Conference PaperPDF Available

Predicting the Side Resistance of Piles Using a Genetic Algorithm and SPT N-Values

  • Toronto Metropolitan University

Abstract and Figures

A genetic algorithm (GA) was developed to predict the side resistance of piles in Ontario with the standard penetration test blowcounts (SPT N-values). Pile foundations commonly support structures by transferring loads deeper into the ground; unfortunately, the soil characteristics, such as the strength and grain-size distribution, are rarely consistent at a site. Due to the spatial variability of the soil, challenges arise to accurately predict the ultimate axial capacity of piles. This research aims to mitigate this problem by implementing genetic programming. Since the 1950s, the Ministry of Transportation of Ontario has gathered approximately 100 pile load tests in various soil conditions, and a total of 23 piles were selected for this study. These piles were either H piles or pipe piles and were subjected to extension load tests. A GA was created to correlate the side resistances to the SPT N-values, and the developed relationships were compared to existing design methods for different soil types. In all, the ultimate goal of this research was to improve local pile design in Ontario's practice.
Content may be subject to copyright.
Predicting the Side Resistance of Piles Using a
Genetic Algorithm and SPT N-Values
Markus Jesswein & Jinyuan Liu
Department of Civil Engineering Ryerson University, Toronto, ON, Canada
Minkyung Kwak
Ministry of Transportation of Ontario, Toronto, ON, Canada
A genetic algorithm (GA) was developed to predict the side resistance of piles in Ontario with the standard penetration test
blowcounts (SPT N-values). Pile foundations commonly support structures by transferring loads deeper into the ground;
unfortunately, the soil characteristics, such as the strength and grain-size distribution, are rarely consistent at a site. Due
to the spatial variability of the soil, challenges arise to accurately predict the ultimate axial capacity of piles. This research
aims to mitigate this problem by implementing genetic programming. Since the 1950s, the Ministry of Transportation of
Ontario has gathered approximately 100 pile load tests in various soil conditions, and a total of 23 piles were selected for
this study. These piles were either H piles or pipe piles and were subjected to extension load tests. A GA was created to
correlate the side resistances to the SPT N-values, and the developed relationships were compared to existing design
methods for different soil types. In all, the ultimate goal of this research was to improve local pile design in Ontario’s
Un algorithme génétique (AG) a été développé pour prédire la résistance au cisaillement des pieux en Ontario avec les
compteurs d'essai de pénétration standard (valeurs N du SPT). Les fondations de pieux soutiennent généralement les
structures en transférant les charges plus profondément dans le sol; Malheureusement, les caractéristiques du sol, telles
que la force et la distribution granulométrique, sont rarement constantes sur un site. En raison de la variabilité spatiale du
sol, il est difficile de prédire avec précision la capacité axiale ultime des piles. Cette recherche vise à atténuer ce problème
en mettant en œuvre la programmation génétique. Depuis les années 1950, le ministère des Transports de l'Ontario a
recueilli environ 100 essais de chargement de pieux dans diverses conditions de sol, et un total de 23 piles ont été
sélectionnées pour cette étude. Ces piles étaient des piles H ou des pieux tubulaires et ont été soumises à des essais de
charge d'extension. Un AG a été créé pour corréler les résistances au cisaillement aux valeurs N du SPT, et les relations
développées ont été comparées aux méthodes de conception existantes pour différents types de sols. Dans l'ensemble,
le but ultime de cette recherche était d'améliorer la conception locale des piles dans la pratique ontarienne.
Deep foundations are designed to support bridges and
buildings, but accurately predicting the capacity of a pile is
a challenge due to the influences from the installation
method, pile geometry, and soil properties. Especially
since glacial tills are commonly found in Ontario, the site
conditions are rarely consistent as the soil strength and
contents can vary spatially. In addition, the standard
penetration test (SPT) is unreliable because it lacks
accuracy. Yet, it is commonly used in site investigations
because it is a cheap and simple measurement technique
and can be applied in various soil conditions, including
gravel and cobble rich soils. Numerous correlations have
been proposed to predict the pile capacity with SPT N-
values, but many of the design methods were developed
by averaging or generalizing the soil conditions. Also, a
limited number of design methods have been proposed for
soils with stiff clays and glacial tills, especially for Ontario.
This investigation addresses the uncertainty in design
by developing a genetic algorithm (GA) to correlate SPT N-
values to the side resistance of driven piles. The Ministry
of Transportation of Ontario (MTO) accumulated a
database with over 100 pile load tests, and results from 23
high-quality extension load tests were selected for this
study. GAs can efficiently correlate several variables
compared to traditional statistical approaches, and a GA
was developed to predict the frictional resistance with
multiple N-values and soil types along the length of a pile.
2.1 Current Design Methods
A pile subjected to axial compressive loading can
experience two mechanisms: the side resistance () and
the tip resistance (). The side resistance is the friction
between the soil and pile walls, while the tip resistance is
developed by the strength of the soil at the pile base. As a
pile reaches its maximum load, these mechanisms
contribute towards the ultimate pile resistance, or capacity
   [1]
In this paper, the focus is on the side resistance, which
depends on the unit side resistance (), pile perimeter ()
and embedment length ().
   [2]
Usually, the side resistance is directly predicted by the
average SPT N-value (
) along the pile length. As shown
in Table 1, many empirical correlations have been
proposed for various soil conditions. Unfortunately, the
average soil conditions may not accurately represent the
soil behaviour.
Table 1. Existing Design Methods for Pile Side Resistance
Soil Type
Equation for (kPa)
Shioi & Fukui
 
Meyerhof (1956)
Pipe Piles: 
H Piles: 
Shioi & Fukui
 
Brown (2001)1
 
Decourt (1982)2
 
   
1 Recommended SPT range is      and 2     .
Several references also report various influences on
the capacity with the pile length (Meyerhof 1976; Vesic
1977; Poulos et al. 2001). This relationship can be due to
load dissipation along the pile length; events related to the
pile installation process, such as whipping; or the
confinement of the effective stress with noncohesive soils.
In all, the side resistance is influenced by many
variables, including the soil type, soil strength, pile
geometry, and installation process. The complex and
nonlinear relationship between the pile and soil can be
predicted with a GA.
2.2 An Introduction to Genetic Algorithms
Genetic algorithms are an optimization approach inspired
by Darwin’s theory of evolution (Banzhaf et al. 1998). In
nature, chromosomes give an organism its attributes to
survive and succeed in an environment. Through
reproduction, organisms can adapt and evolve to their
environment. A GA represents the problem domain as a
chromosome. In this case, a GA was developed to conduct
symbolic regression and predict the side resistance. For
symbolic regression, the genes of a chromosome
represented the components of a function: a variable,
constant, or operator.
A generic GA searches for a solution through five
general steps: chromosome creation, evaluation, selection,
crossover, and mutation. First, multiple attempts for a
problem are made at once in a trial, or generation, by
generating a population of chromosomes with different
attributes. The performance or fitness of a single
chromosome is measured by an objective function. From
the population of chromosomes, potential parents are
selected for the creation of offspring. Typically, during
selection, a preference is given to chromosomes with a
higher fitness. The population size remains constant
throughout every generation, and the previous
chromosomes, or at least a majority, are replaced by new
offspring. The population then evolves through several
generations by reproduction mechanisms, such as
crossover and mutation.
A GA is a stochastic method. If a regression analysis
was repeated with multiple trials and had the same initial
conditions, the GA can complete the analysis with a similar
level of fitness but provide a different solution. A GA is also
data-driven and, depending on the complexity of a
problem, may not find an exact solution or global minimum,
but it is an efficient optimization approach, especially if
aspects of a problem are unknown.
3.1 Overview of the Methodology
For driven piles in heterogeneous soils, the goal was to
improve the predictability of the side resistance. A new
design method was developed by following four steps. (1)
Results from pile load tests and soil measurements were
collected from a database by MTO. (2) For every pile, the
measured side resistance was obtained from load-
displacement responses that were measured at the pile
top. (3) A GA was then developed to nonlinearly correlate
the side resistance in heterogeneous soils with SPT N-
values. (4) In the end, the accuracy of the GA was
compared to existing design methods.
3.2 Testing Sites and Pile Load Tests
For this investigation, borehole logs for the soil conditions
and records on the pile load tests were collected from
MTO. Driven piles were either H piles or steel pipe piles,
and they were subjected to static axial-tension loads to
measure the side resistance. The pile properties and
Figure 1. Location of Studied Sites
Table 2. Details on the Studied Piles
Pile Type1
Embedded Soil
324 OD Pipe
Clayey Silt
324 OD Pipe
Clayey Silt
324 OD Pipe
Silty Clay
HP 310x110
Silty Clay
324 OD Pipe
324 OD Pipe
HP 310x79
HP 310x79
HP 310x110
Layered Clayey Silt
and Silty Sand
324 OD Pipe
Layered Clayey Silt
and Silty Sand
HP 310x110
Layered Clayey Silt
and Silty Sand
HP 310x79
Sand to Silty Sand
HP 310x79
Sand to Silty Sand
HP 310x79
Sand to Sandy Silt
HP 310x110
Sand to Silty Sand
HP 310x110
Sand to Silty Sand
HP 310x110
Sand to Silty Sand
HP 310x110
Silty Sand; Layered
Clay and Silt
324 OD Pipe
Silty Sand; Layered
Clay and Silt
HP 310x110
Layered Sand and
Silty Clay
324 OD Pipe
Sandy Silt to Sand
HP 310x110
324 OD Pipe
1 Steel H pile designations are depth (mm) by weight (kg/m). Steel
pipe piles were filled with concrete before testing, and OD is the
outside diameter (mm); 2 Embedment Length; 3 Classifications
according to MTO standards.
ultimate side resistances are in Table 2. From the 23 piles,
9 were pipe piles, and 14 were H piles. The embedment
lengths varied from 3 m to 45 m, but most of the piles were
between 12 to 25 m long.
Figure 1 shows the locations of the sites. The soils were
generally compact or stiff, but some loose sands were
found. Noncohesive soils were common in the database.
For the pipe piles, two were mainly in cohesive soils, four
dominated in noncohesive soils, and the remaining had
mixed soil conditions. One H pile was fully embedded in
cohesive soils, while nine H piles were in noncohesive
soils. Borehole logs contained the soil type, SPT N-values,
and occasionally, the unit weights at the sites. SPT N-
values were corrected according to CGS (2006) for a
hammer efficiency of 60% and the overburden conditions.
3.3 Measured Side Resistance ()
The load-displacement response was measured with dial
gauges at the top of the piles during the load tests. The
failure load was determined by the criteria from De Beer
(Fellenius 1980). De Beer recommends plotting both axes
of the load-displacement curves with a log-scale, and the
failure load is indicated by the largest change in slope on
the plot (Fellenius 1980).
3.4 Construction of the Genetic Algorithm
3.4.1 Introduction
Each pile was divided into 50 segments to consider the
varying side resistance along their length. The variables in
the analysis included the corrected SPT N-values (), the
soil type (), effective stress (), and pile slenderness ratio
(  ). N-values were corrected for the hammer
efficiency () for cohesive soils, and the overburden
correction was also applied () for noncohesive soils.
The soil type was a binary variable equal to 1 for
noncohesive soils or 2 for cohesive soils. The slenderness
ratio was modified to determine the side resistance at any
depth. It was composed of the embedment length (),
depth to the centre of a pile segment (), and maximum
width or diameter of a pile (). The side resistance of a H
pile will be influenced if the soil creates an unplugged, fully
plugged, or partially plugged condition. If a pile is assumed
to be fully plugged, illogical results may be provided,
especially for noncohesive soils. Since the actual perimeter
of the pile is known, H piles were assumed to be unplugged
for the analysis. The database was divided for H piles and
pipe piles, and the GA performed 5 trials for each pile type
(a total of 10 runs) to regress the variables and test results.
Figure 2. Process of the Genetic Algorithm
As displayed in Figure 2, the GA in this investigation
evolved the chromosomes with the following steps:
creation, evaluation, selection, crossover, mutation, and
constant refinement. The GA was created with Matlab
(Mathworks 2017) and applies the Multi Expression
Programming (MEP) technique (Oltean & Dumitrescu,
2002) to encode and evaluate the chromosomes. MEP was
based on the activation of programs or code with integers
and can efficiently encode or decode functions compared
to other techniques (Oltean & Dumitrescu, 2002). Table 3
Table 3. Settings for the Genetic Algorithm
Parameter Setting
Number of generations
Population size
Function set
  power, exponential,
logarithmic, hyperbolic tangent
Chromosome length
Fitness function
Mean Squared Error (MSE)
Mutation rate (%)
Crossover rate (%)
Crossover type
Uniform with brood
Population size for brood
Brood crossover rate (%)
Population size for brood
constant refinement
Tournament selection size
Initial operator likelihood (%)
Initial variable likelihood (%)
Initial constant likelihood (%)
shows the settings of the GA.
3.4.2 Creating Chromosomes
For a given problem, a potential solution is represented by
a chromosome. In this investigation, the chromosomes with
MEP were linear entities, or arrays, and represented a
function for the unit side resistance of a pile. The genes, or
entries within the arrays, were divided into two
components. The first part indicated the activation of a
function component, such as a variable, constant, or
operator. The second part links the action of the operators.
Figure 3 shows an example to decode an MEP
chromosome. The first row of the chromosome in the figure
has negative integers to represent the operators. In this
example, -1 is for addition. Positive integers designated the
activation of variables and constants, which are
represented in general by and in the figure. A constant
or variable must be the first entry within the chromosome
to prevent illogical errors during evaluation (Oltean &
Dumitrescu, 2002). The last two rows indicate the
locations, or column numbers, for the operators to be
performed. During evaluation, the result of every gene is
stored, and the operators are applied to the results from
Figure 3. Example of Decoding the Chromosome
previous portions of the chromosome. Since variables and
constants are numerical values, their corresponding links
in the chromosome are meaningless.
The GA in this investigation was capable of simple
arithmetic, but it also included power, logarithmic, and
hyperbolic tangent operators because they represent
nonlinear relations. However, the possible combinations
for a GA to search can also increase exponentially if many
operators are added (Banzhaf et al. 1998).
The population of chromosomes were initially created
randomly, but a probability was assigned for the likelihood
of occurrence for the operators, constants, and variables.
The chromosomes were given a maximum length of 20
3.4.3 Evaluation of the Chromosomes
The goal of the GA in this study was to find the function
with the best fitness to represent the unit side resistance of
a pile. During evaluation, the changing shear strength from
the soil conditions were considered by dividing the pile into
several segments. The GA predicted the unit side
resistance for each layer, and the total side resistance of a
pile was found by the summation of the side resistances on
the pile segments. Load-displacement responses were
measured at the top of the piles during testing. Thus, the
fitness function compared the predicted total side
resistance () to the measured side resistance ()
with the mean squared error (MSE):
 
   [3]
Where n is the number of analyzed piles. A lower MSE
indicates a better fit between the measured and predicted
Generally, care is needed to ensure that illogical errors
do not occur during evaluation. Examples include dividing
by zero or taking the logarithm of a negative value. Division
operators may be protected by simply returning the
numerator if a denominator of zero is found (Banzhaf et al.
1998), but Oltean & Dumitrescu (2002) recommend
mutating division into a variable or constant. Other
operators were protected and transformed as suggested by
Brameier & Banzhaf (2007).
3.4.5 Selection of the Parents
A pair of parents were selected for mating using
tournament selection. For each parent, a number, or
tournament size, of chromosomes were randomly
sampled, and the chromosome from this group with the
highest fitness became a parent. Tournament selection
was repeated until the new population size matched the
original size.
3.4.6 Crossover and Mutation
The activation of crossover was assigned a probability. If
crossover was chosen to not occur, the parents were
copied and sent for mutation. Otherwise, uniform crossover
was applied with brood recombination (Figure 4). Uniform
crossover randomly distributes the genes from the parents
Figure 4. Example of Uniform Crossover with Brood
to the offspring. Brood recombination was inspired by
organisms having a litter of offspring (Tackett 1994), and it
attempts to extract the best attributes from two parents.
Two parents performed crossover several times to create
a subpopulation () of new chromosomes. The two
offspring with the best fitness continued for mutation.
Ifevery pair of parents experienced brood recombination,
the total number of created offspring would be multiplied
by the population size, and the computational effort would
be significantly increased (Banzhaf et al. 1998). Thus, the
chance of brood recombination was assigned a probability.
For mutation, a selected gene would be transformed
randomly into a different component type. For example, an
addition operator could become a constant. The chance of
mutation was set to be low at 10 %. After mutation, the
resulting offspring replaced the worst chromosomes from
the original population if they developed a higher fitness.
3.4.7 Constant Refinement
For symbolic regression, the constants are either
evolutional or non-evolutional (Banzhaf et al. 1998). Non-
evolutional constants are kept the same throughout a
generation, but the GA can apply operations to manipulate
the value of a constant within a function. For evolutionary
constants, optimization techniques, such as the
Levenberg-Marquardt algorithm (Marquardt 1963) or
Nelder-Mead simplex method (Nelder and Mead 1965),
can be applied. These methods are mathematically
complex and usually iterative. They may take numerous
trials to terminate on a potential solution, especially with
several variables. Since the population of chromosomes
may be large, the computational effort should be
minimized. Brood recombination was applied in this GA as
a simple approach to refine the values of the constants. For
every chromosome, the values of the constants were
randomly changed for 50 attempts. The values that
provided the best fitness were kept as the new constants.
3.5 Results from the Genetic Algorithm
The GA analyzed both pipe piles and H piles separately
with 2000 chromosomes for 60 generations. The plots in
Figures 5 and 6 show the average and lowest MSE within
the population of chromosomes for pipe and H piles. For
Figure 5. Fitness Performance with Pipe Piles
Figure 6. Fitness Performance with H Piles
each of the 5 trials, the analysis usually terminated with a
similar MSE. As a better links were made, the brood
recombination during crossover and constant refinement
resulted in sudden drops in the best fitness throughout the
At the end of the 5 trials, the function with the best
fitness (BF) was collected for each pile type. The MSE by
itself may seem misleading since the pipe piles had lower
side resistances on average than the H piles. Yet, as
shown in Figures 7 and 8, the best-fit function for pipe and
H piles had a reasonably good R2, and results were also
mainly within ± 25 % of the 1:1 line. For pipe piles, the
function with the lowest MSE had a R2 of 0.73:
      [4]
The function with the best fit for unplugged H piles had a
R2 of 0.82:
  [5]
010 20 30 40 50 60
Fitness (Mean Squared Error)
Trial 1 Average
Trial 1 Best
Trial 2 Average
Trial 2 Best
Trial 3 Average
Trial 3 Best
Trial 4 Average
Trial 4 Best
Trial 5 Average
Trial 5 Best
010 20 30 40 50 60
Fitness (Mean Squared Error)
Trial 1 Average
Trial 1 Best
Trial 2 Average
Trial 2 Best
Trial 3 Average
Trial 3 Best
Trial 4 Average
Trial 4 Best
Trial 5 Average
Trial 5 Best
For pipe piles, the developed function had a predicted
to measured resistance ratio ( ) of 1.09 on
average, and it overestimated the resistance of the piles
dominating in cohesive soils. The overestimation likely
resulted since a limited number of piles were fully
embedded in clay. In general, the GA rarely considered the
soil type for H piles because noncohesive soils dominated
the sites. While observing the applied variables in the final
generation, functions frequently contained the slenderness
ratio for H piles, and Equation 5 does not apply any other
variable. This result may indicate the unreliability of SPT or
the soil plugging and installation effects of H piles. Equation
5 typically underestimated the side resistance with an
average  of 0.94. Equations 4 and 5 may also tend
to be more conservative for piles with higher resistances or
longer lengths.
The function with the lowest fitness, as demonstrated
Figure 7. Comparison of Measured and Predicted Side
Resistances by GA for Pipe Piles
Figure 8. Comparison of Measured and Predicted Side
Resistances by GA for H Piles
with Equation 5, is not always practical, beneficial, or
appropriate. Another function for each pile type was then
selected by Pareto optimization (PO). The results from the
final generations of the 5 trials were pooled together to
create a population of 10000 functions. These functions
were then graphically evaluated by their fitness and
complexity. The complexity is the number of components
in a function, and the Pareto front was created in Figures 9
and 10 by finding the best fitness for each complexity. In
general, a lower complexity, or a shorter function, results in
a higher MSE, but a longer function can have several
operations to create a better fitness. The orange square
markers are points on the Pareto front, and the blue circles
are the remaining results. Any point along the Pareto front
can be a potential solution; thus, the preferred solution
mainly relies on the tolerable error and judgement of the
investigator (Smits & Kotanchek 2005).
For pipe piles, the improvement of the MSE is low for a
complexity between 9 to 15. The function on the Pareto
front with 9 components was selected since shorter
functions had a significantly higher MSE. The
corresponding function is below (R2 = 0.69):
      [6]
The Pareto front in Figure 10 was linear for H piles, and
the sudden increase in fitness at a complexity of 15 may be
due to the volatile nature of brood recombination. A
preference was given to a function containing several
variables. The selected function initially had a complexity
of 9 but was simplified to the following (R2 = 0.76):
        [7]
The fitness of Equations 6 and 7 is displayed in Figures 7
and 8. The functions from the Pareto optimization have a
small difference in R2 compared to the functions with the
lowest MSE. Since Equation 6 included the soil type, it
bears more information on the soil conditions than
Equation 4, and it had a slightly better average  of
1.06. Equations 7 and 5 did not include the soil type, and
Equation 7 tends to overestimate compared to Equation 5
with an average  of 1.16. The effective stress was
Figure 9. Pareto Front from GA results for Pipe Piles
BF: y = 0.75x + 96
R² = 0.73
PO: y = 0.80x + 85
R² = 0.69
0 200 400 600 800 1000
Predicted Total Side Resistance (kN)
Measured Total Side Resistance (kN)
Best Fit Function
Pareto Evaluated Function
BF: y = 0.85x + 62
R² = 0.82
PO: y = 0.72x + 231
R² = 0.76
0 500 1000 1500 2000
Predicted Total Side Resistance (kN)
Measured Total Side Resistance (kN)
Best Fit Function
Pareto Evaluated Function
Pareto Front
Figure 10. Pareto Front from GA results for H Piles
not included in any of the functions from the GA.
3.6 Performance of Existing Design Methods
The side resistances of the piles were calculated with
design methods that were intended for both cohesive and
noncohesive soils: Shioi and Fukui (1982), Decourt (1982),
and Brown (2001). N-values were corrected and limited as
mentioned by the references, and H piles were assumed to
be fully plugged as suggested by Brown (2001). The results
of the predictions are provided in Figures 11 to 13.
The three existing design methods mainly
overestimated the side resistance and gave erratic results.
Especially for the pipe piles, a logical linear relationship
with a decent fitness could not be established between the
measured and predicted values. The approach by Brown
(2001) had the worst performance with an average
 of 2.51 and 2.97 for pipe and H piles, respectively.
The method by Decourt (1982) overestimated the side
resistance by 2.40 times on average for both pile types,
Figure 11. Comparison of Measured and Predicted Side
Resistances by Shioi and Fukui (1982)
Figure 12. Comparison of Measured and Predicted Side
Resistances by Decourt (1982)
Figure 13. Comparison of Measured and Predicted Side
Resistances by Brown (2001)
and it gave the best results among the existing methods.
The greatest over predictions occurred with piles in clays
or very stiff soils.
This preliminary investigation demonstrated the capability
of a simple GA to predict the side resistance of 23 piles with
SPT N-values. Although a small sample size was analyzed,
the GA was given more detail on the soil measurements by
dividing the piles into segments. This GA was then tailored
to consider heterogenous soil conditions, and the
correlated functions were refined with Pareto optimization.
For both pipe and H piles, a function was initially
selected from two different criteria: the best fitness and
Pareto optimization. Equations 6 and 7 were determined
HP: y = 1.14x + 538
R² = 0.48
0 1000 2000 3000 4000
Predicted Total Side Resistance (kN)
Total Measured Side Resistance (kN)
Pipe Pile
H Pile
HP: y = 1.23x + 604
R² = 0.44
0 1000 2000 3000 4000
Predicted Total Side Resistance (kN)
Total Measured Side Resistance (kN)
Pipe Pile
H Pile
PP: Trend?
HP: y = 1.50x + 759
R² = 0.44
0 1000 2000 3000 4000
Predicted Total Side Resistance (kN)
Total Measured Side Resistance (kN)
Pipe Pile
H Pile
Pareto Front
PP: Trend?
PP: Trend?
from the Pareto evaluation, and they are recommended
over the functions from the best fitness. Thus, the following
equation is proposed for pipe piles (R2 = 0.69):
        [8]
The function below is suggested for unplugged H piles (R2
= 0.76):
          [9]
For these two functions, it is suggested, like Meyerhof
(1976), to limit the unit side resistance to 100 kPa. From
the studied piles, the measured unit side resistance did not
surpass this value.
Both Equations 8 and 9 were directly proportionate to
the SPT N-values and indicate a higher unit side resistance
with stiffer soils. They also apply the inverse of the
slenderness ratio. This variable was commonly applied by
the GA, and it can indicate in the equations that the soil
disturbance is lower towards the pile base. The side
resistance could also be higher towards the pile base
during pull-out because every pile had an over-sized base
plate or reinforcement base plate. Since the sites
dominated in the noncohesive soils, the common use of the
slenderness ratio could also indicate the influence of the
effective stress, but it is difficult to evaluate without results
from fully instrumented piles.
The results from the GA were more accurate compared
to the existing design methods. Yet, the existing design
methods solely relied on the SPT N-values and were
intended for weaker soils. Cohesive soils were the main
cause of overestimation, but Equation 6 from the GA
assumes cohesive soils have a lower side resistance than
noncohesive soils. Shioi and Fukui (1982) received the
opposite result. This investigation did not have many piles
in stiff undrained clays; thus, Equations 6 and 7 may be
more appropriate for noncohesive soils and drained clays.
Although the findings heavily rely on the extent of the
site investigations and pile load tests, the performance of
the existing methods demonstrates a need in Ontario for
locally-developed design methods for the pile capacity. The
GA gained practical functions with multiple variables and
soil measurements along the piles. It can likely provide
more accurate results with advanced soil testing, such as
the cone penetration test, or data from fully instrumented
piles load tests. The GA was also efficient at considering
nonlinear relationships, which would be difficult to achieve
with traditional statistics. In all, machine learning
techniques can help address uncertainty in geotechnical
engineering and offer better design methods for future
infrastructure projects in Ontario.
The presented research was made possible with funding
and a master’s scholarship from the National Sciences and
Engineering Research Council of Canada and support and
resources provided by the Ministry of Transportation of
Ontario. The authors would like to thank Mr. David Staseff
from MTO for sharing the database of pile load tests. The
authors also appreciate the help from Mr. Andy Lai, Mr.
Filipe Batista, Ms. Chantel Yung, Ms. Maeeda Khan, and
Ms. Maribel Castro with the collection of data.
Banzhaf, W., Nordin, P., Keller, R.E., and Francone, F.D.
1998. Genetic programming: an introduction, Morgan
Kaufmann Publishers Inc., San Francisco, CA, USA.
Brameier, M. and Banzhaf, W. 2007. Linear genetic
programming, Springer, New York, NY, USA.
Brown, R. P. 2001. Predicting the ultimate axial resistance
of single driven piles (PhD dissertation), University of
Texas, Austin, TX, USA.
Canadian Geotechnical Society (CGS). 2006. Canadian
Foundation Engineering Manual, 4th ed., CGS,
Richmond, BC, Canada.
Decourt, L. 1982. Predictions of bearing capacity based
exclusively on N values of the SPT, In Proceedings of
the 2nd European Symposium on Penetration Testing,
Amsterdam, 1: 29-34.
Fellenius, B. H. 1980. The analysis of results from routine
pile load tests, Ground engineering, 13(6): 19-31.
Marquardt, D.W. 1963. An algorithm for least-squares
estimation of nonlinear parameters, Journal of the
Society for Industrial and Applied Mathematics, 11(2):
Mathworks. 2017. Matlab [Computer software]. Natick, MA,
Meyerhof, G.G. 1956. Penetration tests and bearing
capacity of cohesionless soils, Journal of the Soil
Mechanics and Foundations Division, ASCE, 82(1): 1-
Meyerhof, G.G. 1976. Bearing capacity and settlement of
pile foundations, Journal of the Geotechnical
Engineering Division, ASCE, 102(3): 197-228.
Nelder, J.A. and Mead, R. 1965. A simplex method for
function minimization, The Computer Journal, 7(4):
Oltean, M. and Dumitrescu, D. 2002. Multi Expression
Programming, Babes-Bolyai University, Romania.
Poulos, H.G., Carter, J.P., and Small, J.C. 2001.
Foundations and retaining structures -research and
practice, In Proceedings of the 15th International
Conference on Soil Mechanics and Geotechnical
Engineering, Istanbul, Istanbul, Turkey, 4: 2527-2606.
Shioi, Y. and Fukui, J. 1982. Application of N-value to
design foundations in Japan, In Proceedings of the 2nd
European Symposium on Penetration Testing,
Amsterdam, 1: 159-164.
Smits, G.F. and Kotanchek, M. 2005. Pareto-front
exploitation in symbolic regression, In O’Reilly et al.
(Eds.), Genetic programming theory and practice II,
Springer, Boston, MA, USA.
Tackett, W. A. 1994. Recombination, selection, and the
genetic construction of computer programs (PhD
dissertation), University of Southern California, Los
Angeles, CA, USA.
Vesic, A.S. 1977. Design of pile foundations, synthesis of
highway practice 42, Transportation Research Board,
Washington, DC, USA.
... For over time, many studies based on SPT test results have been proposed to predict the pile load capacity [7][8][9]. However, the disadvantage of the traditional method is that it is difficult to choose the right coefficients for the parameters, and the absence of all the other parameters has resulted in poor performance of these models [10]. ...
Accurate determination of bearing capacity is one of the key issues in pile foundation design. However, the estimation of the pile bearing capacity based on in-site test methods showed time-consuming and costly. Thus, the main object of this study is to apply Machine Learning (ML) method, namely Deep Neural Network (DNN) to predict the axial load capacity. A total of 472 static pile load test reports collected from constructions in Ha Nam province were used to build the model in which ten factors and the destructive load on the pile head is considered the single output variable. The original data is divided into three parts, including the training set (60%), the validation set (20%), and the testing set (20%) to build, validate, and test the model respectively. In particular, the best DNN model architecture is determined by a grid search technique. To verify the performance of DNN models, various methods named root mean squared error (RMSE) and R-squared (R2) were used. The results showed that the DNN model of two hidden layers achieved superior performance (average R2 = 0.897, RMSE = 108.515) in the prediction of pile bearing capacity compared to other DNN network architectures. This study can offer pile foundation design engineers an effective tool to quickly predict the axial load capacity of the driven piles.
... In the specific conditions, the above methods all prove to have many advantages, but, besides that, there are still many problems that need to be carefully considered before being widely applied in construction practice. Specifically, in their study, Jesswein et al. [13] showed that the calculation of pile load capacity according to SPT is not completely reliable because it still lacks precision, although it is cheap and easy to perform. Similarly, the analytical method is considered impractical, since the theoretical approach is based on too many assumptions and simplifications [3]. ...
Full-text available
Accurate prediction of pile bearing capacity is an important part of foundation engineering. Notably, the determination of pile bearing capacity through an in situ load test is costly and time-consuming. Therefore, this study focused on developing a machine learning algorithm, namely, Ensemble Learning (EL), using weight voting protocol of three base machine learning algorithms, gradient boosting (GB), random forest (RF), and classic linear regression (LR), to predict the bearing capacity of the pile. Data includes 108 pile load tests under different conditions used for model training and testing. Performance evaluation indicators such as R-square (R2), root mean square error (RMSE), and MAE (mean absolute error) were used to evaluate the performance of models showing the efficiency of predicting pile bearing capacity with outstanding performance compared to other models. The results also showed that the EL model with a weight combination of w1 = 0.482, w2 = 0.338, and w3 = 0.18 corresponding to the models GB, RF, and LR gave the best performance and achieved the best balance on all data sets. In addition, the global sensitivity analysis technique was used to detect the most important input features in determining the bearing capacity of the pile. This study provides an effective tool to predict pile load capacity with expert performance.
Conference Paper
Full-text available
This paper presents a broad review of shallow and deep foundations and retaining structures and the most significant methods developed to predict their behaviour. Static and some cyclic loading effects are considered, but dynamic behaviour has been excluded. Emphasis has been placed on methods that have been validated or found to be reliable for use in engineering practice. These include some well-tried and tested methods and others that have been suggested and validated in recent times by geotechnical researchers. Recommendations are made about preferred methods of analysis as well as those whose use should be discontinued. In addition, some observations are made about the future directions that the design of foundations and retaining walls may take, as there are still many areas where uncertainty exists. Some of the latter have been identified. RÉSUMÉ: Ce papier a pour objet donner une vue générale des types de fondations profondes et peu profondes et des structures de soutènement ainsi que des méthodes les plus significatives développées pour prédire leur comportement. Les effets de charges statiques et cycliques sont considérés mais le comportement dynamique a été exclu. L'accent a été mis sur les méthodes qui ont été validées ou reconnues fiables dans la pratique de l'ingénierie. Ce qui inclus tant quelques méthodes qui ont fait leurs preuves que des methods qui ont été suggérées et testées récemment par des chercheurs en géotechnique. Des réconisations sont formulées sur les méthodes d'analyse à privilégier et sur celles qui devraient être bandonnées. De plus, quelques pistes quant aux futures directions que devrait prendre le design des fondations et des structures de soutènement sont proposées, car dans de nombreux domaines des incertitudes subsistent. Parmi ces dernières, certaines sont identifiées.
Full-text available
Computational intelligence seeks, as a basic goal, to create artificial systems which mimic aspects of biological adaptation, behavior, perception, and reasoning. Toward that goal, genetic program induction“Genetic Programming”has succeeded in automating an activity traditionally considered to be the realm of creative human endeavor. It has been applied successfully to the creation of computer programs which solve a diverse set of model problems. This naturally leads to questions such as:  Why does it work?  How does it fundamentally differ from existing methods?  What can it do that existing methods cannot? The research described here seeks to answer those questions. This is done on several fronts. Analysis is performed which shows that Genetic Programming has a great deal in common with heuristic search, long studied in the field of Artificial Intelligence. It introduces a novel aspect to that method in the form of the recombination operator which generates successors by combining parts of favorable strategies. On another track, we show that Genetic Programming is a powerful tool which is suitable for real-world problems. This is done first by applying it to an extremely difficult induction problem and measuring performance against other state-of-the-art methods. We continue by formulating a model induction problem which not only captures the pathologies of the real world, but also parameterizes them so that variation in performance can be measured as a function of confounding factors. At the same time, we study how the effects of search can be varied through the effects of the selection operator. Combining the lessons of the search analysis with known properties of biological systems leads to the formulation of a new recombination operator which is shown to improve induction performance. In support of the analysis of selection and recombination, we define problems in which structure is precisely controlled. These allow fine discrimination of search performance which help to validate analytic predictions. Finally, we address a truly unique aspect of Genetic Programming, namely the exploitation of symbolic procedural knowledge in order to provide “explanations” from genetic programs.
Technical Report
Full-text available
In this paper a new evolutionary paradigm, called Multi-Expression Programming (MEP), intended for solving computationally difficult problems is proposed. A new encoding method is designed. MEP individuals are linear entities that encode complex computer programs. In this paper MEP is used for solving some computationally difficult problems like symbolic regression, game strategy discovering, and for generating heuristics. Other exciting applications of MEP are suggested. Some of them are currently under development. MEP is compared with Gene Expression Programming (GEP) by using a well-known test problem. For the considered problems MEP performs better than GEP.
The paper deals with the analysis of results from axial testing of vertical single piles, i. e. the most common field test performed. It aims to show that engineering value can be gained from elaborating on a pile test - during the actual testing in the field, as well as in the analysis of the results. Topics covered include testing methods, interpretation of failure load, influence of errors, and others. Refs.
Symbolic regression via genetic programming (hereafter, referred to simply as symbolic regression) has proven to be a very important tool for industrial empirical modeling (Kotanchek et al., 2003). Two of the primary problems with industrial use of symbolic regression are (1) the relatively large computational demands in comparison with other nonlinear empirical modeling techniques such as neural networks and (2) the difficulty in making the trade-off between expression accuracy and complexity. The latter issue is significant since, in general, we prefer parsimonious (simple) expressions with the expectation that they are more robust with respect to changes over time in the underlying system or extrapolation outside the range of the data used as the reference in evolving the symbolic regression. In this chapter, we present a genetic programming variant, ParetoGP, which exploits the Pareto front to dramatically speed the symbolic regression solution evolution as well as explicitly exploit the complexity-performance trade-off. In addition to the improvement in evolution efficiency, the Pareto front perspective allows the user to choose appropriate models for further analysis or deployment. The Pareto front avoids the need to a priori specify a trade-off between competing objectives (e.g. complexity and performance) by identifying the curve (or surface or hyper-surface) which characterizes, for example, the best performance for a given expression complexity.
Conventional bearing capacity of driven and bored piles in sand and nonplastic silt is limited to short piles above the critical depth of penetration. The bearing capacity of longer piles can be estimated from the limiting values of the point resistance and skin friction using either the friction angle of the soil or preferably the results of static and standard penetration tests directly. For driven and bored piles in clay and plastic silt, conventional bearing capacity theory using the undrained shear strength of the soil represents mainly the failure condition at the pile points. The positive and negative shaft resistance some months after pile installation is governed by the drained shear strength of remolded soil and can be estimated from skin friction factors, provided the earth pressure coefficient at rest of the deposit is known. The ultimate load of pile groups and the settlement of pile foundations are examined.