in the design. As deviations often exist between the assumed properties of
components at the preliminary design stage and their counterparts determined at the detailed
design stage, re-analysis will be entailed and iterative steps such as configuration processing,
numerical modeling, structural analysis, code conformance checking and sizing optimization
are usually involved. It was not easy to code those empirical rules or expert knowledge in a
conventional algorithmic or sequential framework. Previous computer-aided design was
mainly through loose coupling of individual programs on the captioned sub-processes, which
entail intensive knowledge of the structural designer and are prone to human errors during the
data transferring processes. As such, an integrated as well as user-friendly system is valuable
for storing and employing the expertise.

Recent advances in knowledge-based system (KBS) technologies have rendered it possible to

incorporate the heuristic knowledge into the conventional algorithmic structural analysis
models. A KBS can be defined as an interactive computer-based decision-making tool that
emulates the intensive expert knowledge in a specific domain problem. During the last
decade, KBSs have been widely adopted to solve problems in many different disciplines
(Chau 1992, Chau et al. 2002, Chau and Ng 1996, Chau and Yang 1992, Chau and Zhang
1995). For application in structural design, the KBS framework is in addition required to

1
couple symbolic processing and extensive numerical processing. Examples of such systems
by employing various representation schemes are INDEX (Kumar 1995) and LADOME (Lin
and Albermani 2001). However, none of these KBSs appear to have incorporated the recent
structural optimization techniques.

In the past, a myriad of mathematical programming algorithms, which usually assume

continuous design variables and simple constraints, have been developed for the optimum
design of structural systems. Some popular methods, such as calculus-based gradient
techniques, entail the construction or approximation of derivative information and yet may
only attain local optima. In practical structural design problems, owing to the availability of
standard practical sizes and their restrictions for construction and manufacturing purposes,
the design variables are always discrete. In our case for instance, in order to optimize
reinforced concrete structural systems, design variables, which are concrete cross-sectional
area, reinforcement diameter and reinforcement spacing, should be considered as discrete
quantities in practice. However, only very few algorithms have dealt with the optimization of
structures under the actual design constraints of code specifications. In fact, it is more rational
to use discrete variables during the evaluation process of optimization since every candidate
design is a practically feasible solution. This may not be the case when the design variables
are continuous, since some of the designs evaluated in the optimization procedures are merely
mathematically feasible. This issue can have great significance in solving practical problems
of design optimization.

In recent years, genetic algorithms (GAs), which are applications of biological principles into
computational algorithm, have been employed to attain the optimum design solutions
(Goldberg and Kuo 1987). By applying the principle of survival of the fittest into the
optimization of structures, they are particularly amenable to deal with discrete optimum
design problems. Besides, they entail only minimum subsidiary information, i.e. on objective
function value, to direct the search, yet they are able to search through large spaces quickly.
Owing to the processing leverage associated with GAs, the method has a much more global
perspective than many commonly employed optimization techniques. However, as far as
structural optimization is concerned, literature review shows that steel structures are usually
dealt with (Hayalioglu 2001).

In this study, a genetic algorithm (GA) is incorporated into the KBS on the optimum design
of reinforced concrete structures subjected to the actual constraints of the British Standard on
design of concrete structures for retaining aqueous liquid, BS 8007, employing actual
reinforced concrete sections as discrete design variables. When compared with optimum
design of steel structures, it should be noted that reinforced concrete structures involve more
design variables as it deals with more than a single type of material, namely, concrete and
steel reinforcement. Besides, the objective function in the optimization of steel structures can
be the minimum weight or cost which are representing the same situation whilst their
counterparts of reinforced concrete structures have to be minimum material cost. It is because
more than one material are involved and the unit cost of concrete is normally different from
that of reinforcement.

In civil engineering, applications of machine learning are still rare. GAs are considered a
means of machine learning and it is worthwhile to study its capability to implement
knowledge acquisition under a KBS paradigm. Whilst there are numerous applications of
KBS or genetic algorithms, a hybrid KBS with GAs has not been found. As such, the
objective of this study is to develop a microcomputer-based hybrid KBS that can bring all the

2
design stages, comprising structural optimization with GAs in particular, together into a
single and user-friendly environment. It can offer assistance and advice to the design engineer
in making decisions. Increase in efficiency, improvement, standardization and optimization
of design output and automated record keeping are among the benefits of this hybrid KBS.

Programming Development Environment

The KBS development environment for OPTLIQ is VISUAL RULE STUDIO, which acts as
an ActiveX Designer under the Microsoft Visual Basic programming environment. VISUAL
RULE STUDIO is an application development environment that combines expert system
technologies with object-oriented programming, relational database, graphics capabilities and
debugging tools. It furnishes a variety of knowledge representation schemes, different
inference mechanisms and capabilities to interface with external programs in windows
environment. VISUAL RULE STUDIO provides an interactive windows-based user interface
that runs under the conventions of Microsoft Windows.

Architecture of Hybrid System

On the basis of the nature of structural design, the selected knowledge base shell as well as its
capability in coupling all design stages for liquid retaining structures, blackboard architecture
is adopted here (Engelmore and Morgan 1988). The well-organized architecture can facilitate
the communication between diversified knowledge modules involved in the structural design
process as well as future extension. Figure 1 shows the architecture of the KBS and the
relation between various components of OPTLIQ. The knowledge base is mainly composed
of knowledge modules and the blackboard. Knowledge modules correspond to procedural
expertise knowledge in solving design problem and are divided into two groups, namely,
Design Process and Process Control. The design process can be subdivided into subtasks,
each of which can be performed corresponding to a particular expertise within the knowledge
base. Message communications between Design Process knowledge modules and Process
Control knowledge modules have to effect via the blackboard.

The blackboard contains only declarative knowledge and is divided into two groups, namely,
Design Entities and Design Stage. The Design Entities can be considered the breakdown of
design concepts of liquid retaining structures. A typical example of objects in the blackboard
is Wind Load. The properties of Wind Load are expressed by using several attributes, which
is detailed in Figure 2. In this example, the attribute type of terrainCategory is compound,
which means that it can take only one of the specified values “general terrain” or “builtup
terrain”. Its search order is session context only, meaning that, in determining its value, the
system context is searched. The attribute type of basicWindPressure is numeric, which means
that it must have a numerical value. Its search order list is rules only. The class in Design
Stage includes a number of attributes that represent indicators tracking the current stage of
every design context. The knowledge represented in this level will handle the order of
execution via the Process Control knowledge modules.

Knowledge Acquisition and Representation

Knowledge acquisition is mainly from literature, including handbooks, standards and codes,
and interviews with experienced designers on liquid retaining structures. The domain
knowledge of OPTLIQ, such as the minimum percentage of reinforcement, maximum
percentage of reinforcement, minimum grade of concrete, etc. is obtained, which is then
translated into rules or methods. For instance, the following rule group is expressed using the
Production Rule Language, representing knowledge on the determination of bending
moments and shear forces in preliminary design.

3
!RULE GROUP: moment & shear OF BBPreliminaryParameters

RULE to find momentMh : 1 of 4

IF shape OF BBConfigurationRequirement IS rectangular
THEN momentMh OF BBPreliminaryParameters := momentCoefficientMh OF
BBPreliminaryParameters*specificGravityOfLiquid OF
BBConfigurationRequirement*height OF BBConfigurationRequirement^3/1000

RULE to find momentMv : 2 of 4

IF shape OF BBConfigurationRequirement IS rectangular
THEN momentMv OF BBPreliminaryParameters := momentCoefficientMv OF
BBPreliminaryParameters*specificGravityOfLiquid OF
BBConfigurationRequirement*height OF BBConfigurationRequirement^3/1000

RULE to find shearVh : 3 of 4

IF shape OF BBConfigurationRequirement IS rectangular
THEN shearVh OF BBPreliminaryParameters := shearCoefficientVh OF
BBPreliminaryParameters*specificGravityOfLiquid OF
BBConfigurationRequirement*height OF BBConfigurationRequirement^2/100

RULE to find shearVv : 4 of 4

IF shape OF BBConfigurationRequirement IS rectangular
THEN shearVv OF BBPreliminaryParameters := shearCoefficientVv OF
BBPreliminaryParameters*specificGravityOfLiquid OF
BBConfigurationRequirement*height OF BBConfigurationRequirement^2/100

Interfacing Facilities
Microsoft Visual Basic offers the interfacing facilities where the execution of external
program can be achieved by using a command “SHELL external program name”. All data
communications are effected in a fully integrated fashion. Once the execution of the external
algorithmic program is done, the KBS resumes its design session environment. Following
preparation of the input file in model generation, the nonlinear finite element analysis
package ABAQUS is executed. The output data files are saved in the knowledge base for
retrieval and manipulation. The KBS is not only aimed to act as a front-end to this finite
element package, but also to encapsulate knowledge on the entire design process. In order to
achieve the optimum design, structural re-analysis is often inevitable.

Whilst existing algorithmic models generally deal with numerical data input only, the novice
user usually finds it more convenient to express information in a natural language. The
numerical model generator functions to convert these linguistic variables entered by the user
to numerical format conforming to stipulations of the analysis package. The code
conformance checking module is used to check the code requirements of BS 8007 (British
Standards Institution 1987). The communication between the programs and the knowledge
base is performed mainly through the objects and attributes. When a consultation pertinent to
extensive data is made to the system, external Access database files are accessed. The
database is composed of structural properties of reinforced concrete sections, moment and
shear coefficients for various configurations in preliminary design, structural properties of
proposed alternatives and final member details in detailed design.

4
Explanation Facilities
Explanations of liquid retaining structure types, procedures to design various design loading,
various code provisions and expert comments regarding design of liquid retaining structures
are included in the explanation facilities via the Help button. The explanations consist of
built-in specific texts together with associated values of design parameters generated by the
knowledge base during system run time.

User Interface Facilities

The user interface facilities allow the user to specify all design requirements and acquire
output result from interactive design consultation. Graphical user interfaces, consisting of
layers of display screens and pop-up windows are used for messages transfer and hence input,
handling and interpretation of data and information have been greatly simplified. During the
design process, subjected to conformance with Process Control knowledge modules, the user
has the control over the sequence of the actions. Communications with the system are
directed mostly through selection of appropriate values or parameters from menu and through
replying answers to the queries asked by the system.

Major Design Tasks

Major design tasks performed by the KBS are preliminary synthesis, detailed specification,
numerical model generation, nonlinear structural analysis, code conformance checking, and
member sizing optimization. The system applies some engineering heuristics such as
approximate analysis using span depth ratio, crack width computation and moment and shear
coefficients for different length and width ratios to evaluate each alternative. Through the GA
optimization module, the alternative with the minimum cost will be recommended by the
system as the selected proposed alternative.

During the detailed specification stage, the system generates default loading and support
conditions, which can be modified by the user. If the liquid retaining structure is chosen to be
underground, the user is required to enter level of ground surface, level of water table,
specific weight of soil and active soil pressure coefficient. If the structure is above the ground,
the wind loading is calculated according to the Code of Practice for Wind Effects in Hong
Kong (Hong Kong Building Development Department 1983). Various load combinations
according to BS 8110 (British Standards Institution 1985) are considered.

An iterative process of numerical model generation, finite element analysis, code

conformance checking and optimization of structure are then involved. Upon receiving the
requisite messages from the knowledge base, a structural model tailor-made for ABAQUS is
automatically prepared by the model generator, thus relieving the user of the cumbersome
task to manipulate a large amount of data manually. The system evaluates the structural
stability from the analysis results and furnishes post-processing. The system then proceeds to
check the structural members according to BS 8007. The computed crack width has to be less
than the prescribed crack width.

Structural optimization of the structure involves sizing the components under the constraints
of the structural adequacy as well as the crack width requirements, which is effected by GAs
as depicted in the following sections. The optimum section is then compared with the original
section used in the preceding structural analysis. This phase of design typically involves
several iterations until convergence is accomplished. Figure 3 is the flowchart showing the
overall design algorithm of OPTLIQ.

5
Optimization by GAs
Holland (1975) has put forward the GAs as an optimization method, which apply the concept
on the artificial survival of the fittest coupled with a structured information exchange using
randomized genetic operators taken from the nature. GAs differ from traditional optimization
algorithms in aspects such as working on coded design variables, population processing,
probabilistic operators, and separation of domain knowledge from search. GAs exploit
efficiently useful information subsumed in a population of solutions with better performance,
by employing operations to generate a new and improved population of strings from an old
population. This iterative process to generate and test a population of strings mimics a natural
population of biological creatures. GAs search from a population of strings and climb many
peaks in parallel simultaneously, hence lowering the probability of locating local optima.
GAs require that alternative solutions be coded as strings, which may comprise concatenation
of some substrings so that each substring represents a design variable. Individuals and the
characters are termed chromosomes and artificial genes, respectively.

The reproduction operator applies the principle of survival of the fittest in the population.
Strings with better objective function values, representing more highly fit, receive more
offspring in the mating pool. The crossover operator leads to the recombination of individual
genetic information from the mating pool and the generation of new solutions to the problem.
In the present work, a two-point crossover is utilized. A mutation operator is applied so as to
avoid being trapped in local optima and to preserve the diversification among the population
in the search. This operator is applied to each offspring in the population with a
predetermined probability, termed the mutation probability.

Formulation of Optimum Design Problem

The set of design variables, including the thickness of concrete slab, the diameter of
reinforcement bar and the bar spacing, is determined so that the total material cost of the
structure comprising n groups of member,

n
min C ( x) = ∑ U i *Vi + Ri *Wi (1)
i

is minimized subject to crack width and stress constraints. In eq. (1), Ui and Vi represent the
unit cost and the concrete volume of member i respectively. Ri and Wi are the unit cost and
the weight of steel reinforcement of member i respectively.
The serviceability limit state or crack width constraint is

W a −W max ≤ 0 (2)

where Wa is the actual crack width and Wmax is the prescribed maximum crack width
depending on the exposure environment. Wa, is determined using the following formula:-

3a cr ε m
Wa = (3)
a −c
1 + 2( cr )
h−x

where acr is the distance from the point considered to the surface of the nearest longitudinal
bar, εm is the average strain for calculation of crack width allowing for concrete stiffening

6
effect, c is the minimum cover to the tension reinforcement, h is the overall depth of the
member and x is the depth of the neutral axis (British Standards Institution 1987).

The stress constraints, representing the ultimate limit states of flexure and shear resistance,
are expressed in terms of the following equations (British Standards Institution 1985) for
members subject to bending and shear force:

M au − M ult ≤ 0 (4)
V a − Vult ≤ 0 (5)

where Mau is the actual ultimate bending moment, Mult is the nominal ultimate moment
capacity of the reinforced concrete section, Va is the actual ultimate shear force and Vult is the
nominal ultimate shear capacity of the section. The ultimate moment capacity is determined
by the following equations, depending on whether concrete or steel stresses is more critical.

Fy
M ult = As Z or M ult = 0.157 Fcu bd 2 whichever is the lesser (6)
1.15

where Fy is the yield strength of reinforcement, As is area of tension steel, Z is the lever arm,
Fcu is the characteristic concrete strength, b is the width of section and d is the effective depth
of section.

Ultimate shear capacity of the section (Vult = vcbvd) is represented by shear strengths vc for
sections without shear reinforcement, which depend upon the percentage of longitudinal
tension reinforcement [100As/(bvd)] and the concrete grade:-

v c = 0.79[100 As /(bv d )]1 / 3 (400 / d ) 1 / 4 / γ m (7)

where bv is breadth of section, γm is a safety factor equal to 1.25, with limitations that
[100As/(bvd)] should not be greater than three and that (400/d) should not be less than one.
For characteristic concrete strengths greater than 25 N/mm2, the values given by the above
expression is multiplied by (Fcu/25)1/3.

The above is then converted into an unconstrained problem by employing a transformation

based on the violations of normalized constraints (Hayalioglu 2001). The normalized form of
constraints can be expressed as follows:

Wa
−1 ≤ 0 i=1,…..n (8a)
Wmax
M au
−1 ≤ 0 i=1,…..n (8b)
M ult
Va
−1 ≤ 0 i=1,…..n (8c)
Vult

The unconstrained objective function ϕ(x) is written as

n

∑ {(W
Wa M au V
ϕ ( x) = C ( x)[1 + K − 1) + + ( − 1) + + ( a − 1) + }] (9)
i max M ult Vult

7
where K is a penalty constant to be selected depending on the problem and

Wa W
( − 1) + = max( a − 1,0) (10a)
W max W max
M au M
( − 1) + = max( au − 1,0) (10b)
M ult M ult
Va V
( − 1) + = max( a − 1,0) (10c)
Vult Vult

The penalty parameter largely depends upon the degree of constraint violation. In this case,
values of 10 and 100 have been attempted and it is found that the results are not sensitive to
these two values.

The minimum of the unconstrained function ϕ(x) is searched by the GA, with the best
individual having the maximum fitness. In this study, the fitness value is acquired by
subtracting from the summation of the maximum and minimum values of the objective
function. This ensures that all the fitness values are non-negative and individuals acquire
fitness values in accordance with their actual merit. The expression becomes

F j = [ϕ ( x) max + ϕ ( x) min ] − ϕ j ( x) (11)

where Fj is the fitness of the j-th individual in the population, ϕ(x)max and ϕ(x)min are the
maximum and minimum values of ϕ(x) among the current population respectively and ϕj(x)
is the objective function value computed for the j-th individual. Computation of the fitness of
an individual requires the values of crack width and stresses in the structural system from the
results of the finite element analysis.

Optimum Design Algorithm

Figure 4 shows the flowchart of GA for optimum design of reinforced concrete sections in
liquid retaining structures. It is programmed under Microsoft Visual Basic programming
environment. The starting population is first randomly constructed. The binary codes for the
design variables of each individual are decoded and their sequence numbers in the available
slab thickness, bar diameter and bar spacing are found. Based on the responses of the
structure, the value of unconstrained function ϕ(x) for each individual is computed. The
maximum and minimum values of this function in the population are determined and hence
the fitness value for each individual is determined.

By applying the reproduction operator, the individuals are copied into the mating pool
according to their fitness. A proportionately higher probability of reproduction selection, sj, is
given to those strings with higher fitness values Fj according to the following distribution
Fj
sj = Psize
(12)
∑Fj
j

where Psize is the population size. As the number of individuals in the next generation is also
the same, the individuals with small fitness die off.

After the mating pool is created, individuals are coupled to generate offspring using a two-
site crossover. A set of crossover parameters, consisting of a match and two cross sites, are
generated randomly. The genetic operation of crossover is performed on each mated pair with

8
a certain probability, referred to as crossover probability. Suppose that two strings X and Y of
length 11 are the mating pair with the following genes

X = x1 , x 2 , x 3 , x 4 , x 5 , x 6 , x 7 , x 8 , x 9 , x10 , x11 (13a)

Y = y1 , y 2 , y 3 , y 4 , y 5 , y 6 , y 7 , y 8 , y 9 , y10 , y11 (13b)

Two cross sites cs1 and cs2 are randomly selected and two new strings are created by
swapping all characters between positions cs1 and cs2 inclusively from one individual in the
pair to the other. For instance, if the cross sites generated are 2 and 7, the resulting crossover
yields two new strings X’ and Y’ following the partial exchange.

X ' = x1 , x 2 , | y 3 , y 4 , y 5 , y 6 , y 7 , | x 8 , x 9 , x10 , x11 (14a)

Y ' = y1 , y 2 , | x 3 , x 4 , x 5 , x 6 , x 7 , | y 8 , y 9 , y10 , y11 (14b)

Mutation is then applied to each offspring in the new population, which flips the gene of an
offspring from 1 to 0 or vice versa at random position. The initial population is then replaced
by the new population.

The above steps are then iterated until the distance between the maximum and the average
fitness values of the current population is less than a certain threshold. The optimum values
obtained from interactive optimization can then be added to the knowledge base, which is
effectively extended through machine experimentation. Machine learning can be effected
since the final optimum structural section under the specified loading and geometrical
configuration from the finite element structural analysis is added to the database containing
the heuristics during the preliminary design. The system can also be used as a means for
testing the available empirical knowledge from the model run.

Application Case
A typical example of liquid retaining structure is used to illustrate the application of OPTLIQ.
A rectangular shape liquid retaining structure with two compartments located above the
ground, having a volume of 100 m3, a depth of 5 m and breadth/width ratio of 1.2, is designed
under a severe exposure environment, i.e. the maximum design crack width is 0.2mm.

Preliminary Design
Upon execution of the system, the main menu screen is displayed as shown in Figure 5.
During the preliminary design stage, heuristics are used to evaluate different alternatives. The
system searches the databases on moment and shear coefficients and on sectional properties
and suggests an initial member thickness of 225 mm with reinforcement diameter 10 mm at
100 mm spacing as the most suitable alternative. The user can choose between system’s
selection and user’s selection and, in this case, the system’s selection is opted.

Detailed Specifications and Analysis

Detailed design specifications, as shown in Figure 6 are input for the selected alternative. The
iterative process of numerical model generation, structural analysis, code conformance
checking and optimized member sizing is commenced next.

Structural Optimization
The practically available values of the design variables are given in the lists L1, L2 and L3,
representing slab thickness, bar diameter and bar spacing respectively.

9
L1 = (200, 225, 250, 300, 350, 400, 450, 500, 600, 700, 800, 900, 1000) (mm). (15a)
L2 = (10, 12, 16, 20, 25, 32, 40) (mm). (15b)
L3 = (100, 125, 150, 175, 200, 225, 250, 275, 300) (mm). (15c)

There are 13, 7 and 9 different values for the three design variables respectively. A binary
coding is adopted for the design variables because it is easy to handle. The total length of the
string becomes eleven, with two substrings of length four, representing the slab thickness and
the bar spacing, and one substring of length three, representing the bar diameter. The
population size, the crossover probability and the mutation probability are selected as 10,
0.95 and 0.01, respectively. These values are consistent with other empirical studies with
high crossover probability, low mutation probability, and moderate population size, although
it is found that GAs are not highly sensitive to these parameters.

The individual numbers and the strings generated randomly are shown in the first and second
columns of Table 1 respectively. The third, fourth and fifth columns display the
corresponding values of the three design variables for each individual in the population,
which can be acquired by decoding the first substring of length four, second substring of
length three and last substring of length four, respectively. Column six gives the costs of the
structure for the design represented by the individual strings. Values of the objective function,
which account for the possible violation of constraints, are given in column seven. It can be
seen that constraints are violated for individuals 1, 4, and 10. Table 1 shows that the
individual 3 is the best fit since it has the least cost ($45296) among the ten individuals and
has not violated any constraints. Fitness values are computed for all the ten individuals as
shown in column eight. By applying the proportionate probability of reproduction selection,
the actual count of individuals in the mating pool is shown in column nine. As shown in
column ten, a mating pool is created where individuals 1,2,5,6,8, and 9 get one copy each,
individuals 3 and 7 get two copies, and 4 and 10 die off. The crossover operator is then
applied and the crossover parameters generated, including the mating pair and the cross sites,
are shown in columns eleven, twelve and thirteen. Since there are ten individuals in the
population, there are five matching pairs selected randomly. Individual 1 gets 6, 2 gets 8, 3
gets 7, 4 gets 5, and 9 gets 10. The population after the crossover becomes the initial
population of generation 2, which is processed as shown in Table 2.

Figure 7 shows the relationship between the minimum cost versus the number of generations
for the population size of 10. The minimum cost of $38687, representing a reinforced
concrete section of member thickness 300 mm with reinforcement diameter 25 mm at spacing
225 mm, is found after 5 generations. It is interesting that near-optimal results are obtained
after only 5 generations (approximately 48 new function evaluations) even though the size of
the search space is huge (211 = 2048). The number of points explored is small and represents
only 2.3% or so of the total search space.

Since the population size may play an important role in the value of the minimum cost and in
the number of generations produced, another population size of 8 members has also been
performed. In that case, as shown also in Figure 7 the same minimum cost is found after 6
generations. Again, in order to acquire the near-optimal results, the number of points
explored is small and represents only 2.2% or so of the total search space.

10
Design Report
Figure 8 shows the final design report that provides the type of structure, location, volume,
selected reinforced concrete section, total number of node, total number of element,
calculated crack width, maximum bending moment, shear force and the corresponding
member number at both ultimate and serviceability limit states. The values acquired from the
final design, together with detailed specifications, can be added to the knowledge base to
improve the existing heuristic knowledge.

Conclusions
A coupled microcomputer KBS on optimum design of liquid retaining structure (OPTLIQ)
was implemented to combine expert knowledge with GA optimization, object-oriented
programming, graphics capabilities, KBS technologies, conventional algorithmic models and
relational databases under a windowing environment. The prototype system undertakes all
major design stages including preliminary synthesis, detailed specification, numerical model
generation, finite element analysis, code conformance checking, and member sizing
optimization. The incorporated GA encapsulating reproduction, crossover, and mutation
operators locates the optimal solution quickly after examining a minute portion of the discrete
design alternatives in the design of liquid retaining structures. The system will be an ideal
research tool to validate and enhance our empirical knowledge, which in turn may lead to
more efficient and optimized structural design. It can act as a repository of empirical
knowledge provided by experienced specialists.

References
British Standards Institution (1985). BS 8110: Structural use of concrete, British Standards
Institution, London.
British Standards Institution (1987). BS 8007: Design of concrete structures for retaining
aqueous liquids, British Standards Institution, London.
Chau, K.W. (1992). “An expert system for the design of gravity-type vertical seawalls.”
Engineering Applications of Artificial Intelligence, 5(4), 363-367.
Chau, K.W., Cheng, C.T., and Li, C.W. (2002). “Knowledge management system on flow
and water quality modeling.” Expert Systems with Applications, 22(4), 321-330.
Chau, K.W., and Ng, V. (1996). “A knowledge-based expert system for design of thrust blocks
for water pipelines in Hong Kong.” Water Supply Research and Technology - Aqua, 45(2), 96-
99.
Chau, K.W., and Yang, W.W. (1992). “A knowledge-based expert system for unsteady open
channel flow.” Engineering Applications of Artificial Intelligence, 5(5), 425-430.
Chau, K.W., and Zhang, X.N. (1995). “An expert system for flow routing in a river network.”
Advances in Engineering Software, 22(3), 139-146.
Engelmore, R. and Morgan, T. (1988). Blackboard systems, Addison_Wesley, Wokingham.
Goldberg, D.E., and Kuo, C.H. (1987). “Genetic algorithms in pipeline optimization.”
Journal of Computing in Civil Engineering, ASCE, 1(2), 128-141.
Hayalioglu, M.S. (2001). “Optimum load and resistance factor design of steel space frames
using genetic algorithm.” Structural and Multidisciplinary Optimization, 21(4), 292-299.
Holland, J.H. (1975). Adaptation in natural and artificial systems. University of Michigan
Press, Ann Arbor, Mich.
Hong Kong Building Development Department (1983). Code of Practice for Wind Effects in
Hong Kong, Hong Kong Building Development Department, Hong Kong.
Kumar, B. (1995). Knowledge Processing for Structural Design, Topics in Engineering Vol.
25, Computational Mechanics, Southampton.

11
Lin, S., and Albermani, F. (2001). “Lattice-dome design using a knowledge-based system
approach.” Computer-aided Civil and Infrastructure Engineering, 16(4), 268-286.

12
Figure captions

Figure 1. Architecture of the hybrid knowledge-based system

Figure 2. Characteristics of class Wind Load

Figure 3. Flowchart showing overall design algorithm of OPTLIQ

Figure 4. Flowchart of genetic algorithm for optimum design of discrete reinforced concrete

sections in liquid retaining structures

Figure 5. Screen showing the main menu

Figure 6. Screen showing summary of design specifications

Figure 7. Minimum cost versus number of generation

Figure 8. Screen showing the final design report

13
No Population Thick Bar Bar C(x) ϕ(x) F count Mating pool Mate cs1 cs2
ness size spacing
1 11100101110 1000 12 100 83116 452151 264973 1 11100101110 6 1 9
2 01101101011 400 32 100 86951 86951 630173 1 01101101011 8 9 10
3 01001100111 300 32 250 45296 45296 671828 2 01001100111 7 9 10
4 11011010110 1000 25 225 90550 597195 119929 0 01001100111 5 1 9
5 10010000011 600 40 150 104345 104345 612779 1 10010000011 4 1 9
6 00001001010 1000 20 100 97342 97342 619782 1 00001001010 1 1 9
7 01001010011 300 25 150 45689 45689 671435 2 01001010011 3 9 10
8 10011110110 600 40 225 84586 84586 632538 1 10011110110 2 9 10
9 11100000010 1000 40 125 146246 146246 570878 1 11100000010 10 4 6
10 00000011011 1000 10 100 80671 671828 45296 0 01001010011 9 4 6

Table 1. Details of computations in generation 1

14
No Population Thick Bar Bar C(x) ϕ(x) F count Mating pool Mate cs1 cs2
ness size spacing
1 10001001010 500 20 100 59786 59786 395541 1 10001001010 8 7 9
2 01101101011 400 32 100 86951 86951 368376 1 01101101011 4 3 5
3 01001100111 300 32 250 45296 45296 410031 2 01001100111 9 8 11
4 00010000011 200 40 150 74300 74300 381027 1 00010000011 2 3 5
5 11001100111 900 32 250 90365 410031 45296 0 01001100111 6 6 9
6 01100101110 400 12 100 38048 123360 331967 1 01100101110 5 6 9
7 01001010011 300 25 150 45689 45689 409638 2 01001010011 10 0 9
8 10011110110 600 40 225 84586 84586 370741 1 10011110110 1 7 9
9 11101000010 1000 20 125 92897 170001 285326 1 11101000010 3 8 11
10 01000010011 300 10 150 26239 370620 84707 0 01001010011 7 0 9

Table 2. Details of computations in generation 2

15
 Knowledge Requirements
Interview Knowledge from of BS 8007
with Expert Engineer Literature and BS 8110

Microsoft Windows Environment

Visual Basic Programming Environment

Input/
Procedural Knowledge Acquisition Model Generation
Output
Methods Module Module
ASCII
Data Files

Knowledge Base Shell

(Visual Rule Studio)
Finite
System Element
Inference Knowledge Base
Interfaces Structural
Engine
Analysis
Blackboard
Knowledge (Context)
Package
Module (ABAQUS)
Backward
Chaining Design Design
Process Entities
GA
Microsoft
Optimization
Access
Forward Process Design Module
Chaining Control Stage Databases
(Moment
Coefficients,
Sectional
Code Properties,
Conformance Final
Checking Member
User Interfaces Explanation Module Module Details)

User

Figure 1.

16
CLASS BBWindLoad
WITH terrainCategory COMPOUND
general terrain,
builtup terrain
WHEN CHANGED
BEGIN
FORGET BBWindLoad.basicWindPressure
END
SEARCH ORDER CONTEXT
WITH methodOfSelection COMPOUND
found by system,
user defined
WHEN CHANGED
BEGIN
FORGET BBWindLoad.shapeFactorCs
FORGET BBWindLoad.shapeFactorCsx
FORGET BBWindLoad.shapeFactorCsy
FORGET BBWindLoad.heightAspectFactorChx
FORGET BBWindLoad.heightAspectFactorChy
FORGET BBWindLoad.forceCoefficientCfx
FORGET BBWindLoad.forceCoefficientCfy
FORGET BBWindLoad.basicWindPressure
END
SEARCH ORDER CONTEXT
WITH shapeFactorCs NUMERIC
SEARCH ORDER RULES
WITH shapeFactorCsx NUMERIC
SEARCH ORDER RULES
WITH shapeFactorCsy NUMERIC
SEARCH ORDER RULES
WITH heightAspectFactorChx NUMERIC
SEARCH ORDER RULES
WITH heightAspectFactorChy NUMERIC
SEARCH ORDER RULES
WITH forceCoefficientCfx NUMERIC
SEARCH ORDER RULES
WITH forceCoefficientCfy NUMERIC
SEARCH ORDER RULES
WITH topLevel NUMERIC
WHEN CHANGED
BEGIN
FORGET BBWindLoad.basicWindPressure
PURSUE BBWindLoad.basicWindPressure
END
SEARCH ORDER CONTEXT
WITH basicWindPressure NUMERIC
SEARCH ORDER RULES
WITH designedWindPressureX NUMERIC
SEARCH ORDER RULES CONTEXT
WITH designedWindPressureY NUMERIC
SEARCH ORDER RULES CONTEXT

Figure 2.

17
 start

Input structural specifications: shape,

geometry, exposure condition

No
Warning Are inputs within
message available ranges?

Yes

Estimate maximum moment and shear from

heuristic moment & shear coefficient database

Effect optimization by Genetic Algorithms

Change default
properties &/or unit
costs of materials? Yes

No

Select alternative

Input: support specification, imposed

load, wind load, load combination

View design
specification
Generate numerical model

Perform finite element analysis

Revise
Effect optimization by Genetic Algorithms selected
member
size

Get best alternative with minimum total material cost

No
Is best alternative same
as selected alternative?

Yes

Generate final member details

Produce design report

Stop

Figure 3.

18
 Start

Input: number of member group, population size,

crossover probability, and mutation probability

Generation = 1

Randomly generate population comprising binary codes

Compute cost, crack width and stresses

of individual designs in population

Compute violation coefficients (p, q & r)

and objective function ϕ(x)

Compute individual fitness and actual count

Store best individual

Generation =
generation + 1
Create mating pool

Apply two-site crossover operator

Stop
Apply mutation operator

No Yes
Print best individual
has converged?
together with cost

Figure 4.

19
Figure 5.

20
Figure 6.

21
 50000
Minimum cost ($)

45000
Population size
=8
40000
Population size
35000 = 10

30000
1 3 5 7 9 11
Number of generation

Figure 7.

22
Figure 8.

23