International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117

Contents lists available at ScienceDirect

International Journal of
Rock Mechanics & Mining Sciences
journal homepage: www.elsevier.com/locate/ijrmms

Indirect estimation of the ultimate bearing capacity of shallow

foundations resting on rock masses
Shervin Tajeri a, Ehsan Sadrossadat b,n, Jafar Bolouri Bazaz a
a
Department of Civil Engineering, Ferdowsi University of Technology, Mashhad, Iran
b
Young Researchers and Elite Club, Mashhad Branch, Islamic Azad University, Mashhad, Iran

art ic l e i nf o a b s t r a c t

Article history: The success of a foundation design for structures is to precisely estimate the bearing capacity of un-
Received 16 August 2014 derlying soils or rocks. To avoid the elaborate in-situ experimental methods, several approaches pre-
Received in revised form sented by various researchers for the estimation of the bearing capacity factor. Despite this fact, there still
19 August 2015
exists a serious need to develop more robust predictive models. The aim of this paper is to propose a
Accepted 18 September 2015
novel formulation for the ultimate bearing capacity of shallow foundations resting on/in rock masses,
Available online 27 September 2015
using a powerful evolutionary computational technique, namely linear genetic programming. Thus, a
Keywords: comprehensive set of data is collected to develop the model. In order to evaluate the validity of the
Rock mass properties obtained model, several analyses are conducted and compared with those provided by other researchers.
Ultimate bearing capacity
Consequently, the results clearly demonstrate the proposed model accurately characterize the bearing
Shallow foundation
capacity factor and reach a notably better prediction performance than the traditional models.
Evolutionary computation
Linear genetic programming & 2015 Elsevier Ltd. All rights reserved.

1. Introduction groups: (1) codes, (2) analytical methods, (3) semi-empirical

Shallow foundations support structures at a shallow depth
below the ground surface and transmit applied loads to the un-
derlying materials such as soils, rocks or intermediate geo-mate-
rials. In general, any foundation design must satisfy at least two
important criteria 1–3: (1) obtaining sufficient bearing capacity of
underlying layer against ultimate failure, and (2) achieving ac-
ceptable total or differential settlements under working loads.
Although, the design of foundations resting on or in rock masses is
commonly controlled by the settlement criterion, the bearing ca-
pacity of rock mass must be estimated to evaluate the stability 4.
Therefore, in order to provide a precise and efficient design of a
foundation, it is crucial to account for the bearing capacity of the
rock mass beneath it. According to the rock mass properties and
the beneath layer of it, the failure may occur in several
mechanisms 5. Bearing capacity failure in an overloaded rock
foundation is one of them. The mode of bearing capacity failure
mainly depends on the ratio of space between joints (S) to foun-
dation width (B), joint condition (open or closed) and direction,
rock type as summarized in Table 1and schematically represented
in Fig. 15–7.
The most usually utilized approaches to determine the bearing
capacity (qult) of foundations on rocks can be classified into four
n
Corresponding author.
E-mail address: ehsan.sadrossadat@gmail.com (E. Sadrossadat).
methods, and (4) in-situ and full-scaled testing methods 4. Codes
often propose conservative values for estimating the allowable
bearing pressure or ultimate bearing capacity 8–10. These pre-
sumed values are derived from local experience and geology from
a particular site, however, the engineer should ensure that they are
applicable to the particular conditions relevant to the considered
site 4. On the one hand, analytical methods are based on bearing
capacity theories, including limit equilibrium methods, using in-
itial assumptions and relate qult to footing geometry and rock
properties such as those equations provided by 6,11,12. On the other
hand, semi-empirical and empirical methods are often obtained by
the correlation between qult and rock mass properties, based on
empirical observations and experimental test results such as
equations made by 13–15. General forms of mostly utilized and
traditional equations proposed by various researchers in the lit-
erature are summarized in Table 24,7–9.
As represented in Table 2, analytical methods include terms of
physical and mechanical properties of rock mass and geometry of
the foundation but not include terms of rock type, classification
and qualitative parameters of rock mass Also, semi-empirical and
empirical methods often relate qult to quantitative and qualitative
of rock mass and are not prepared for the geometry of foundations
or space between joints (Table 2).
The bearing capacity of a shallow foundation on a jointed rock
mass mostly depends on geometry of foundation, the ratio of joint
spacing to foundation breadth or loading width, as well as rock
mass qualities such as joint conditions (open or closed), rock type

http://dx.doi.org/10.1016/j.ijrmms.2015.09.015
1365-1609/& 2015 Elsevier Ltd. All rights reserved.
108 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117
Table 1
Bearing capacity failure modes in layered and jointed rock mass.
Ratio of joint spacing (S) to footing Description
width (B)
S
<1 Just jointed rock mass
B columns”
S
<1 Just jointed rock mass
B
S
>1 Just jointed rock mass
B
S
>1 Thick rigid layer of rock mass over weaker
B layer
S
<1 Thin rigid layer of rock mass over weaker
B layer
and rock mass strength 5,12,15–18. In regarding to the equations in
the literature, there is not a comprehensive model including si-
multaneously both quantitative and qualitative parameters, such
as foundation geometry and RMR. Thus, the complexity of analysis
of bearing capacity behavior and accounting for the influences of
different parameters on the bearing capacity factor implies that
there is the necessity for a more comprehensive model.
By progressing in computational software and hardware sys-
tems, several computer-aided modeling and soft computing
techniques such as artificial neural networks (ANNs), adaptive
neuro-fuzzy system (ANFIS), fuzzy inference system (FIS), support
vector machine (SVM) and genetic programming (GP) have been
realized by various researchers in several civil engineering do-
mains. Such computing techniques have a lot of features that have
made them attractive choice for predicting different problems. The
first feature is they are data-driven self-adaptive methods. That
means they do not require many prior assumptions about the
models of the problem under study. They automatically learn from
data to determine the structure of a prediction model. These
techniques become more attractive because of their capability of
information processing, such as non-linearity, high parallelism,
robustness, fault and failure tolerance and their ability to gen-
eralize. Besides, these techniques have been successfully employed
to solve problems in civil engineering field 19–29.
The aim of this paper is to utilize a powerful branch of genetic
programming (GP), namely linear genetic programming (LGP), to
derive a more comprehensive predictive model for the ultimate
bearing capacity of shallow foundations resting in/on jointed rock
masses. A comprehensive and reliable set of data including 102
rock socket, centrifuge rock socket, plate load and large-scaled
footing load test results are collected to develop the model. In
order to verify the robustness of the obtained model several va-
lidation and supplementary study phases are conducted.
2. Genetic programming
Genetic programming, as a subset of evolutionary computa-
tional intelligence approaches, considers the synthesis of
B B
Thin or Thick Rock
Mass
S S
Weak Layer
Fig. 1. A typical sketch of a shallow foundation resting on a (a) layered or (b) jointed rock mass.
Joints Direction Bearing capacity failure mode
Open Vertical to sub-vertical Uniaxial compression of the “rock
Closed 90° to 70° General wedge shear failure
Wide Horizontal to sub-horizontal Splitting failure
Flexure failure
Punching failure
Darwinian ideas of genetic inheritance, natural variation and se-
lection to solve complicated problems. In general, in genetic pro-
gramming (GP), inputs and corresponding output data samples are
known and the main goal is to generate predictive models relating
them without any prior assumptions 30,31. Typically in GP, a po-
pulation of individuals initialized and members of the population
are ranked according to a fitness function. Those members with
the highest fitness ranking are given a higher chance to become
parents for the next generation, the offspring. The approach that is
utilized to generate offspring from the parents, is termed the re-
production heuristic. Then selected members are transformed, by
chance, into new members via mutation and recombination or
crossover. These steps repeat until the convergence conditions are
satisfied and the fittest member is selected 32,33.
2.1. Linear genetic programming
There are several branches of GP where individuals, i. e. pro-
grams or encoded solutions, are represented in different ways.
These are tree-shaped, graph-shaped and linear encodings 30.
Tree-shaped expressions or encodings typically define a root node
that represents the output. Each node can have one or more child
nodes. Some nodes represent operations on children, unary op-
erations such as abs, exp, and log, or binary operations such as add
(þ), mult (  ), and div (/). Nodes without children are called leaf
nodes (or terminals) that represent input values or evolved con-
stant values within the system. The graph-based expressions are
similar to the tree-shaped ones, but child nodes are no longer
unique or multiple nodes may refer to the same node as their
children 34.
Linear genetic programming (LGP) is a new subset of GP with a
linear structure similar to the DNA molecule in biological gen-
omes. In LGP, expressions of a functional programming language
(such as LISP) are substituted by programs of an imperative lan-
guage (such as C/C þ þ ) 31,35. Fig. 2 represents a comparison of
structure of a program evolved by (a) tree-based, (b) graph-based
GP, and (c) Linear GP. As shown in this figure, a linear genetic
program can be seen as a data flow graph generated by multiple
usage of register content. In classical tree-based and graph-based
Jointed Rock Mass
 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117 109

Table 2
General forms of equations made by analytical methods.

Reference Equation Method Factor

Terzaghi (1943) qult = cNc + 0.5γBNγ + γDNq Analytical Nc = 2Nϕ0.5 (Nϕ + 1) ; Nγ = Nϕ0.5 (Nϕ − 1); Nq = Nϕ2
c E 0.1qu (Kulhawy and Goodman, 1987)
Carter and Kulhawy (1988) qult = qu (m + s) Semi-Empirical m = mi exp (
GSI − 100
); s = exp (
GSI − 100
)
28 9
Goodman (1989) qult = qu (
1 S
(N ( )(Nϕ− 1) / Nϕ − 1)) Analytical Nϕ = tan2 (45 + )
ϕ
Nϕ − 1 ϕ B 2

Bowels (1996) qult = qu × (RQD)2 Empirical

qult: bearing capacity of shallow foundation on rock; D: depth of foundation below ground surface; c: the cohesion intercepts for the rock mass; ϕ: angle of internal friction
for the rock mass; γ: effective unit weight of the rock mass; B:breadth or width of foundation; NØ, Nc, Nq and Nγ: non-dimensional bearing capacity factors as exponential
functions of ϕ; RQD: Rock Quality Designation; GSI: geological strength index (rock mass classification); m and s: material constant in the Hoek and Brown failure criterion; S:
discontinuity spacing; S/B: ratio of joint spacing to foundation width and qu: Unconfined compressive strength of rock.

GP, the data flow is more rigidly determined by the graph struc-
ture of the program 34,35.
There are some main reasons for using LGP. In LGP the in-
dividuals are linear-shaped (strings). Computers do not naturally
run tree-shaped programs. Therefore, slow interpreters have to be
used as a part of classical tree-based or graph-based GP. Thus, the
use of an expensive interpreter is avoided in LGP. Consequently, a
LGP system can run several orders of magnitude faster than
comparable interpreting systems. The enhanced speed of the lin-
ear variants of LGP permits conducting many runs in realistic
timeframes. This leads to deriving consistent and high-precision
models with little customization 27,36–39. One of the most pro-
blems that arise in GP system is “bloat”, that is the tendency of GP
individuals to grow in size without increasing in quality. This leads
to take a lot of memory and slows down the evolutionary process.
To avoid bloating, LGP use variable-size chromosomes that are
limited to a maximum number of instructions (genes) 36.
2.2. Genetic operators
Crossover (recombination) occurs between two or more in-
struction blocks whereas mutation occurs on a single instruction
or chromosome. The crossover (recombination) operation works
by exchanging continuous sequences of instructions between
parents. As it is seen in Fig. 3(a), a segment of random position and
arbitrary length is selected in each of the two parents (f(0) and g
(0)) and exchanged. If one of the two children would exceed the
maximum length, crossover is aborted and restarted with ex-
changing equally sized segments. Two commonly used types of
standard LGP mutations are micro and macro mutation. The micro
mutation can change an operation type (e.g. change v[1]þ v[2] to v
[1]  v[2]), or can change the arguments of an operation (e.g.
change v[1] þ1 to v[1]þ v[1]). The macro mutation can delete an
operation (e.g. change from v[1] þv[2] to v[2]), or insert an op-
eration (e.g. change from v[1] þv[2] to v[1] þ(v[1]  v[1])) 27,31,36.
Fig. 3 represents a typical crossover and mutation operation in
LGP.
f[0] = v[0] - v[1] Crossover f΄[0] = v[0] + v[3]
g[0] = v[2] + v[3] g΄[0] = v[2] - v[1]
Mutation
f[0] = v[0] - v[1] f΄[0] = v[0] / v[1]
Fig. 3. Typical variation operations in LGP: (a) crossover and (b) mutation.
3. Numerical simulation of bearing capacity
In order to reach reliable estimations of the bearing capacity of
shallow foundations on rock masses, the influence of several
parameters should be incorporated into the model development.
Regarding to the general forms of the existing equation for indirect
estimation of qult of shallow foundations on rock masses, mainly
depends on embedment or excavation depth (D), angle of internal
friction for the rock mass (φ), effective unit weight of the rock
mass (γ), breadth or width of foundation (B), ratio of joint spacing
to foundation width (S/B) and unconfined compressive strength of
rock (qu). The rock mass qualitative parameters such as rock
quality designation (RQD) index and geological strength index
(GSI) are widely used to develop empirical and semi-empirical
equations for the evaluation of rock mass properties 7–9,13,15,16,40.
The RMR index is another qualitative parameter that has found
wide applications in various types of geological engineering pro-
jects. This parameter was introduced by 41,42 to provide reliable
estimation of rock mass properties. The RMR value is sum of six
parameters: (1) unconfined compressive strength of intact rock
material (qu), (2) rock quality designation (RQD) index, (3) joint or
discontinuity spacing (S), (4) joint condition, (5) ground water
condition and (6) adjustment of joint orientation. Thus, the RMR
parameter implicitly includes the effect of several important
parameters for characterizing the rock mass properties. In this
paper, RMR employed as an input parameter in order to take into
account the effects of both qualitative parameters as well as other

Fig. 2. A comparison of a program structure evolved by tree-based (a), graph-based GP (b), and Linear GP (c) for the expression of y¼ f(v[i])¼ (v[1] þ 2)2/v[2].
110 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117
influencing quantitative parameters to predict the bearing capa-
city of shallow foundations on rock masses. Consequently, the
proposed formulation of qult is considered as a function in terms of
following parameters:
⎛ only at the end of the data analysis and when the best model is
S ⎞
qult = f ⎜ RMR, γ , qu, D, B, , ϕ⎟
⎝ B ⎠
where qult (MPa) is the ultimate bearing capacity of the shallow
foundation resting on rock mass, RMR, rock mass rating, γ (kN/m3),
effective unit weight of the rock mass, qu (MPa), the unconfined
compressive strength of rock, D (m), the embedment or excavation
depth, B (m): breadth or width of the foundation, S/B, the ratio of
joint spacing to foundation width (equivalent diameter) and ϕ
(degree) is the angle of internal friction for the rock mass.
3.1. Experimental database
A comprehensive database including 102 elaborate and ex-
pensive experimental data from different studies, is employed to
develop the LGP-based model. The database contains results of 49
rock socket tests (six centrifuge rock socket tests), forty plate load
tests and thirteen load test on scaled footings and the results of
experiments on different sizes of circle footings. All the rock
sockets in the database are circular for which the end bearing
capacity (tip resistance) could be isolated, separating it from the
shaft resistance of the rock sockets. The major bearing capacity
values have been reported by 18. It should be noted that all bearing
capacity values in the database are axial capacity that interpreted
from load test results. Load test results are indicated in the form of
curves describing foundation displacement as a function of applied
load. The failure is defined as the point in which excessive dis-
placements take place under a relatively small increase in loading,
typically associated with an abrupt change in the load-displace-
ment curve characteristics. In the collected database, the ultimate
bearing capacity (qult) values are initially obtained or interpreted
from load-displacement curves using 43 interpreting method.
Different parts of the employed database have been used by var-
ious researchers to develop models or to conduct different studies
on modeling the bearing capacity of foundations resting on or in
rock masses and also described in National Cooperative Highway
Research Program (NCHRP) report 651 (2010) 7,18,29,44,45. The de-
scriptive statistics and the set of collected data incorporated in the
LGP model are given in Table 3 and Table 4, respectively.
3.2. Data classification
Overfitting is one of the essential problems in generalization of
the soft computing techniques. Overfitting occurs when the
models developed perform very well on training and validation
data, but do not perform so well on the unseen, applied data. It is a
case in which the error on the learning set is driven to a very small
value, but when new data presented to the model, the error be-
comes very large. An approach to avoid overfitting is to test in-
dividuals from the run on a validation set to find a better
Table 3
Descriptive statistics of parameters in database used to develop LGP-based Models.
Parameter RMR γ (kN/m3) qu (MPa)
Mean 62.69 21.57 4.33
Standard deviation 24.37 2.43 8.17
Sample variance 593.76 5.90 66.82
Kurtosis  0.43  0.40 17.42
Skewness  0.69 0.41 3.81
Minimum 15.00 16.82 0.24
Maximum 100.00 28.30 55.00
generalization 30,46. In another way, overfitting occurs when the
models developed perform very well on training and validation
data, but do not perform so well on the unseen, test data. hy-
pothetically with the same reason, another data set should be used
obtained based on train and validation data sets. For this purpose,
(1)
the available data set is classified into three sets: (1) training
(learning), (2) validation (check), and (3) test subsets. The training
set is used to fit the models, the validation set is used to estimate
prediction error for model selection and the test set is used for
evaluation of the generalization error of the final chosen model.
The training, validation and test data are usually taken as 50–70%,
15–25% and 15–25% of all data, respectively, for the development
of models in artificial intelligence (AI) techniques. These ratios are
selected based on previously suggested values 27,31,38. In the pre-
sent study, Among 102 data, 80% data set are randomly taken as
training (learning) and validation (check) process (60%, i.e. 61 data
vectors for the training process and 20%, i.e., 2 data as the vali-
dation data). The 20% remaining data set, i. e. 21 data, are used for
the testing of the obtained model.
4. Development of the LGP-based model
Several runs are performed to provide LGP-based models for
the prediction of the ultimate bearing capacity (qult) of shallow
foundation resting on/in jointed (non-fractured) rock masses. The
LGP parameters are changed for each run. The parameters are
selected based on previously suggested values 27,38,47,48. The de-
velopmental processes were controlled by considering evolu-
tionary parameters such as population size, probability of cross-
over, probability of mutation and selecting arithmetic operators
and mathematical functions. The success of the LGP algorithm
usually depends on increasing the initial and maximum program
size parameters so three optimal population size 500, 1000 and
1500 values are used for the prediction. Two levels are considered
for each probability of crossover and mutation 0.5 and 0.95. During
the process the complexity of evolved functions increases, there-
fore, in order to prevent decreasing the speed of algorithms two
optimal values (128, 512) are considered for the maximum pro-
gram size as trade-offs between the running time and the com-
plexity of the solutions. Two values (10, 20) are set for the number
of demes. Note that demes are semi-isolated subpopulations that
evolution proceeds faster in them in comparison to a single po-
pulation of equal size 31. In the present study, to obtain a model
match or similar to those provided by various researchers in the
literature and to reach a simple equation for the aim of hand
calculating, four basic arithmetic operators (þ ,  ,  ,/) and basic
mathematical functions (√) are utilized to formulate the LGP-
based models. The termination criterion for each run is considered
based on the number of generations that run has gone without
improving. The number of generations without improvement is set
to 1000. To evaluate the fitness of the evolved program, the
average of the squared errors is used. The resulting run is then
D (m) B (m) S/B φ (Degree) qult (MPa)
4.59 0.57 5.75 30.94 10.08
4.97 0.51 7.33 4.93 12.84
24.75 0.26 53.67 24.28 164.94
1.19 5.13 4.26 0.39 11.46
1.27 1.80 2.11 0.12 2.96
0.00 0.05 0.35 20.00 0.25
20.73 2.74 36.69 45.00 75.60
Table 4
The sets of collected data incorporated in the LGP model.

Case No. Reference Rock Type Description Type of Load Test RMR γ(kN/m3) D(m) B(m) qu(MPa) S/B ϕ(o) qult(MPa)

1 Abu-Hejleh and Att- Weathered Claystone Good Rock Rock Socket 70 19.89 4.91 1.07 0.63 1.86 29.3 2.63
2 wooll (2005) Blue and sandy claystone, thinly bedded, very hard Good Rock Rock Socket 70 19.89 4.27 1.22 0.80 1.62 29.3 2.54
3 Blue and clayey sandstone, well cemented, very hard Good Rock Rock Socket 70 19.89 9.17 1.37 10.49 2 29.3 15.23
4 Pierre shale, very well cemented, very hard Very Good Rock Rock Socket 82 19.01 6.10 1.22 22.98 1.75 25 26.33
5 Blue claystone with occasional interbeds of sandstone and silt-stone Good Rock Rock Socket 70 21.27 1.10 0.79 1.21 3.46 30 6.94
6 Claystone, weathered Good Rock Rock Socket 70 19.89 1.52 0.76 0.48 3.2 30 2.25
7 Baker (1985) Hardpan (hard-bearing till). Till has a qu comparable to that of rock Good Rock Rock Socket 70 18.86 18.29 1.28 1.38 2.38 35 5.84
8 Till Good Rock Rock Socket 68 17.45 20.73 1.92 0.57 1.03 35 2.29
9 Burland and Lord Hardpan (hard-bearing till). Till has a qu comparable to that of rock Good Rock Rock Socket 80 20.28 18.29 0.76 1.11 2.4 35 4.79

S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117
10 (1970) Grade IV chalk, rubbly, partly-weathered chalk with bedding and Poor Rock Plate Load Test 20 19.25 1.82 0.91 0.66 0.67 28 0.59
jointing
11 Grade V chalk, structure less remoulded chalk containing small lumps Very Poor Rock Plate Load Test 15 19.25 3.42 0.86 0.89 0.71 28 0.50
of intact chalk
12 Grade IV chalk, rubbly, partly-weathered chalk with bedding Very Poor Rock Plate Load Test 15 19.25 3.42 0.86 1.14 0.71 28 0.60
13 Grade III chalk, rubbly to blocky unweathered chalk. Very Poor Rock Plate Load Test 15 19.25 7.77 0.86 1.25 0.71 30 0.60
14 Butler and Lord (1970) Lower grey chalk marl Fair Rock Plate Load Test 15 19.25 2.13 0.14 0.90 0.36 28 0.48
15 Lower grey chalk marl Fair Rock Plate Load Test 72 19.25 5.18 0.14 0.82 2.14 28 3.35
16 Occasional hard lumps of intact chalk and flint stones in a weathered Fair Rock Plate Load Test 60 19.25 5.49 0.14 0.82 2.18 30 2.40
chalk matrix
17 Occasional hard lumps of intact chalk and flint stones in a weathered Fair Rock Plate Load Test 55 19.25 9.14 0.14 0.82 2.18 30 1.00
chalk matrix
18 Occasional hard lumps of intact chalk and flint stones in a weathered Fair Rock Plate Load Test 70 19.25 10.67 0.14 0.82 2.18 30 0.96
chalk matrix
19 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 68 19.25 12.19 0.14 0.55 2.18 30 0.96
chalk matrix
20 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 35 19.25 7.62 0.14 0.60 2.18 30 1.15
chalk matrix
21 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 35 19.25 10.36 0.14 0.61 2.18 30 1.46
chalk matrix
22 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 35 19.25 6.91 0.14 0.58 2.18 30 1.61
chalk matrix
23 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 40 19.25 2.34 0.14 0.50 2.18 30 1.11
chalk matrix
24 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 50 19.25 3.86 0.14 0.53 2.18 30 1.03
chalk matrix
25 Occasional hard lumps of intact chalk and flint stones in a weathered Poor Rock Plate Load Test 50 19.25 5.39 0.14 0.90 2.18 30 0.46
chalk matrix
26 Lower grey chalk marl Poor Rock Plate Load Test 35 19.25 5.79 0.14 0.90 2.14 28 2.07
27 Lower grey chalk marl Poor Rock Plate Load Test 35 19.25 4.27 0.14 0.90 2.14 28 2.00
28 Lower grey chalk marl Poor Rock Plate Load Test 15 19.25 11.73 0.14 0.87 2.14 28 3.50
29 Diabase breccia, highly fractured, RQD ¼ 10% Fair Rock Rock Socket 20 26.51 19.00 1.20 15.00 0.51 35 8.90
30 Limestone, intact, RQD ¼ 100% Fair Rock Rock Socket 75 23.56 13.50 1.20 2.50 2.29 37 8.90
31 Glos and Briggs (1983) Sandstone, horizontally bedded, shaley, RQD ¼ 74% Fair Rock Rock Socket 55 22.57 14.02 0.61 8.36 1 30 10.10
32 Sandstone, horizontally bedded, shaley, with some coal stringers, RQD Fair Rock Rock Socket 60 22.57 15.24 0.61 9.26 1 30 13.10
¼ 88%
33 Goeke and Hustad Clay-shale, with occasional thin limestone seams Good Rock Rock Socket 78 20.56 8.80 0.76 0.81 4.01 24 4.69
(1979)
34 Hummert and Cooling Shale, thinly bedded with thin sandstone layers Good Rock Rock Socket 65 25.74 0.30 0.46 3.82 2.65 25 9.33
(1988)
35 Jubenville and Hep- Shale, unweathered Good Rock Rock Socket 65 21.50 1.52 0.31 1.08 6.39 40 2.98
worth (1981)
36 Lake (1970) Grade V chalk, completely weathered, structurelessremoulded chalk Good Rock Plate Load Test 70 19.25 7.70 0.14 0.46 14.18 28 2.39
containing small lumps of intact chalk
37 Lake and Simons Chalk Very Good Rock Plate Load Test 90 16.82 12.12 0.14 1.04 14.15 38.5 12.26

111
38 (1970) Chalk Very Good Rock Plate Load Test 80 17.61 7.70 0.14 1.04 14.15 38.5 5.27
 112
Table 4 (continued )

Case No. Reference Rock Type Description Type of Load Test RMR γ(kN/m3) D(m) B(m) qu(MPa) S/B ϕ(o) qult(MPa)

39 Chalk Very Good Rock Plate Load Test 92 18.08 14.48 0.14 1.04 14.15 38.5 14.75
40 Leung and ko (1993) Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 28.30 4.19 1.06 2.10 2.86 20 6.51
triFotinguge model)
41 Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 25.15 4.19 1.06 6.70 2.86 20 16.10
triFotinguge model)
42 Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 26.72 4.19 1.06 4.20 2.86 20 10.90
triFotinguge model)
43 Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 23.58 4.19 1.06 5.40 2.86 20 15.70
triFotinguge model)
44 Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 24.37 4.19 1.06 8.50 2.86 20 23.00

S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117
triFotinguge model)
45 Gypsum mixed with cement is used as pseudo rock Good Rock Rock Socket (Cen- 70 23.58 4.19 1.06 11.30 2.86 20 27.70
triFotinguge model)
46 Lord (1997) Chalk, Grade C, medium high density Very Poor Rock Plate Load Test 15 19.25 0.00 0.87 0.31 0.7 30 0.30
47 Chalk, Grade C, medium high density Very Poor Rock Plate Load Test 15 19.25 0.00 0.87 0.92 0.7 30 0.50
48 Chalk, Grade B & C, low density Very Poor Rock Plate Load Test 15 19.25 0.00 0.87 0.24 0.7 30 0.25
49 Chalk, Grade B & C, low density Very Poor Rock Plate Load Test 18 19.25 0.00 0.87 0.56 0.7 30 0.50
50 Chalk, Grade D, structureless or remoulded mélange, o 35% com- Poor Rock Plate Load Test 20 19.25 0.00 0.87 0.50 0.7 30 0.50
minuted chalk matrix,4 65% coarse fragments
51 Maleki and Hollberg Marlstone with shorite crystals Good Rock Plate Load Test 62 27.67 0.00 0.15 13.80 13.03 28 20.00
(1995)
52 Mallard (1977) Chalk, weak, weathered, fractured with open fissures Good Rock Plate Load Test 80 19.25 8.69 0.45 0.91 5.49 30 5.00
53 McVay, Ko and Otero Limestone Good Rock Rock Socket 70 23.56 5.49 2.74 1.92 0.72 40 4.51
54 (2006) Limestone Good Rock Rock Socket 70 23.56 5.49 2.74 8.47 0.72 40 5.75
55 Nitta et al. (1995) Granite, weathered Good Rock Plate Load Test 80 25.52 0.00 0.30 1.07 6.1 41.3 18.00
56 Pellegrino (1974) Tuff Good Rock Plate Load Test 70 22.17 0.00 0.30 4.72 6.6 29.83 10.53
57 Tuff Good Rock Plate Load Test 72 22.17 0.00 0.30 4.03 6.6 29.83 10.00
58 Tuff Good Rock Plate Load Test 70 22.17 0.00 0.30 4.03 6.6 29.83 11.16
59 Tuff Good Rock Plate Load Test 75 22.17 0.00 0.30 3.35 6.6 29.83 12.00
60 Tuff Good Rock Plate Load Test 65 22.17 0.00 0.30 2.00 6.6 29.83 5.92
61 Pells and Turner (1979 Strong sandstone, medium to strong - core sections can be broken by Good Rock Footing 65 25.15 0.00 0.08 14.00 26.42 34 75.60
& 1980) hand with difficulty and lightly scored with a steel knife, slightly
fractured
62 Strong sandstone, medium to strong - core sections can be broken by Good Rock Footing 65 25.15 0.00 0.05 11.61 36.69 34 72.80
hand with difficulty and lightly scored with a steel knife, slightly
fractured
63 Strong sandstone, medium to strong - core sections can be broken by Good Rock Footing 70 24.05 0.60 0.15 9.99 12.95 28 25.00
hand with difficulty and lightly scored with a steel knife, slightly
fractured
64 Strong sandstone, medium to strong - core sections can be broken by Good Rock Rock Socket 70 24.05 0.94 0.29 6.00 6.83 28 13.80
hand with difficulty and lightly scored with a steel knife, slightly
fractured
65 Strong sandstone, medium to strong - core sections can be broken by Good Rock Rock Socket 70 24.05 1.36 0.29 6.00 3.15 27 7.67
hand with difficulty and lightly scored with a steel knife, slightly
fractured
66 Strong sandstone, medium to strong - core sections can be broken by Good Rock Footing 80 22.48 0.00 0.62 0.30 4.96 27 4.50
hand with difficulty and lightly scored with a steel knife, slightly
fractured
67 Very Weak sandstone - rock structure is evident but frequentzones of Good Rock Footing 80 22.48 0.00 0.37 0.30 6.5 27 3.75
sugary sandstone - crumbled by hand, highly weathered and fractured
68 Fresh shale Intact Rock Footing (Model Footing) 90 23.67 3.32 0.41 35.00 1.47 27 23.57
69 Radhakrishna and Siltstone, medium-hard, fragmented Fair Rock Rock Socket 60 19.65 7.32 0.70 9.00 0.86 32 13.10
Leung (1989)
70 Spanovich and Garvin Shale Fair Rock Footing 60 23.67 0.00 0.46 1.45 1.99 36 4.44
71 (1979) Shale Fair Rock Footing 70 23.67 0.00 0.61 1.45 2.5 36 6.62
72 Shale Fair Rock Footing 50 23.67 0.00 0.76 1.45 2.01 36 3.47
73 Thorne (1980) Shale Fair Rock Rock Socket 50 22.52 6.90 0.45 34.00 1.35 27 28.00
74 Sandstone Fair Rock Rock Socket 50 25.15 7.07 0.45 12.50 1.35 34 14.00
75 Sandstone, fresh, defect free Fair Rock Rock Socket 70 25.52 7.90 0.45 27.50 2.03 34 50.00
76 Shale, occasional recemented moist fractures and thin mud seams, Fair Rock Rock Socket 50 21.60 7.70 1.22 55.00 0.5 27 27.80
intact core lengths 75-250 mm
77 Ward and Burland Grade I chalk, hard and brittle Poor Rock Plate Load Test 40 19.25 8.07 0.86 2.07 0.35 30 1.13
78 (1968) Grade II chalk, medium-hard chalk, joints more than 0.66 ft apart and Poor Rock Plate Load Test 20 19.25 7.40 0.86 1.60 0.35 30 1.00
closed
79 Grade III chalk, unweathered chalk, joints 0.2 - 0.66 ft apart, open up to Poor Rock Plate Load Test 20 19.25 5.44 0.86 0.91 0.71 27 0.75
0.01 ft
80 Grade IV chalk, weathered chalk with bedding and jointing, joints0.033 Poor Rock Plate Load Test 15 19.25 3.18 0.86 0.57 0.71 27 0.40
- 0.2 ft apart and open up to 0.066 ft
81 Webb (1976) Diabase, highly weathered Fair Rock Rock Socket 60 18.04 12.20 0.62 0.52 0.99 35 1.32
82 Williams (1980) Mudstone, moderately weathered Very Good Rock Rock Socket 70 22.52 0.00 0.30 1.14 6.6 35 3.68
83 Mudstone, moderately weathered Very Good Rock Footing 81 22.52 0.00 0.60 0.54 5.08 30 4.51

S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117
84 Mudstone, moderately weathered Very Good Rock Footing 81 22.52 0.00 1.00 0.57 3.05 30 4.98
85 Mudstone, moderately weathered Very Good Rock Footing 90 22.52 0.00 0.10 0.60 18.29 30 7.20
86 Mudstone, moderately weathered Very Good Rock Footing 100 22.52 0.25 0.10 0.44 30.48 29 10.57
87 Mudstone, moderately weathered Very Good Rock Rock Socket 85 22.52 2.01 0.30 0.65 8.13 31 5.16
88 Mudstone, moderately weathered Very Good Rock Rock Socket 95 22.52 1.00 0.10 0.75 21.34 33 9.26
89 Mudstone, moderately weathered Very Good Rock Rock Socket 88 22.52 1.00 0.30 0.67 8.13 31 4.87
90 Mudstone, moderately weathered Very Good Rock Rock Socket 100 22.52 1.52 0.10 0.57 24.38 30 12.48
91 Mudstone, moderately weathered Very Good Rock Rock Socket 100 22.52 0.20 0.10 0.45 21.34 29 10.19
92 Mudstone, moderately weathered Very Good Rock Rock Socket 100 22.52 2.20 0.10 0.52 9.14 30 13.09
93 Mudstone, moderately weathered Very Good Rock Rock Socket 85 22.52 1.80 0.60 1.93 3.56 37 9.02
94 Mudstone, moderately weathered Very Good Rock Rock Socket 70 22.52 3.00 1.00 1.40 2.44 36 3.39
95 Mudstone, moderately weathered Very Good Rock Rock Socket 95 22.52 1.00 0.10 2.98 18.29 39 32.47
96 Mudstone, moderately weathered Very Good Rock Rock Socket 95 22.52 2.00 0.10 1.83 24.38 37 29.28
97 Mudstone, moderately weathered Very Good Rock Rock Socket 90 22.52 0.50 0.10 2.27 18.29 38 23.49
98 Mudstone, moderately weathered Very Good Rock Rock Socket 92 22.52 0.30 0.10 2.12 21.34 37 26.74
99 Mudstone, moderately weathered Very Good Rock Rock Socket 90 22.52 0.20 0.10 1.53 3.05 36 10.19
100 Mudstone, moderately weathered Very Good Rock Rock Socket 90 22.52 0.90 0.30 2.19 8.13 38 17.97
101 Mudstone, moderately weathered Very Good Rock Rock Socket 90 22.52 1.50 0.30 1.97 7.11 37 13.58
102 Wilson (1976) Weak clayey mudstone, cretaceous, bedding planes dipping at only a Fair Rock Rock Socket 50 22.52 6.00 0.67 1.09 1.36 45 4.79
few degrees and occasional vertical jointing

113
114 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117

accepted as the production run. After test on testing data set, the n
∑i = 1 (hi −h¯ i )(ti−t¯i )
model is selected based on the fitness value and the maximum R=
n n
∑i = 1 (hi −h¯ i )2 ∑i = 1 (ti−t¯i )2 (3)
correlation coefficient between predicted and measured data. For
the LGP-based analysis, the Discipulus software 49 was used for the
development of each model.
The best LGP programs obtained at the end of training in C þ þ n
∑i = 1 (hi −ti )2
and are given in Appendix A. These programs can be run in C þ þ RMSE =
n (4)
environment. The resulting code may be linked to the optimizer
and compiled or it may be called from the optimization routines
50
. In order to facilitate the use of the derived codes via hand ⎡1 n ⎤
calculations, they are simplified and converted into a functional MAE = ⎢ ∑ hi − ti ⎥
⎢⎣ n ⎥⎦ (5)
representation by successive replacements of variables 36. The i=1

optimal LGP-based formula of qult is as follows, namely qult, LGP :
S S
qult, LGP = 0.064 qu ϕ (2 + RMR RMRqu BγD + γD
B B
A major distinction of LGP for determining the bearing capacity
lies in its powerful ability to model the mechanical behaviour
without requesting a prior form of the existing relationships or
any assumptions like analytical methods. It is worth mentioning
that the statistical regression models, e.g. semi-empirical or em-
pirical models, are mostly obtained after controlling only few
equations established in advance. Thus, these models cannot effi-
ciently consider the interactions between the dependent and in-
dependent variables. On the other hand, LGP introduces com-
pletely new features. The best equation generated by the LGP
technique is determined after controlling numerous linear and
nonlinear preliminary models, e.g., the proposed design equation
is selected among a total of 875,722,182 programs evolved and
evaluated by the LGP method during the conducted 340 runs and
it is not established by chance.
5. Results and discussion
5.1. Performance analysis and validation of LGP-based model
According to Smith's recommendation 51, if a model gives a
correlation coefficient (R)40.8, there is a strong correlation be-
tween the predicted and measured values. In addition, the error
values should be considered in all cases. In this study, to evaluate
the performance of the models correlation coefficient (R), root
mean squared error (RMSE) and mean absolute error (MAE) as
suggested by several researchers. These parameters were calcu-
lated using the following equations:
where, hi and ti are the actual and predicted output values for the
ith output, respectively. hi and ti are, respectively, the average of
the actual and predicted outputs, and n is the number of samples.
(2)
It can be observed from Fig. 4 that the LGP model with high R
and low RMSE and MAE values is able to predict the target values
with an acceptable degree of accuracy. The performance of the
model on the training and testing data suggests that it has both
good predictive abilities and generalization performance. A com-
parison of the measured and predicted qult values by LGP is shown
in Fig. 4.
Besides, new criteria recommended by 52 were checked for the
external validation of the model on the testing data sets 27,38. It is
suggested that at least one slope of regression lines (k or k')
through the origin should be close to 1. Also, the performance
indexes of m and n should be lower than 0.1. Recently, 53 in-
troduced a confirm indicator (Rm) of the external predictability of
models. For Rm 4 0.5, the condition is satisfied. Either the squared
correlation coefficient (through the origin) between predicted and
experimental values (Ro2), or the coefficient between experimental
and predicted values (Ro'2) should be close to R2, and to 1. The
considered validation criteria and the relevant results obtained by
the models are presented in Table 5. As it is seen in this table, the
derived model satisfies all of the required conditions. The valida-
tion phase ensures the derived LGP model is strongly valid.
In order to have an idea about the prediction performance of
the proposed model against a classical model, a comparative study
is conducted. For this aim, the obtained results are compared with
those provided by the following well-known models developed by
12,15
. Besides, a precise way of observing the systematic mis-
matches between the predictions and observations is to consider
the residuals, defined as the residual error (RE) between observed
and predicted values. A comparison of the bearing capacity pre-
dictions made by different models is represented in Fig. 5. Note

80 70
R = 0.96 R = 0.95
Predicted qult (MPa) by LGP

70 RMSE = 5.04
Predicted qult (MPa) by LGP

RMSE = 3.48 60
MAE = 2.39 MAE = 3.1
60 50
50
40
40
30
30
20
20

10 10

0 0
0 20 40 60 80 0 20 40 60 80
Experimental qult (MPa) Experimental qult (MPa)

Fig. 4. Measured versus predicted qult values using the LGP model: (a) training (learning and validation) data, (b) testing data.
 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117 115

Table 5 as quantitative parameters.

Statistical parameters of the LGP model for the external validation.
5.2. Parametric study
Item Formula Condition The LGP
model
In order to ensure the validity of LGP-based model trends, the
1 R 0.8 < R 0.954 behavior of variables should be assessed and compared with those
2 ∑n (hi × ti ) 0.85 < k < 1.15 1.095
k= i=1 provided by other researchers in the literature. For this purpose, a
h2
i comparative parametric analysis is performed. The parametric
3 ∑n (hi × ti ) 0.85 < k′ < 1.15 0.861
k′ = i=1 2 analysis represents the response of the qult the model to the input
t
i variables. The methodology is based on changing only one input
4 R2 − Ro2 m < 0.1 -0.081
m= variable at a time while the other variables are kept constant at the
R2
5 n < 0.1 -0.058
average values of their entire data set. This procedure is repeated
R2 − Ro ′2
n= using another variable until the model response is obtained for all
R2
6 0.5 < Rm 0.664 of the predictor variables 27,38. Fig. 6 shows the tendency of the qult
Rm = R2 (1 − R2 − Ro2 )
predictions to the variations of the influencing parameters, i.e.,
where
Ro2 = 1 −
∑n (ti − h o )2
i=1 i
hio = k × ti
Should be close to 0.985
RMR, γ, qu, D, B, S/B, and ϕ. In order to determine and compare the
1
∑n (ti − t¯i )2 capability of the proposed formula, parametric analysis considered
i=1
∑n (hi − t o )2 Should be close to 0.96 for the LGP-based model and those provided by 12,15. As it is seen
Ro′2 = 1 − i=1 i and o
ti = k′ × h i 1
∑n (hi − h¯i )2 in Fig. 6, the results of the LGP-based model are conform to the
i=1
traditional models and experimental results in the literature
7,9,13,15
.

90 5.3. Sensitivity analysis

80 LGP (This work) Experimental
70 Considering the contribution and relative importance of each
R = 0.95
60 RMSE = 3.85
parameter is another important concern in selecting an input
qult (MPa)

50 MAE = 2.53 parameter for the aim of future experiments, model developments
40 Max | RE | =15.17 or field investigations. Sensitivity analysis aims to describe how
30 much model output values are affected by changes in model input
20 values. Thus, the best thirty programs, obtained by LGP, are eval-
10 uated through a sensitivity analysis and results are summarized in
0 Fig. 7. This is a common approach in the GP-based analyses 27,38,54.
0 10 20 30 40 50 60 70 80 90 100 As represented in Fig. 7, the frequency column represents what
Test No.
percentage of the best thirty programs from the project contained
the related input. It is important to note that the greater the value
90 in frequency, the more impact removal have, i. e. the output var-
80 Carter and Kulhawy (1988) Experimental iation is more dependent on the input variable with more fre-
70 quency. As it can be seen, the LGP model is sensitive to qu (MPa)
R = 0.81 and S/B more than other variables. The results generally conform
60
RMSE = 11.21
qult (MPa)

50 MAE = 6.5 to those noticed by 7,12.

40 Max | RE | =48.95
30
20 6. Conclusion
10
0 This paper aimed at developing a new formulation for indirect
0 10 20 30 40 50 60 70 80 90 100
estimation of the bearing capacity of shallow foundations resting
Test No.
on/in rock masses, using a novel and powerful evolutionary
computational technique, namely linear genetic programming.
250 Unlike other intelligent methods such as ANN, FIS and ANFIS, the
Goodman (1989) Experimental
LGP method provides simplified equations that can be readily used
200
for the design purposes via hand calculating. A comprehensive and
R = 0.88
RMSE = 21.22 reliable database, collected from the literature, is used for the
qult (MPa)

150
MAE = 6.75 model development. Regarding to the available variables in the
Max | RE | =137.02
100 database and traditional equations in the literature, a new im-
proved formulation is obtained that relates the bearing capacity
50 factor (qult) to RMR, γ, qu, D, B, S/B, and ϕ. In order to facilitate the
use of the obtained model, it is converted and represented as a
0 formula. The optimal LGP-based model is selected after several
0 10 20 30 40 50 60 70 80 90 100
Test No.
assessment procedures and validated with different criteria. The
results of the sensitivity and parametric analyses are generally
Fig. 5. A comparison of experimental versus predicted qult values using different expected cases from an engineering viewpoint. These studies are
models.
conducted to evaluate the robustness of the obtained model. The
results indicate that the derived model has a notably better per-
that another major advantage of the proposed model over the formance than traditional and widely used Goodman’s and Carter
other models is that it considers the important effect of rock mass and Kulhawy's models. Moreover, the derived model takes into
classification through using RMR and employing qualitative as well account the important effect of both quantitative and qualitative
116 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117

22
19
20
LGP (This Work) 17
18

qult (MPa)
15
qult (MPa)

16 13
LGP (This Work)
14 11
9
12
0 2 4 6 8
15 35 55 75 95
D (m)
RMR

18.5
21
18.2
LGP (This Work)
19
qult (MPa)

17.9

q ult (MPa)
17
17.6
LGP (This Work)
15
17.3
13
17
0 0.5 1 1.5 2 2.5
17 19 21 23 25 27
B (m)
γ (kN/m3)

200 LGP (This Work) 70

60 LGP (This Work)
Goodman (1989)
150 50 Goodman (1989)
Carter and Kulhawy (1988)
qult (MPa)

40
qult (MPa)

100
30
20
50
10

0 0
0 10 20 30 40 50 0 6 12 18 24 30 36
S/B
qu (MPa)

22

20
qult (MPa)

18

LGP (This Work)

16
Goodman (1989)
14
20 25 30 35 40 45
(Degree)

Fig. 6. A comparative parametric analysis of qult in the LGP and traditional models.

Frequency parameters simultaneously. At last, it should be noted that the

1 predictive capability of the proposed model is mostly limited to
the range of the data used for its calibration and also the for-
0.8
mulation depends on the available variables in the database. To
0.6 cope with these limitations, the model can be easily retrained and
improved to make more accurate predictions for a wider range or
0.4 other variables such as the angle of the load, the slope near the
foundation, the geometric characteristics of joints, etc. by includ-
0.2
ing the data with other different test conditions.
0
RMR γ (kN/m3) qu (MPa) D (m) B (m) S/B φ (Degree)

LGP Model Appendix A. The optimum LGP program for the prediction of
qult, namely “qult, LGP” in the paper
Fig. 7. Contributions of the each variable impact on the LGP model.

The following LGP program can be run in the Discipulus in-

teractive evaluator mode or can be compiled in C þ þ environment.
(Note: v[0], ..., v[6] respectively represent RMR, qu, S/B, and φ. f[0]
 S. Tajeri et al. / International Journal of Rock Mechanics & Mining Sciences 80 (2015) 107–117 117

holds the output.) 20. Padmini D, Ilamparuthi K, Sudheer KP. Ultimate bearing capacity prediction of
float Discipulus C Function(float v[]) shallow foundations on cohesionless soils using neurofuzzy models. Comput
Geotech. 2008;35:33–46.
{ 21. Kuo YL, Jaksa MB, Lyamin AV, Kaggwa WS. ANN-based model for predicting the
long double f[8]; bearing capacity of strip footing on multi-layered cohesive soil. Comput Geotech.
long double tmp ¼ 0; 2009;36:503–516.
22. Kolay E, Kayabali K, Tasdemir Y. Modeling the slake durability index using
intcflag ¼ 0; regression analysis, artificial neural networks and adaptive neuro-fuzzy
f[0] ¼f[1] ¼f[2] ¼f[3] ¼ f[4]¼ f[5] ¼f[6] ¼f[7] ¼0; methods. Bull Eng Geol Environ. 2010;69:275–286.
L0:f[0] þ ¼v[2]; 23. Alavi AH, Gandomi AH. A robust data mining approach for formulation of
geotechnical engineering systems. Eng Comput. 2011;28:242–274.
L1:f[0]* ¼v[0]; 24. Sattari MT, Apaydin H, Ozturk F. Flow estimations for the Sohu Stream using
L2:f[0]* ¼v[3]; artificial neural networks. Environ Earth Sci. 2012;66:2031–2045.
L3:f[0]* ¼v[1]; 25. Tasdemir Y, Kolay E, Kayabali K. Comparison of three artificial neural network
approaches for estimating of slake durability index. Environ Earth Sci.
L4:f[0]* ¼v[4]; 2013;68:23–31.
L5:f[0] þ ¼v[1]; 26. Ceryan N, Okkan U, Kesimal A. Prediction of unconfined compressive strength
L6:f[0]* ¼v[0]; of carbonate rocks using artificial neural networks. Environ Earth Sci.
2013;68:807–819.
L7:f[0]* ¼v[0]; 27. Sadrossadat E, Soltani F, Mousavi SM, Marandi SM, Alavi AH. A new design
L8:f[0] ¼sqrt(f[0]); equation for prediction of ultimate bearing capacity of shallow foundation on
L9:f[1] þ ¼v[1]; granular soils. J Civ Eng Manag. 2013;19:S78–S90.
28. Manouchehrian A, Gholamnejad J, Sharifzadeh M. Development of a model for
L10:f[1]* ¼v[3]; analysis of slope stability for circular mode failure using genetic algorithm.
L11:f[0] þ ¼f[1]; Environ Earth Sci. 2014;71:1267–1277.
L12:f[0]¼ sqrt(f[0]); 29. Ziaee SA, Sadrossadat E, Alavi A, Mohammadzadeh Shadmehri D. Explicit for-
mulation of bearing capacity of shallow foundations on rock masses using ar-
L13:f[0] þ ¼v[5]; tificial neural networks: application and supplementary studies. Environ Earth
L14:f[0] þ ¼v[5]; Sci. 2014:1–15.
L15:f[0]* ¼v[5]; 30. Banzhaf W, Nordin P, Keller R, Francone F. Genetic programming: an introduction
on the automatic evolution of computer programs and its application. San Fran-
L16:f[0]* ¼v[2]; cisco: Morgan Kaufmann; 1998.
L17:f[0]* ¼ v[6]; 31. Brameier M, Banzhaf W. Linear genetic programming. New York: Springer; 2007.
32. Koza JR. Genetic programming, on the programming of computers by means of
L18:f[0] ¼sqrt(f[0]);
natural selection. Cambridge: Mass: MIT Press; 1992.
L19:f[0]* ¼0.064; 33. Coello Coello C, Lamont GB, Van Veldhuizen DA. Evolutionary algorithms for
L20: solving multi-objective problems. 2nd ed,New York: Springer; 2007.
34. M Schmidt, H Lipson. Comparison of Tree and Graph Encodings as Function of
if (!_finite(f[0])) f[0]¼ 0;
Problem Complexity GECCO '07. In: Proceedings of the 9th annual conference
return f[0]; on genetic and evolutionary computation, New York: 2007: 1674–1679.
} 35. Brameier M, Banzhaf W. A comparison of linear genetic programming and
neural networks in medical data mining. IEEE Trans Evol Comp. 2001;5:17–26.
36. Oltean M, Grosan C. A comparison of several linear genetic programming
techniques. Complex Syst. 2003;14:1–29.
References 37. Francone FD, Deschaine LM. Extending the boundaries of design optimization
by integrating fast optimization techniques with machine–code–based, linear
genetic programming. Info Sci. 2004;161:99–120.
1. EE De Beer. Bearing capacity and settlement of shallow foundations on sand. In: 38. Alavi AH, Ameri M, Gandomi AH, Mirzahosseini MR. Formulation of flow
Proceedings of symposium on bearing capacity and settlement of shallow number of asphalt mixes using a hybrid computational method. Constr Build
foundations, Duke University Durham, NC: 15–33. Maters. 2011;25:1338–1355.
2. Das BM. Principles of foundation engineering. 3rd edition,Boston: PWS Publish- 39. Gandomi A, Alavi A, Sahab M. New formulation for compressive strength of
ing; 1995. CFRP confined concrete cylinders using linear genetic programming. Mater
3. Chen W, Duan L. Bridge engineering handbook. New York: CRC Press; 1999. Struct. 2010;43:963–983.
4. Lo KY, Hefny AM. Foundations on rock. In: Rowe RK, editor. Geotechnical and 40. Hoek E, Brown ET. The Hoek–Brown failure criterion-a 1988 update. 15th Cana-
geoenvironmental engineering handbook. Berlin Heidelberg: Springer; 2001. dian Rock Mechanics Symposium. University of Toronto; 1988.
p. 305–332. 41. Bieniawski ZT. Determining rock mass deformability. Int. J. Rock Mech. Min. Sci..
5. Sowers GF. Introductory soil mechanics and foundations: geotechnical engineer- 1978;15:335–343.
ing. 4th ed,New York: Macmillan; 1979. 42. Bieniawski ZT. Engineering rock mass classifications: a complete manual for en-
6. Bishoni BL. Bearing capacity of a closely jointed rock [PhD dissertation]. United gineers and geologists in mining, civil, and petroleum engineering. New York:
States: Georgia Institute of Technology; 1968. Wiley; 1989.
7. Paikowsky S, Cannif M, Lensy K, Aloys K, Amatya S, Muganga R. NCHRP design 43. Hirany A, Kulhawy FH. Conduct and interpretation of load tests on drilled shaft
and construction of shallow foundations for highway bridge structures. Wa- foundations: detailed guidelines. Palo Alto: Electric Power Research Institute;
shington, DC: Transportation Research Board of the National Academies; 2010 1988.
report 651: LRFD. 44. Prakoso WA. Reliability-based design of foundations on rock masses for trans-
8. DE Becker, D MooreI. Canadian Foundation Engineering Manual. 4th ed. Ca-
mission line and similar structures. Cornell University; 2002.
nadian Geotechnical Society; 2006.
45. Muganga RK. Uncertainty evaluation of displacement and capacity of shallow
9. AASHTO. LRFD bridge design specifications. 4th ed,Washington, DC: AASHTO;
foundation on rock [MSc Thesis]. Lowell: University of Massachusetts; 2008.
2007.
46. Yılmaz I, Yuksek AG. An example of artificial neural network (ANN) application
10. Look BG. Handbook of Geotechnical Investigation and Design Tables. London:
for indirect estimation of rock parameters. Rock Mech Rock Eng. 2008;41:781–
Taylor and Francis; 2007.
795.
11. Terzaghi K. Rock defects and loads on tunnel supports. In: Proctor RV, White TL,
47. FD Francone. Discipulus Owner's Manual. Littleton. Register Machine Learning
editors. Rock tunneling with steel supports. Youngstown, OH: Commercial
Shearing and Stamping Co; 1946. p. 17–99. Technologies, Inc., Colorado, USA: 1998.
12. Goodman RE. Introduction to rock mechanics. 2nd ed,New York: Wiley; 1989. 48. Baykasoğlu A, Güllü H, Çanakçı H, Özbakır L. Prediction of compressive and
13. Bowles JE. Foundation analysis and design. 5th ed,New York: McGraw-Hill; 1996. tensile strength of limestone via genetic programming. Expert Syst Applic.
14. Hoek E, Brown ET. Underground excavations in rock. London: The Institution of 2008;35:111–123.
Mining and Metallurgy; 1980. 49. Conrads M, Dolezal O, Francone FD, Nordin P. Discipulus-fast genetic program-
15. Carter P, Kulhawy FH. Analysis and design of foundation socketed into rock. New ming based on AIM learning technology. Littleton, CO: Register Machine Learning
York: Empire State Electric Engineering Research Corporation and Electric Technologies Inc.; 2004.
Power Research Institute; 1988 report no. EL-5918. 50. Deschaine LM, Patel JJ, Guthrie RD, Grumski JT. Using genetic programming to
16. Peck RB, Hanson WF, Thornburn TH. Foundation engineering. 2nd ed,New York: develop a C/C++ simulation model of a waste incinerator. Seattle, WA: Society for
Wiley; 1974. Computer Simulation's Advanced Technology Simulation Conference; 2001.
17. FH Kulhawy, RE Goodman. Design of foundations on discontinuous rock. In: 51. Smith GN. Probability and statistics in civil engineering. London: Collins; 1986.
Proceedings of the International Conference on Structural Foundations on Rock: 52. Golbraikh A, Tropsha A. Beware of q2!. J Molec Graph Modell. 2002;20:269–276.
International Society for Rock Mechanics; 1980: 209–220. 53. Roy PP, Roy K. On some aspects of variable selection for partial least squares
18. Zhang L, Einstein H. End bearing capacity of drilled shafts in rock. J Geotech regression models. QSAR Comb Sci. 2008;27:302–313.
Geoenviron Eng. 1998;124:574–584. 54. Gandomi AH, Alavi AH, Mirzahosseini R, Moghdas Nejad F. Nonlinear genetic-
19. Shahin M, Jaksa M. Neural networks prediction of pullout capacity of marquee based models for prediction of flow number of asphalt mixtures. J Maters Civ
ground anchors. Comput Geotech. 2005;32:153–163. Eng. 2011;23:248–263.