Inverse Design of NATM Tunnels by Neural Networks

download Inverse Design of NATM Tunnels by Neural Networks

of 18

Transcript of Inverse Design of NATM Tunnels by Neural Networks

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    1/18

    12thICSGE10-12 Dec. 2007

    Cairo - Egypt

    Ain Shams University

    Faculty of Engineering

    Department of Structural Engineering

    Twelfth International Colloquium on Structural and Geotechnical Engineering

    INVERSE DESIGN OF NATM TUNNELS BY NEURAL

    NETWORKS

    F. M. EL-NAHHAS

    Professor, Head of Department of Structural Engineering,Ain Shams University

    H. E. A. ALI

    Associate Professor, Department of Structural Engineering,

    Ain Shams University

    S. M. EL-ARABY

    Assistant Professor, Department of Structural Engineering,

    Ain Shams University

    H. E. ABDELBARY

    PH.D. Research Student, Department of Structural Engineering,

    Ain Shams University

    ABSTRACT

    Tunnels are often designed using uncertain geotechnical data. Insufficient boreholes,

    natural variation, difficulties in extracting undisturbed samples and lack of realistic

    testing procedures are common impediments rendering precise prediction of the

    associated deformations and lining stresses practically unachievable. Faced with these

    uncertainties, geotechnical engineers usually opt to re-appraise the assumed parameters

    by inverse analysis using the monitoring measurements. The parameters obtained from

    such analyses can then be realistically implemented in subsequent geotechnical

    assessments.

    In this paper, artificial neural networks (ANNs) are used in an attempt to simulate theback analysis of the tunnels to obtain realistic parameters that can be used in the finite

    element analysis to achieve more accurate design of tunnels. A large database of actual

    models for a real case study of Tunnel in Algeria is used to develop and to verify the

    ANN model. The designed tunnel deformations determined by utilizing ANNs through

    the finite element analyses are compared with the measured deformations. The results

    indicated that ANNs are useful technique for inverse design of tunnels.

    KEYWORDS

    Tunnels, NATM, Neural Networks, Rock Excavation, and Inverse (back) Analysis.

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    2/18

    1 INTRODUCTION

    In order to improve tunnel ground support behaviour and reduce costs associated with

    over-design or rehabilitation due to under-design, there is always a need for better

    modelling capability of support and lining behaviour prior to construction. Hence, the

    need exists to model the rock mass behaviour well past the yielding point, not just thetypical elastic stress modelling that currently occurs in tunnelling [1].

    Incompatibility often prevails between the predicted and the real behaviour of the model

    and only during construction, it is possible to re-evaluate the input data and to

    continuously refine the numerical model using the field measurements and inverse

    analysis techniques. In this study, the inverse analysis conducted by the artificial

    intelligence techniques (in particular Artificial Neural Networks, ANN) is investigated.

    Over the last few years, the use of Artificial Neural Networks (ANNs) has dramatically

    increased in many engineering fields [2], [3]. In particular, ANNs have been applied to

    various geotechnical engineering problems, including the design of tunnel andunderground openings [4], and have demonstrated considerable degree of success in that

    field.

    2 CASE STUDY

    The case study involves The Bouira Tunnel, which forms part of the construction lot

    Lakhdaria-Liaison RN5 of Autoroute Est-Ouest and is located near the village of

    Djebahia in the north of Algeria. It is a twin tube motorway tunnel with three lanes and

    walkways on both sides in each tube, as shown in Fig 1. The tubes are running parallel,

    with a distance between tunnel axes of 34.5m. Excavation cross section varies between

    153 and 156m2.

    Fig. 1: Cross-Section of the Bouira Tunnel at East Portal

    The tunnels are running below a pass located at the end of Oued Djelada bordered by

    Djebel Harchaoua (Elevation +770 m) in the southwest and by Djebel Guenadir(Elevation +720 m) in the Northeast, as shown in Fig 2 [5]. The village of Ain Cheriki

    is situated on this pass which shows an elevation of approximately 560m. The majority

    of the tunnel alignment runs in a straight line from Northwest to Southeast before

    turning to the east at the eastern tunnel portal. Overburden varies between 8 to 68m.

    Table (1) shows the chainages of the portals for each tunnel.

    Table 1: The Tunnel Portals are Located at the following Chainages:

    Tunnel Western

    Portal

    Chainage (m)

    Eastern Portal

    Chainage

    (m)

    Total

    Length

    (m)

    T1 182+985 184+230 1245

    T2 183+025 184+208 1183

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    3/18

    Cut and cover sections have originally been planned for the western portal for T1

    between chainage PK 182+985 to PK 183+040 and for T2 between chainage PK

    183+025 to PK 183+075. The rest of the tunnels have been planned to be excavated

    using the New Austrian Tunnelling Method (NATM).

    Fig. 2:The Geological Map for Bouira Tunnel

    2.1 The Tunnel and Geology

    The area in which the tunnel alignment is situated forms part of folded mountain chain

    (Rifo-Tellin Chain), which extends over the entire length of the Maghreb and is part

    of the Alpine chain. In the area of northern Algeria it is named Tell Atlas. According

    to Pique [6] the region of interest should be part of the external zones of the Rifo-

    Tellian mountain chain. The external zones are further divided into the ultra-Tellian

    nappe, the Tellian units sensu strictu and the peni-Tellian and southern units. Due to the

    lack of the regional geologic maps, an exact allocation of the region of interest to one ofthese geological units is not possible up to now. This area has been exposed to

    polyphase, complex tectonic deformation resulting in folding and faulting. Due to the

    continuing convergence of Africa and Eurasia the area is at present still submitted to

    active compressive faulting [6]. Three lithological formations of Senonian age are

    described as follows:

    - Foliated marl, strongly fractured and folded, at the base and the middle of the

    excavation.

    -Compact marl and hard limestone, intercalations of compact marls and of hardlimestone with a thickness of several meters.

    -Clay / alterated Marl, light colored, dense marls, very superficially weathered.The longitudinal geological section, as shown in Fig 3, shows the three lithogies: sound

    marl, foliated marl and clay/alterated marl. All three lithological unites may be

    considered as a succession of clay and deformed marl of Senonian age.

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    4/18

    Fig. 3:The Longitudinal Section of the Tunnel with Geology

    2.1 Monitoring

    Two types of monitoring sections are considered. Monitoring section type A contains 5

    points for convergence measurements and monitoring section B with 5 points for

    convergence measurements and multiple rod extensometers. Distance between

    monitoring sections type A is generally 100m. Three monitoring sections type B are

    considered for each tunnel, Geoconsult 2004 [7]. Fig.4 shows the plan and cross-section

    for the monitoring system.

    Fig. 4: Schematic showing Instrumentation through Bouira Tunnel

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    5/18

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    6/18

    3.1 Calculation Steps

    Table 2 shows calculation phases and the construction stages used in the finite element

    simulations. The twin tunnels were explicitly modelled in the finite element simulations

    while the study was carried out on tunnel (T1), as the construction of tunnel (T2) was

    not started at the time when the field data measurements were recorded.

    Table 2: PLAXIS Calculations Phases and the Corresponding Construction Stages

    Calculation Phase Construction Stage

    0 Initial Condition

    1 T1 Heading Excavation

    2 T1 Heading Support

    3 T1 Bench Excavation

    4 T1 Bench Support

    3.2 Rock Parameters and Excavation Sequence

    The elastic-plastic Mohr Coulomb model was used to characterise the rock constitutive

    behaviour in PLAXIS simulation of the tunnel construction. Mohr Coulomb model

    involves five input parameters, i.e. E and for soil elasticity; and c for soil plasticity

    and as an angel of dilatancy.

    A parametric study was carried out as shown in Figures 8 to 19 to obtain the most

    influential parameters on the tunnel design. Based on the parametric analysis, Table 3

    shows the range of the effective parameters that was used through the analysis to obtain

    the horizontal deformations in order to learn the neural network model.

    Fig. 8: Effect of the Unloading Ratio for Fig. 9: Effect of the Unloading Ratio for

    Heading Excavation on the Vertical and Bench Excavation on the Vertical and

    Horizontal Total Deformations Horizontal Total Deformations

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    7/18

    Fig. 10: Effect of the Ko of Foliated Marl Fig. 11: Effect of Ko of Marl Shale on

    on the Vertical and Horizontal Total the Vertical and Horizontal Total

    Deformations Deformations

    Fig. 12: Effect of the Ko of Shear Zone Fig. 13: Effect of E of Foliated Marl of

    on the Vertical and Horizontal Total on the Vertical and Horizontal Total

    Deformations Deformations

    Fig. 14: Effect of the E of Marl Shale Fig. 15: Effect of E of Shear Zone

    on the Vertical and Horizontal Total on the Vertical and Horizontal Total

    Deformations Deformations

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    8/18

    Fig. 16: Effect of the C of Foliated Marl Fig. 17: Effect of Phi of Foliated Marl

    on the Vertical and Horizontal Total on the Vertical and Horizontal Total

    Deformations Deformations

    Fig. 18: Effect of the C of Marl Shale Fig. 19: Effect of Phi of Marl Shale

    on the Vertical and Horizontal Total on the Vertical and Horizontal Total

    Deformations Deformations

    Based on the previous parametric study, it was found that the following parameters have

    the most influential on the horizontal and vertical deformations of the tunnel; Unloading

    Ratio for Heading Excavation, Unloading Ratio for Bench Excavation (The Unloading

    Ratio is the ratio between the part of the stresses released due to the excavation and theremained stresses that will be imposed on the lining), Youngs Modulus E for

    Foliated Marl Layer, Youngs Modulus E for Compact Marl Layer, Cohesion C for

    Compact Marl Layer and Friction for Compact Marl Layer.

    Table 3, based on this parametric study, shows the range of the influential parameters

    that was used through the analysis to obtain the horizontal deformations in order to train

    the neural network model.

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    9/18

    Table 3:Range of the most Influential Parameters

    Parameters Range Units

    Unloading Ratio for

    Heading Excavation

    0.2 0.75 -

    Unloading Ratio for

    Bench Excavation0.2 0.75 -

    Youngs Modulus E for

    Foliated Marl Layer400 900 MPa

    Youngs Modulus E for

    Marl Shale Layer

    100 600 MPa

    Cohesion C for Marl

    Shale Layer

    50 100 kPa

    Friction for Marl ShaleLayer

    11 - 22 Deg.

    The following figures from 20 to 25 showing the obtained deformations corresponding

    to the variations of the most influential parameters after the heading excavation (step

    no.1)

    Fig. 20: Effect of the Unloading Ratio for Fig. 21: Effect of the Unloading Ratio

    Heading Excavation on the Vertical and for Bench Excavation on the Vertical

    and Horizontal Deformations Horizontal Deformations

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    10/18

    Fig. 22 Effect of the E of Foliated Marl Fig. 23 Effect of the E of Marl Shale

    on the Vertical and Horizontal on the Vertical and Horizontal

    Deformations Deformations

    Fig. 24: Effect of the C of Marl Shale Fig. 25: Effect of Phi of Marl Shale

    on the Vertical and Horizontal on the Vertical and Horizontal

    Deformations Deformations

    4 DEVELOPMENT OF NEURAL NETWORK MODEL

    The Pioneering work in developing ANN models shows that neural network concepts

    can be extended to constitutive relations [10], [11]. A neural network is trained using

    the input and output values deduced from the finite element analysis to learn the

    material behaviour. If the training data contains enough relevant information, the trained

    neural network can generalize material behaviour to new loading cases.Among various types of neural networks, multi-layer feed-forward network is known to

    be most suitable to describe non-linear functions [12], and so far, has been the only type

    of neural network used to describe material constitutive behaviour.

    The simulation of the neural network model through this work was carried out by

    personal computer-based software NeuroSolutions Version 4.3 (2003) developed by

    NeuroDimension, Inc [13].

    The data used to calibrate and validate the neural network model were obtained from the

    parametric study of the most influential parameters affecting on the tunnel design as

    well as the corresponding deformations associated with the tunnel excavation. As

    explained earlier, the data cover a wide range of variation in soil parameters and

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    11/18

    sequence of excavation (unloading ratio). The database comprised a total of 77

    individual cases.

    4.1 Inputs and Outputs of the ANN-based Model

    Ten input parameters were used in the ANNs model, as follows:- Horizontal and vertical deformations for point no.1 (P1)- Horizontal and vertical deformations for point no.2 (P2)- Horizontal and vertical deformations for point no.3 (P3)- Horizontal and vertical deformations for point no.4 (P4)- Horizontal and vertical deformations for point no.5 (P5)While the output layer consists of the following:

    - Unloading Ratio for Heading Excavation

    - Unloading Ratio for Bench Excavation

    - Youngs Modulus E for Foliated Marl Layer

    - Youngs Modulus E for Compact Marl Layer

    - Cohesion C for Compact Marl Layer

    - Friction for Compact Marl Layer

    4.2 Data Division and Processing

    It is common practice to divide the available data into two sub-sets; a training set, to

    construct the neural network model, and an independent validation set to estimate model

    performance in the deployed environment [14]. However, dividing the data into only

    two subsets may lead to model overfitting. As a result, and as discussed later, cross

    validation [15] is used as the stopping criterion in this study and, consequently, the

    database is randomly divided into three sets: training, testing and validation. In total,

    70% of the data are used for training and 15 % are used for validation set and 15% for

    the testing set.

    4.3 Model Architecture

    Determining the network architecture is one of the most important and difficult tasks in

    the development of ANN models. It requires the selection of the number of the hidden

    layers and the number of the nodes in each of these. It has been shown that a network

    with one hidden layer can approximate any continuous function, provided the sufficient

    connection weights are used [16]. Consequently, one hidden layer is used in this study.

    The utilized network paradigm is a Multi-Layer Perceptron type (MLP) consisting of

    multiple layers of PEs connected in a feedforward pattern. Back-propagation of errors

    technique is used to train the MLP. The calculated backpropagation components of errorpass backwards from the end of the network to the beginning. The gradient search

    components adjust the weights contained in the synapses and axons. The network is

    constructed using the medium complexity setting which have one hidden layer. The

    criterion axon reads the desired file from the attached file component and determines

    the error in the network. The output axon generates the actual network outputs. The

    input axon does nothing but accept the input from the file component.

    The number of the nodes in the input and output layers are restricted by the number of

    the model inputs and outputs. The input layer of the ANN model developed in this work

    has ten nodes; one for each of the model inputs [i.e., horizontal and vertical

    deformations for point no.1 to point no. 5 (P1 to P5). The output layer has six nodes

    [i.e., unloading ratio for heading excavation, unloading ratio for bench excavation,

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    12/18

    youngs modulus E for foliated marl layer, youngs modulus E for compact marl

    layer, cohesion C for compact marl layer, friction for compact marl layer]. Inorder to obtain the optimum number of hidden layer nodes, it is important to strike a

    balance between having sufficient free parameters (weights) to enable representation of

    the function to be approximated, and not having too many so as to avoid overtrainingand to ensure that the relationship determined by the ANN can be interpreted in a

    physical sense. Overtraining is not an issue in this study, as crossvalidation is used as

    the stopping criterion. However, as just discussed, physical interpretation of the

    connection weights is important, and hence the smallest network that is able to map the

    desired relationship should be used. In order to determine the optimum network

    geometry, ANNs with one, two, three and four hidden layer nodes are investigated.

    Fig. 26:The Skeleton of the Neural Network Model

    4.4 ANN Model Training

    The process of optimizing the connection weights is known as training or learning.

    The method most commonly used for finding the optimum weight combination for

    feedforward neural networks is the back-propagation algorithm [17], which is based on

    first-order gradient descent. Feedforwards networks trained with the backpropagationalgorithm have already been applied successfully for many geotechnical engineering

    problems [18] and [19], and are thus used in this work.

    In this study, the general strategy adopted for finding the optimal parameters that

    control the training process is as follows; for each trial number of hidden layer nodes,

    random initial weights and biases are generated. The neural network is then trained with

    different combinations of momentum terms and step sizes in an attempt to identify the

    ANN model that performs best on the testing data. The momentum terms used in this

    study are 0.2, 0.4, 0.6, 0.8 and 1.0 whereas the step sizes used are 0.01. 0.05. 0.1. 0.3.

    0.6 and 1.0

    4.5 Stopping CriteriaStopping criteria are those used to decide when to stop the training process. They

    determine whether the model has been optimally or suboptimally trained. As discussed

    earlier, the crossvalidation technique [15] is used in this work as the stopping criterion,

    as it is considered to be the most valuable tool to ensure that overfitting does not occur

    [20] and as sufficient data are available to create training, testing and validation sets.

    The testing set measures the ability of the model to generalize, and the performance of

    the model using this set is checked at many stages of the training process, and training is

    stopped then the error of the testing set starts to increase. The testing set is also used to

    determine the optimum number of hidden layer nodes and the optimum internal

    parameters (step size, momentum and initial weights).

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    13/18

    4.6 ANN Model Validation

    Once the training phase of the model has been successfully accomplished, the

    performance of the trained model is validated using the validation data, which have not

    been used as part of the model building process. The purpose of the model validation

    phase is to ensure that the model has the ability to generalize within the limits set by thetraining data, rather than simply having memorized the input-output relationships that

    are contained in the training data.

    5 INVERSE MODELING: UPDATING TUNNEL DESIGN USING

    REALISTIC ROCK PARAMETERS

    After testing the trained network, it was shown that the correlation coefficient (R2)

    between the networks output and the desired value is reached up to 0.88, 0.61 and 0.67

    for data sets of training, cross validation and testing, respectively, which indicates a

    successful performance as shown in (Table 4).

    Table 4: Performance of Network for Different Output Parameters

    R2

    Data Set Unloading

    Ratio for

    Heading

    Excavation

    Unloading

    Ratio for

    Bench

    Excavation

    E for

    Foliated

    Marl

    Layer

    E for

    Marl

    Shale

    Layer

    C for

    Marl

    Shale

    Layer

    forMarl

    Shale

    Layer

    Training 0.66 0.69 0.48 0.43 0.46 0.88

    Cross

    Validation0.30 0.61 0.03 0.06 0.35 0.55

    Testing 0.51 0.003 0.46 0.53 0.67 0.29

    In Figures 27 to 44, the actual values for the most influential are plotted against the

    networks output for different data sets of training, testing and cross validation.

    Fig. 27 Fig. 28

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    14/18

    Fig. 29 Fig. 30

    Fig. 31 Fig. 32

    Fig. 33 Fig. 34

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    15/18

    Fig. 35 Fig. 36

    Fig. 37 Fig. 38

    Fig. 39 Fig. 40

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    16/18

    Fig. 41 Fig. 42

    Fig. 43 Fig. 44

    5.1 Comments on the results

    As shown in the aforementioned figures, it is clear that the influential parameters in

    tunnel design can be predicted by using a Neural Networks model. The training of the

    Neural Networks model is the key of obtaining more accurate results. This effect is

    appeared in Figures 29, 30, 35 and 42. In order to improve the accuracy of the results, a

    larger database should be used to train the model.

    5.2Validation of the modelThe neural network model is validated by comparing the results of the model with thedeformation and stress measurements that are deliberately excluded from the learning

    database, as indicated in Table 4 and Figures 27 through 44.

    As this work is a part of an ongoing research for using the artificial neural networks in

    the back analysis of tunnels, the authors believe, at this stage, that the accuracy of the

    prediction can be enhanced by adopting the results of well-trained neural networks into

    the learning database. This hypothesis shall be investigated/validated at the concluding

    stage of this research.

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    17/18

    6 SUMMARY AND CONCLUSIONS

    At present, tunnels tend to be designed with considerable uncertainties. In this paper, a

    neural network approach was presented to determine realistic soil/rock parameters that

    to be used in the tunnel design.

    A two hidden layer MLP was trained using data sets supplied by FE models. The

    number of neurons, step sizes and momentum of the layers were optimized using

    genetic algorithms during the training process. The testing of the trained network

    showed that the neural network model can predict successfully the soil/rock parameters

    that will be used in the FE model in the tunnel design.

    7 ACKNOWLEDGMENT

    The Authors would to thank Mr. Issam Ghalayini and Dr. Sherif Wissa, Directors of the

    Geotechnical and Heavy Civil Engineering Department - Dar Al-Handasah (shair and

    partners) for their kind help and permit to use data from their files on the Bouira tunnel.

    8 NOTATIONS

    Ko At Rest Coefficient of Earth Pressure

    E Elasticity of Modulus [MPa]

    C Cohesion [kN/m2]

    Angle of Friction [deg.]

    R2 Coefficient of Correlation

    9 REFERENCES

    [1] Crowder, J.J., and Bawden, W. F. Review of post-peak parameters and

    behaviour of rock masses: Current trends and research [Online]. (2004)

    Rocnews, Fall 2004, available from: http://www.rocscience.com/library/rocnews

    [2] Ali, H. A.; 2001, "Neuronet Based Liquefaction Potential Assessment and

    Stress-Strain Behavior Simulation of Sandy Soils", Ph. D. Thesis, Kansas State

    University, Manhattan, Kansas, USA.

    [3] Fayed, Ayman, 2002, "Interaction between Deep Braced Excavation and Groundfor Metro Subway Stations", Ph. D. Thesis, Ain Shams Univ., Cairo, Egypt.

    [4] Nour El-Din, S.M. (2003). Neuronet Prediction of Surface Settlement

    Associated with Soft Ground Tunnelling. M.Sc. Department of Structural

    Engineering, Ain shams University.

    [5] DAR Al-Handasah, United Kingdom, 1999. Autoroute Est-Ouest, Troncon

    Lakhdaria-Liaison RN5, Lot Tunnel du PK182+985 au PK184+230

    Tunnels de Bouira, Algeria, Rapport de Presentation.

    [6] Pique, A. Geology of the Northwest Africa Beitragze zur Regionalen Geologie

    der Erde, Band 29. Engl. Test by Carpenter, M.S.N., Gebruder Borntraeger

    (Berlin, Stuttgart). (2001)

  • 8/12/2019 Inverse Design of NATM Tunnels by Neural Networks

    18/18

    [7] Geocounsult, Austria, 2004, Autoroute Est-Ouest, Troncon Lakhdaria-Liaison

    RN5, Lot Tunnel du PK182+985 au PK184+230 Tunnels de Bouira,

    A;geria, Review of Optical Displacement Monitoring Temporary Support

    Performance, Tunnel T1 Report.

    [8] El-Sayed, S. M., 2001, Elasto-Plastic Three Dimensional Analysis of ShieldedTunnels, with Special Application on Greater Cairo Metro, Ph. D. Thesis, Ain

    Shams Univ., Cairo, Egypt.

    [9] Plaxis, (2002), Finite Element Code for Soil and Rock Analysis, Version 8.2,

    Plaxis B. V., P.O.Box 851, 3160 AB RHOON, Netherlands.

    [10] Ghaboussi J, Garret J. H, Wu X. (1991). Knowledge-based modeling of

    material behavior with neural networks. Journal of Engineering Mechanics

    Division (ASCE), 117(1):132-153

    [11] Ghaboussi J, Sidarta D. E. (1997). New method of material modeling using

    neural networks. Proceedings of the 6thinternational Symposium on Numerical

    Models in Geomechanics. 393-400

    [12] Reed RD, Marks RJ. (1999). Neural Smithing Supervised Learning inFeedforward Artificial Neural Networks. The MIT Press: Cambridge. MA.

    [13] NeuroSolution Version 4.3, NeuroDimension, Inc. 1800 N. Main Street, Suite

    D4, Gainesville, FL 32609.

    [14] Twomey, J. M., and Smith, A. E. (1997). Validation and verification.

    Artificial neural networks for civil eningineers: Fundamentals and applications,

    N. Kartam, I. Flood, and J. H. Garrett, eds. ASCE, New York, 44-64.

    [15] Stone, M. (1974). Cross-validatory choice and assessment of statistical

    predictions. J. R. Stat. Soc., B 36, 111 147

    [16] Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward

    networks are universal approximates. Neural Networks, 2, 359-366.

    [17] Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning internal

    representation by error propagation. Parallel distributed processing, Vol. 1,

    Chap. 8, D. E. Rumelhart and J. L. McClelland.

    [18] Goh, A. T. C. (1994). Seismic liquefaction potential assessed by neural

    networks. J. Geotech. Geoenviron. Eng., 120(9), 1467-1480.

    [19] Najjar, Y. M., and Basheer, I. A. (1996).A neural network approach for sirte

    characterization and uncertinity prediction. Geotechnical Special Publication,

    58(1), 134-148

    [20] Smith, M. (1993). Neural networks for statistical modeling., Van Nostrand-

    Reinhold, New York.