Contents lists availmble at ScienceDirect ind FRONTIERS China University of Geosciences (Beijing) Geoscience Frontiers journal homepag! ww_elsevier.comilocatelast Research paper New design equations for estimation of ultimate bearing capacity of shallow foundations resting on rock masses Amir H. Alavi*!, Ehsan Sadrossadat” “peparment of Cian Envinmental Engineering. Miigan Stat Universi ast Lansing MI 6882, USA “pspartmen of Chl Engnerng Mastbad Bench, blame Arad Unies, Marte, re Received 4 Sepuembet 2016 Rock masses are commonly used asthe underlying layer of important structures suoh as bidges dams and transportation eonsteuctions. The success ofa foundation design for sch strictures mainly depends fn the accuracy of estimating the bearing capacity of rock beneath them. Several traditional numerical approaches ae proposed forthe estimation of che bearing capacity of foundations resting on rock masses {avoid performing elaborate ané expensive experimental studies. Despite this fat, there stl exists 3 ‘serols need to develop more robust precctive models. This paper proposes new nonlinear prediction ‘models forthe ultimate bearing capacity of shallow foundation resting on non-fractued rock masses using 2 novel evolutienary corsputational approach, called linear genetic programming. A comprehen- sive set of rok socket, centrifuge rock socket, plate load and large-scale footing load est results i used tadevelop the models, inorder to verify the validity ofthe models, the sensitivity analysis scondseted and discussed, The results indicate that the proposed models accurately characterize Une Dearing capacity of shallow foundations. The cortelation coefficients between the experimental and preccted bearing ‘capacity values are equal 0.95 and 0.96 forthe best LCP models. Moreover, te derived models reach notably beter prediction performance than the traditional equations. © 2014, China University of Geosciences (Beijing) and Peking University Production and hosting by isevier BV. This is an open acces article under the CC BY-NC-ND license (tp /ereativecommons.arg Ticensesfbyenc-nd0). 1. Introduction ‘width (B), joint condition, rock type, and the condition of the underlying layer of rock mass (Bishon’, 1968; Sowers, 1979). The Foundations are commonly used as the lowest parts of the most widely used approaches to determine the bearing capacity civil engineering structures to transmit the applied loads to the underlying soil or rock. According to the properties of the rock ‘mass and its beneath layer, the failure of racks under applied Toads may occur through several mechanisms (Sowers. 1979). Comprehensive descriptions about these failure mechanisms are provided in Canadian Foundation Engineering Manual and the National Cooperative Highway Research Program (NCHRP) re- ports (Becker, 1996; Palkowsky et al, 2004, 2010; Canadian Geotechnical Society, 2006). It is well-known that the bearing capacity failure of shallow foundations on jointed rack masses depends on the ratio of space between joints (5) to foundation esadcossadat@giailcon,ehsan sacrossadat@gsnal con (E Sadcossada). Sup: egemsu egal) Peee-enew under responsbity of Cina University of Geosciences (eine IND eens (stp jezeatveconsmansorgicensesiby-ne-nd/ 40). of foundations on rocks can be classified into three groups: (1) analytical methods, (2) semi-empirical methods, and (3) in-situ and full-scaled testing methods. The analytical and semi- empirical methods are widely used for the beating capacity, prediction, particularly in the pre-design phases. The analytical methods such as finite element and limit equilibrium methods relate the beating capacity to the footing geometry and rock properties (Terzaghi, 1946: Bishoni, 1968; Sowers. 1978; Goodman, 1989), The general forms of the anaiytical models are given in Table 1, The semi-empitical methods often propose a correlation between the bearing capacity and rack mass proper- ties based on the empirical observations and experimental test results (Carter and Kulhawy, 1988; Bowles, 1996; Hoek and Brown, 1997, 1988), Some of the main equations obtained by the empitical approaches are summarized in Table 2 Generally, the models obtained by the analytical, nite element and empirical approaches have both advantages and disadvantages2 A Ali Sosa Geoscience Fronts 7208) 91-99 ‘Cnc forms of equations made by he asic ets eee Feuaon abi todd eer Teg StS tan =o OAR, TOR ROE naan) Nap i = am (35+4) ‘Bishan (1968) has = ake For eeetla footings: ¢— 1 Ne Brvorn()(0 +) Ngleat 6) 420 For square fonings « ~ 085 Y Soets (179) a nuns) ‘ - Goodman (1989) Forfracedrsks a = 1 y= wnt (4s 0) ( lu (gy (mts) i) ‘Gas bearing capaci of shallow foun on oc: D depth a fundalion Below ground vac the cedoniveceps Toth ack mass fangs ofntarnal fico ot (Giao etal, 2012; Chen et al, 2013) For the case of bearing capacity ‘of a shallow foundation om a jointed rock mass, parameters such as ‘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 and rock mass strength are influencing (Sowers, 1979: Paikowsky et al, 2010). As represented in Table 1, the analytical ‘methods only include the physical and mechanical properties of rock mass and geometry of foundation. Thus, they do not take into ‘account the important role ofthe rock type and its qualitative mass parameters such as rock quality designation (ROD), rock mass rat- ing (RMR) and geological strength index (CSI). On the other ban the empirical methods often relate the bearing capacity to quali- {ative and rock mass classification parameters and do not account for the geometry of the foundations or space between joints (Table 2). The drawbacks of the existing analytical and empirical, ‘methods imply the necessity of developing new models correlating the beating capacity factor to both quantitative and qualitative parameters Computational intelligence (CI) techniques are considered as, alternatives to traditional methods for tackling real world prob- lems. They automatically lear from data to determine the struc tute of a prediction model. Artificial neural network (ANN), fuzzy inference system (FS), adaptive neuro-fuzzy system (ANFIS), and Support vector machines (SVM) are well-known branches of Cl ‘These techniques have been successfully employed to solve Problems in engineering field (e.g. Das and Basudhar, 2005; Gull and Ercelebi, 2007: Das and Basudhat, 2008; Gullu and Ercelebi, 2008; Zorls et al, 2008; Cabslar and Cevik, 200936; Sivapullaiah et al, 2009; Yaghouby et al, 2009; Al-Anazi and Gates, 2010; Kulatilake et al, 2010; Yaghouby et al. 2010ab, 2012; Golly, 2012, 2013), Besides, these techniques have been used £0 predict the bearing capacity of shallow foundations resting on soil layers (Soleimanbeigi and Hataf, 2005; Padmini et al, 2008; Kuo et al, 2009; Kalinlt etal, 2011), Despite the good performance of ANNs, AIS, ANFIS, SVM and many ofthe other CI methods, they are nso; Ne: beating capac lao, 5 scotia spacing: Ss aloof spacing loundaion vid; uoconined empress eng ofr considered black-box models. That is, they are not capable of ‘generating practical prediction equations, In order to cope with the Limitations of the existing methods, a robust Cl approach, namely genetic programming (GP) has been introduced (Koza, 1992), GP is an evolutionary computational approach. It uses the principle of Darwinian natural selection to generate computer programs for solving a problem. GP has several advantages over the conventional and other similar techniques. A notable feature of GP and its vatiantsis that they can produce prediction equations ‘without a need to pre-define the form of the existing relationship (Ganaker et al, 2009; Guven et al, 2009; Gandom! et al, 2010) Alavi et al, 2011; Alavi and Gandomi, 2011; Gandomi et al, 2ola; Azamathulla and Zahiri, 2012). This technique has been shown to be a powerful tool for the prediction of the bearing ca- pacity of shallow foundations on soils (Adarsh et al, 2012; Shahnazari and Tutunchian, 2012; Tsai et al, 2012; Pan and Tsai 2013}, However, application of GP and other CI techniques to the ‘modeling ofthe beating capacity of shallow foundations resting on rock masses is conspicuous by its absence. This paper proposes 2 novel subset of GP, namely linear genetic programming (LGP) 0 derive precise predictive equations for the ultimate bearing ca- pacity of shallow foundation resting injon jointed (non-fractured) rock. A comprehensive and reliable set of data including 102 pre- viously published rock socket, centrifuge rock socket, plate load and large-scaled footing load test results is collected to develop the ‘models. The robustness ofthe proposed models is verified through different validation phases, 2, Evolutionary computation Evolutionary computation (EC) is a subdivision of Cl inspired by the natural evolution. Some of the subsets of EC are evolutionary strategies (ES) and evolutionary programming (EP). These tech- niques are collectively known as evolutionary algorithms (EAS). Bowes {15ST Ge a ROD Heck sd Brown (1997, 1988) a= es ingey Canter an Klay (1968) aux = ev m= mn (Sg) = 9 ( Tb srock quay deignaton (ook man cavniicalon) se taken at postive} minor principal sess GS geological sveng.h index ek mass asiicin) mand male constants ia the Hee and own (ale eseton:‘AN Alo Sarosedat Geoscience omer 7 ( V. Repeating steps Il through IV until termination or conver- ce conditions are satisfied. Data samples are known Processing li Selected in each ofthe two patents and exchanged. If ne ofthe ee eu / (Brameier and Banzhaf, 2001; Gandomi et al, 2010). The muta- soducion tion operation occurs on a single instruction. Two commonly tions (individuals) are computer programs rather than binary TT 3. Numerical simulation of bearing capacity In order to reach reliable estimations of the bearing capacity of shallow foundations on rock mass, the impact of several parame- ters should be incorporated into the model development. The sxeneral forms of the existing prediction equations, represented in Tables 1 and 2, indicate that the ultimate bearing capacity of shallow foundations on rock mass mainly depends on the foun- dation width and properties of the rock beneath it. The present study takes into account the effects of both quantitative and qual- iMative parameters to predict the bearing capacity, Referring to the form of the existing models, two prediction models (LGP 1 and LGP 2) are developed using the LGP technique. Consequently, the pro- pposed models for the prediction of the ultimate bearing capacity (aus) ate considered to be the functions of the following pauneter *) a 1 tnitiaizing population of randomly generated programs and calculating their fess values danscrs ~F(RMB, 4. Running 4. Tournament. In this step four programs are selected fom the population randomly. They are compaied based on thie tess Two programs ae then picked a8 che ay. ucry = F(m, 8.40: 4) @) winners an two ae the losers, ut, Transforning the winner programs, After that wo winner whee, Programs are copied ané transformed probally nt {wo new programs va crosover and mutation operators. RoR: Rock mas rating (rock mass casstcaion), Iv. Replacing the loser programs inthe tournament ith the Sate constat i ie Hoek and Brown fae exterion twansformed winner programs. The winners ofthe foua- _¢; Material constant inthe Hock and Brown fale citron, tment remain unchanged du tka) Unconfined compressive strength of rack a () 7 © [mae eS ee = i L2sf0]+=2; SO | sae 2] | assgoyones “] COE 7] Ci ov, igure 2. Acompson of progcam strate eed by) ee ated, 9) grapvused GF, ad (eeu GP or he expeson of ~ f= (ell 4 2292AH. Alo Sass Geoscience Fo Saumiel vara dex aR we ~ we Standard 2437112 008 v0.73 25825 ‘devon 59176 126 a0 2914708 5367 2428 7194810 Minimo 1500 003 000003 Sa Sa Maxima 10000 345 08114870 57435 SB: Ratio of joint spacing to foundation width (equivalent ameter) #0): Angle of internal friction forthe rock mass 31, Experimental database ‘A comprehensive database including 102 elaborate experi- ‘mental data from different studies is gathered to develop the L¢P-based models (Ward and Burland, 1968; Burland and Lord, 1970; Lake and Simons, 1970; Webb, 1976; Wilson, 1976; ‘Mallard, 1977; Goeke and Hustad, 1979: Pells and Turner, 1979, 3980; Spanovich and Garvin, 1878; Thome, 1980; Williams, S08 E. RMR E ate Ze En] ‘SB Ze F zy . 7 (208) 51-38 1980; Jubenville and Hepworth, 1981; Glos and Briggs, 1983: Baker, 1985; Hummert and Cooling, 1988; Radhakrishnan and Leung, 1989: Leung. and Ko, 1993: Maleki and Hollberg, 1995; Nitta et al, 1995; Carrubba, 1997; Lord, 1997; Abu-Hejleh and AAttwooll, 2005; MeVay et al. 2005). The database contains the results of 49 rock socket tests (6 centrifuge rock socket tests), 40 plate load tests and 13 load tests on scaled footings. The database includes the results of experiments on circle and square footings of different sizes tested on various types of masses such as sandstone, claystone, shale, chalk, and basalt, The guu values of rock mass is initially obtamed or interpreted from load—displacement curves proposed by Hicany and Kulhawy (1988), Different parts of the employed database have been Used by other researchers to develop prediction models for quit (WASHTO, 2007; Paikowsky et al, 2010}. The descriptive statisties of the experimental data are given in Table 3, Fig. 2 presents the frequency histograms of the parameters used for the model development 32, Data classification Inorder to avoid overfitting, the available data sets are randomly, classified into three subsets: (1) training (learning). (2) validation (check), and (2) test subsets, The training set is used to fit the models and the validation set is used to estimate the prediction extor for model selection. Finally, the test set is employed for the evaluation of the generalization ability ofthe final chosen model ‘The training, validation and test data are usually taken as 50-70%, 15-25% and 15~25% of all data, respectively (Shahin and Jalsa o _ dy (kst “Vo g eat eae 2 Frequency 2 “0 ae oo) a (Ks‘AN. Alavi Sadrosedat Geoscience omters 72018) 31-85, Terfornance inden Mathematical denon Correlation colceat BY ne are ~ nase = ERS The aa outputs forthe Mh outpa pThe clclted usps fore uu Te number a sales The averae fhe aca up The 2005; Alavi et a, 2011). Inthe present study, 80% ofthe data sets are taken for the training and validation processes (61 data vectors forthe training process and 21 datasets asthe validation data), The remaining 20% of the data sets are used for the testing of the ob- {ained madels, 4. Development of LGP-based models Several runs are performed with different combination of the input parameters to obtain the best models for the prediction of 4 The LGP parameters are changed for each run. The param- clers are selected based on both some previously suggested g 1s z g Predicted g,y (ksf) by LGP ‘300 1000~—«1500~—~—2000 Experimental quy(ksf) (b) 2000 Predicted ,y,(ksf) by LGP 2 72000 30010001500, Experimental g,(ksf) values (Francone, 1998-2004; Baykasoglu et al, 2008: Alavi and Gandomi, 2011; Alavi et al, 2011) and making several pre- liminary runs to check the overall performance of the algorithm. The modeling process is controlled via evolutionary parameters such as popuilation size, probability of crossover, 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 Thus, three optimal population sizes of 500, 1000 and 2000 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. Hence, to prevent decreasing the speed of algorithms, two optimal values (64, 128) are considered for the maximum program size as trade- offs between the running time and the complexity of the solu- tions, Two values (10, 20) are set for the number af demes. It is worth mentioning that demes are semi-isolated subpopulations that evolution proceeds faster in them in comparison to a single population of equal size (8rameier and) Banzhaf, 2007), Different arithmetic operators and mathematical functions are utilized to formulate the LGP-based models, The termination terion 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. There are 3x 2x 2 x 2 ~24 different combinations of the parameters. All ‘Frequency 00 0 Relative error of all data (LGP 1) “0 ° o 0 2 ° 200 Relative error of all data (LGP 2)* AH. Ala Sarssdatj Geoscience Fo Wodel Saini index Data MAE 35a ones ab of these combinations were tested and 10 replications for each, ‘combination were carried out, This makes 240 runs for each of the combinations of the predictor variables. To evaluate the fitness of the evolved programs, the average of the squared errors is used. The Discipulus computer program is used to implement the LGP algorithm (Conrads et al,, 2004), The pa- rameters used to evaluate the performance of the models are shown in Table 4 ‘The best LGP programs obtained at the ené of the training process are given in C++ in Appendix A. These programs can be Lar 1 ra Work) R095, Residual 1LGP 2 (is Wark) Retoe RMSE=7854, MA () 7 (208) 51-38 run in any C environment. The resulting code may be linked to the ‘optimizer and compiled or it may be called from the optimization routines {Deschaine, 2000}. In order to facilitate the use of the derived codes via hand calculations, they ate simplified and converted into functional representations by successive re- placements of variables (Oltean and Grosan, 2003). The optimal LcP-based formulations of qu are as follows with different pa- rameters, namely LGP 1 (Eq, (3)) and LGP 2 (Eq. (4) ducer = O45{MR-=245)-0255 (0354.6 + (05q,02¢-5?)) ® 5 ue .cra= Gooden 0999) Res RISA U2, MAE 109%, pralueso2E-M carer an Kita (1988) ‘Test No. ‘Test No.‘AN Alo Sarosedat Geoscience Pontes 7 ( Fig. 4 shows a comparison and the relative error between the ‘measured and predicted qu, values by the LG models forthe entire data, 5. Results and discussion 5.1 Performance analysis and validation Ditferent statistical indices should be considered to evaluate the performance of a numerical cortelation, Smith (1985) proposed the following criteria for judging the performance of a model based on several observations: + if a model gives [Rj > 0.8, a strong correlation exists and the model is suitable, Im addition, the error values should be considered in all cases, (Alavi et al, 2017), The performance measures for the training validation, testing, and al data are summarized in Table 5. tean be observed from fig. 4 and Table 5 that the LGP models, with R > 08 and low RMSE and MAE values, are able to predict the target values vith an acceptable degree of accuracy. As itis, the model (LCP 2, Ea (4)) developed using m. s, qu $/B, and ¢ has a better performance that derived using RMR. +. qu 5/B. and 4. addition to the performance indices described, the values ofthe residuals (standard eror between the observed and predicted values) are visualized in Fig 5, In order to check whether the calcu lated R really exists between the observed and predicted datasets or not, the p-values at 5% significance level were also obtained and presented in this figure. Furthermore, this figure presents a comparative study between the results obtained by proposed models and those provided by the well-known models of Carter and Kuilhawy (1988) and Goodman (1989). As can be observed from Fig. 5, the proposed formulas notably outperform the existing models, Besides, itcanbe seen thatthe p-values for the LGP models are approximately equal to 0. This indicates that there isa sound correlation between the observed) and predicted data ses. tis worth mentioning that ‘most of the existing models are derived based on traditional stais- tical analyses (eg. regression analysis). The major limitation ofthis type of analysis is that the structures of the models are designated after controlling only few equations established in advance. On the other hand, the best equations generated by the LGP technique are etermined after controling numerous linear and nonlinear pre- liminary models. For instance, the proposed design equation (LCP 1) is selected among a total of 560,892,763 programs evolved and evaluated by the [GP method during the conducted 240 runs. 52. Sensitivity analysis As discussed before, the effect of all considered parameters (ie, RMR, 5, m, qu, SJB, and ¢) on qx is well understood. Herein, a sensitivity analysis is conducted to provide a more in-depth un- derstanding of the contribution of these important parameters to the prediction of guy. For the sensitivity analysis, the frequency values of the input parameters are obtained, A frequency vahie equal to 100% for an input indicates that this input variable has been appeased in 100% of the best thirty programs evolved by LGP (Francone, 1998-2004; Alavi et al, 2011; Gandomi etal, 20118}, Herein, the average and maximum impact values ate also deter: mined, The average and maximum impact values, respectively, show the average and the maximum effect of removing all in- stances ofthat input from the thirty best programs of the project. ‘The greater the value, the more impact removal bad. The sensitivity analysis results are summarized in Fig 6. As expected (Paikowsky et al, 2004, 2010), Fig. 6 indicates that both LGP-based models en TeFrequeney Average Impact wMaximum Impact = Percentage %) RRR SB rm 7 3 ‘Frequency Average Impact w Maximum Tnpoct Percentage (%) 8 SB om . ° are more sensitive to qy and 5/8 compared to other variables. The good agreement between of the results of the LGP sensitivity analysis and those reported by various researchers (Carter and Kulhawy, 1888; Goodman, 1989; PaikowsKy et al. 2004, 2010) guarantees the robustness and applicability of proposed models, 6. Conclusion This paper aimed at developing new nonlinear LGP-based ‘models for the estimation of the que of shallow foundations on non-fractured rock masses. A comprehensive database collected from the literature is used for the madel development. The optimal LGP-based models are selected after several assessment pro- cedures. The validation of the models is verified with different criteria, The parametric and sensitivity analyses are conducted to evaluate the robustness of the models. The results indicate that the proposed models provide precise estimations of que The R valies of GP 1 model on the taining, validation and testing data are equal to (0.85, 0.86 and 0.95, respectively. The corresponding values for LGP 2 model are equal to 0.97, 0.96 and 0.98, respectively, Moreover, the p-values at 5% significance level are approximately equal to 0 for both of the models. Considering the better performance of the second model (LGP 2, Eq, (4)}, m seems to be a more efficient parameter than RMR for the prediction of guy. Both ofthe devived ‘models have notably better performances than the existing tradi- tional models, In addition, che LGP models simultaneously incor- porate the effect of both the quantitative and qualitative parameters, Unlike other intelligent methods such as ANN, FIS and ANHIS, LG? method provides simplified equations that can be readily used for the design purposed via hand calcating Furthermore, the results of the LGP analysis provide transparent programs, in this case C}-+ codes, for further use in the bearing capacity prediction, as well as optimization purposes Appendix A. Supplementary data Supplementary data related to this article can be found online at ntps|/dXxdos.org/ 10.1016) gs.2014.12.008,8 ‘AM Noi Saressadt / GeascenceBromters 7 2015) 81-88 References ‘Msi0.200 RD Be Des Stitins ot ASO, Won, ‘Avge, N, Ato, WI), 2005, Colors’ Al Load Tests on Drie Shalt Sostcted in Wek ack Symsest ae Fre Needs Repo No COT 2605-4 September Caorao Deparment of Tansporttion Reseach Branch, Aa’ Dany R isha 6 Mein RT. 200, tion of Ute ‘iio ttelgence Arle I 20490, p10 [N-Anan Ay aes, LD-2010,Asippert wor machine agorthm to ass i laies and cel peed in hetrogenots tesco Sgsecrng Ge- sae Canon, A, 201, Arahat for formulation of ectechnical engineer ystems. Engineering Computations, 283} Aa AHL Aves M, Ganéom. AH. Mizsossins NLR 2271 Formultion of fow mbes ung ya compa ta Crain ‘Aeamatula, WOkdy Zaha, AL 2012 Hew charge preiion in compo Channels wing inar gene prosramming, Jornal of Hysclony sS4-a55c {taker EN. 1965, Companson of caisson laud tests on chcag harepan Deed Pics an cons ake JER (e6) Pre, Sern a the ASCE Notional Eenvention, ASCE Reston Va. pp. 28-10 Santa W. Nowéin Kees fe Raeore,F 1998 Genetic Programming: an Tneducion on the Atami Evoluon ef computer Progam and is sayogis A cus H, cane it Osbk 1, 208, Preon of ‘and tensile strength of limestone via genetic programming. Expert Systems ‘eth Apoiations 39 (12) 23 ‘ecer DE 96, ughtenth Canadian Geoteenical Callum: imitate design ter ouseasons Pat Development forthe Natal ung Coe of Canad, onaan Geotechnical Jounal 3 (6) 9841007 Bispon BL fo0n Bearing Capa of #CselyJoted Rock PhD. Dissertation Insitute of Tetnolony Geer Bowles E996 Foundton Ness Desig, th @ Mcrawil Ine, New ‘erie Computation 51) 1-26 “ arama ania W, 2007 Linear Genetic Programming Springs Since Busines Meda, New York und] ad A 170, Te lad drain havi of ile Ca a Sa and) oratory messuements is Proc conf an inst ‘estizations in Solr and Rocks, Nay 1313, 959, Sesh Geotechnical 5 ‘ny London pp. 5, Cabslae AP. Cele A. 7005, Modeling dining fate and shear modula of Cabalae AE. Cevk, A, 20098. Genetic progranming-isedaenuaton relation Ship, applaton of cent eartquakesin Turkey. Computers & Cosceces Canals Ceotcneal Sciey, 2006. Conan Foun Manat Monten quebe) othe ‘anahe Rayksapt, A iH, 2009, rein of cmmpressive ane tele Stengh of Canntp bara wa neural ntwor nd gee expression Pe {aming NeualCotspuing and Appian 1 (6) 1051-104 ‘eotecncat otra 342) 2300040, carr atu, F888 Analy nd Design of Fondation: Sacked ito Wook Repr: No Ls, tne Sate Hectne Engineering Research Cope. {ation an etic Power Research ste New York cag tee BIW Hang cil 01 ated eo aig Highway 18 in Alishan, Tawa. Engineering Geology 118 (3~4), 61—14 ‘hen A Dats jaa Man 82013 Buckling behavior a sno web for compose weed Fost wih clatlly testuned loaded edge under Canoes spc te: Sy of Cnpotte ces fu age rogcamming Based on AIM Learning Technology. Rezister Machine Learning ‘as, Sua, 2008 Undine aera lad 7 stifeal neural nework. Computers and Geotechmics 33 (8) 454459. sta neural netwark Engieering Geology 1003-4) 142=15, Descnine Li, 2000 Unng Cnet Prem f Deveep 2 Sulton Woda a Wate nae Scene opts (rat Te Rpt Technologies. Colorae, USA Lieto ‘ation by integrating fas optimization cechniques with machinecode—basee, Goodman, RE, 1969. inlcoduction to Rock Mechanics, second john Wiley & ‘ull, H- 2013, On the prediction of shear wave velocity at lca ste of strong can SN at ete yt set nd id Ra tt ac ct ao ie we a Bak SA ct adn ta a Hung. We, Hwang, C.Liou, in, VS. Yang. HL. 2012, Medeling aguier-system eS A aan ty. a Rene teat one anne tard ee ws ued we ey Et iol i sm cee attend ‘Chi-chi earthquake on the subsequent raingallinduced landsliées. Engineering pp. 1949-1956, ~ rock, Journal of‘AN. Alo & Sadrossedat/ Geoscience omers 7208) 31-85 9 "ext Toyo, aan Im Proceeding of 2M fnternatonal Workshop on Rock Foundations pp 449-455, Mallar Dy. 1977 Pca, session 1— chalk: roeedings ICE Conference ‘on Pes n Weak ck Pp 177-100 " Dried Shas in Florida Limestone, 2006 Geotechnical Research in Progress ‘Meeting fod Departnent of raspataton. August 160%. Foi. a ana, Sonoda, Mason, 905 Beating capac of aft ck sen ahh barca tide ed ad lpn ee ‘ken, Goin, C, 2003 A compan of several near gene programming cigs Coipiex yess 14 (4) 1-29, acinn Be amparrs Ke Suchet, RP, 2008. Ute beating cacy pe Computers ané Geotechie 35. 33-48, . resin: Onl, Kes Chena CON i 200 SCI Reper Sir las and Resistance factor Design (HPD) fr Deep Pounce Darcton Research Board af the Natonal Aces Faizowsg 5, Com M- Len K Aloys KAmaty5- Mugena.R, 2010 NCHRP ‘Report G3 LED Design snd constrain of Sal Foundatons for Hah Fan CP. Tat HC. 201. Improving semi-enpicel equations futinate bane acy of alow foandatons sing sl computing paloma Engineering pplestons of Are tteligence 36 0) 47boaeT et FIN. Ture RM, 1972 Hite soluoas forthe design aod abays of Tocksockeee pls. Canadian Geotecisl aural 16-48] et ®IN. Ter RM. 1980 End earn an sock with pater refereace co Sandstone. Inv Prcteings of the Intemational canerence On Sutera Foundations on Rock, Syne, sala a1. p, 11-180 adhalibnan, Reg. CF 180 Inada: Behar a roksocketed ples Schmit. M, Lipson, H. 2008, Comparison of tree ané graph encodings as function ‘a problem compen in: Geneue a Evluonary compurationConerence {ctecoen pp erate. shahnazar H, Totunchian, MA. 2012, Prediction of ultimate beating capacity of “halo feundaens en cohesnless sls a1 evluvanay approach. XSce Sinan A 8 19 The za of grec Sri opin "duced sellin catbonate lay sng ara neral network, Geomecats ne Engineering 1 (0, Sth, GX. 1985. Pabublity and Statsts i Cv Fagineesing. Colles, onde {eosynthets erations 12 (6). 321-332 Sowers, Ge, 1978. Inteductory Sel Mechanis and Foundtions: Geetechnic Engineer, fourth ed. MacSlas, Nev Yoke Teraath Tab. Roce defects and Toads on tonne sspport In: Proctor RV. ‘inte, TL (Bes) Rock Tunneling with Steel Support Commercal Shestng "home, Ft, Cpa of ts dx Sie Feeding ‘sal HC. Tyan YW. Wu. YW. Lin ¥H. 2012 Determining lamae bearing capac ‘of shallow fundatiens using a genet programming system Nedra Computing {nt Appian pnd ott 004 sosao12-152 Wad We Burland], 1968 Aesent of the deformation properties of ointed eck the mass lo Proceedings ofthe Intemational symposium of Rock Mechanses Octobe: 22nd-2st-24th Magee span pp. 3584 ‘Webb: IL 1876 Behavior of bed pes in weathered dbase. Ceetechaique 25 ‘wise. T, 2009, Global Optimization Algorithms ~ Theory and Application. Ger ‘lay eweise de (selepubished [nlne) Aalbers \wullams A" 9H0- Design and Pecirmance of Pes Sockted Into Weak Rock ‘PhD, Dissertation Monash Unveay, Melboune, Asta, ‘ison Lc 1975, ess a bred and deven ples rtaceus mudsane at Fort “Biabe, Sut Ai. eotecnigue 25 (1) 5-13 Yaghocby F. Aveta, A Sleiman, R. 200% classification of caréie abr. ais ising seice fextres of hen rate varantysgnal World Apple Scenes Jota 8 (11, 19471584 ‘aghauy Fata, A ghoul M, 2010 An arta dasiicaton method ‘chine Sse ease n-TEMBE Preceding, rng SC pp 1928-1931, Yaghowby F Ayatollah Yaghouby. Ne 2012 Robust gene programing ase eteion of tal Gluten using RK intsvas Expert Systes 29.2 Yaohasby Aytlahi A. Yashouby, M. Ala, AN, 20108 Tverd automat detection, of ta Hatin a Hybrid compotatonal approach, Comptes Siology aod Nedeine 40-12), S19-950. ‘oa Ke Cohen. OcskoglF, Nearing HA, Ack S, 2008, Prediction