Stephen L. Chiu Rockwell Science Center Thousands Oaks, California 91360 Fuzzy MopEL IDENTIFICATION BASED ON CLUSTER ESTIMATION ApstRacT We present an efficient method for estimating cluster centers of numerical data. This method can be used 10 determine the number of clusters and their inital values {for initializing iterative optimization-based clustering algorithms such as fuzzy Comeans. Here we use the Cluster estimation method as the basis of a fast and robust algorithm for identifying fuzzy. models, A benchmark problem involving the prediction of a cha- otic time sertes shows this model identification method compares favorably with other, more computationally intensive methods. We also illustrate an application of this method in modeling the relationship between aw ‘tomobile trips and demographic factors. © 1994 John Wiley and Sons, Ine. IntRopucTION Clustering of numerical data forms the basis of ‘many classification and system modeling algo: rithms. The purpose of clustering is to distil natural groupings of data from a large data set, producing a concise representation of a system’s behavior. In particular, the fuzzy C-means (FCM) clustering algorithm (Dunn, 1974; Bez dek, 1974; Bezdek ct al., 1987) has been widely studied and applied. The FCM algorithm is an iterative optimization algorithm that minimizes the cost function SS ttn where n is the number of data points, ¢ is the number of clusters, x, is the kth data point, », is the ith cluster center, jz, is the degree of mem. bership of the kth data in the 7th cluster, and m is a constant greater than I (typically m = 2) ‘The degree of membership 41, is defined by Journal of Intelligent and Fuzzy Systems, Vol 2, 267-278 (1994) © 1994 ohn Wiley & Sons. Ine a 2 (ia) Starting with a desired number of clusters ¢ and an initial guess for each cluster center ¥,, i= 1,2,....€, FCM will converge to a solution for y, that represents either a local minimum or a saddle point of the cost function (Bezdek et al., 1987). The quality of the FCM solution, like that of most nonlinear optimization problems, de- pends strongly on the choice of initial valu (ie., the number © and the initial cluster centers) ‘Yager and Filev (1992) proposed a simple and fective algorithm, called the Mountain Meth- od, for estimating the number and initial loc tions of cluster centers. Their method is based on making a grid of the data space and computing a potential value for each grid point based on its distances to the actual data points; a grid point with many data points nearby will have a high potential value, The grid point with the highest potential value is chosen as the first cluster cen: ter. The key idea in their method is that once the first cluster center is chosen, the potential of all grid points are reduced according to their dis: tance from the cluster center. Grid points near the first cluster center will have greatly reduced potential. The next cluster center is then placed the grid point with the highest remaining potential value, This procedure of acquiring new cluster center and reducing the potential of sur- rounding grid points repeats until the potential of all grid points falls below a threshold. Al- though this method is simple and effective, the computation grows exponentially with the d mension of the problem. For example, CCE 1064-1246/94/050267.ing problem with 4 variables and each dimension having a resolution of 10 grid lines would result in 10° grid points that must be evaluated. We present a modified form of the Mountain Method for cluster estimation. We consider each data point, not a grid point, as a potenti: center. Using this method, the number of effee tive “grid points” to be evaluated is simply equal to the number of data points, independent of the dimension of the problem. Another advantage of this method is that it eliminates the need to specify a grid resolution, in which trade-offs between accuracy and computational complexity must be considered. We also improve the com: putational effcieney and robustness of the origi nal method via other modifications Although clustering is generally associated With classiication problems, here we use the cluster estimation method as the basis of a fuzzy model identification algorithm. ‘The key to the speed of this model identification algorithm is that it docs not involve any iterative nonlinear ‘optimization; in addition, the computation grows only linearly with the dimension of the problem. We use a benchmark problem in chaotic time series prediction to compare the performance of this algorithm with the published results of other algorithms. We also show an appli timating the number of automobile trips gener- ated from an arca based on its demographics. cluster tion: es Custer Estimation Consider a collection of m data points (2y.%.--..%,) in an M-dimensional_ space Without loss of generality, we assume that the data points have been normalized in euch dimen- sion so that their coordinate ranges in cach dimension are equal, i.e.. the data points are bounded by a hypercube. We consider exch data point as a potential cluster center and define a measure of the potential of data point x, as where and r, is a positive constant. Thus, the measure of potential for a data point is a function of its 268 cw distances to all other data points. A data point with many neighboring data points will have a high potential value. The constant r, is effectiv ly the radius defining a neighborhood; data points outside this radius have little influence on the potential. This measure of potential differs from that proposed by Yager and Filev in two ways: (1) the potential is associated with an actual data point instead of a grid point; (2) the influence of a neighboring data point decays exponentially with the square of the distance instead of the distance itself. Using the square of the distance eliminates the square root operation that otherwise would be needed for determining the distance itself After the potential of every data point has been computed, we select the data point with the highest potential as the first cluster center. Let Xx} be the location of the first cluster center and, Pt be its potential value. We then revise the potential of cach data point x, by the formula P&P, — Pye Htemnil 8) where and 1, is a positive constant. Thus, we subt aan amount of potential from each data point function of its distance from the first cluster The data points near the first cluster center will have greatly reduced potential, and therefore are unlikely to be selected as the next cluster center. The constant r, is effectively the radius defining the neighborhood that will have measurable reductions in potential. To avoid obtaining closely spaced cluster centers, we set r, to be somewhat greater than r,; a good choice is r= 15 When the potential of all data points have been revised according to eq. (3), we select the dat point with the highest remaining potential aay the second cluster center. We then further reduce the potential of each data point according to their distance to the second cluster center. In general, after the kth cluster center has been obtained, we revise the potential of cach data point by the formula center PE P,— Pre thn where x is the location of the kth cluster center and P; iy its potential valueIn Yager and Filev’s procedure, the process of acquiring new cluster center and revising po- tentials repeats until PEcePt where ¢ is a small fraction, The choice of & is an important factor affecting the results; if ¢ is 100 large, too few data points will be accepted as, cluster centers; if ¢ is too small, too many cluster centers will be generated. We have found it difficult to establish a single value for e that works well for all data patterns, and have the fore developed additional criteria for accepting: rejecting cluster centers. We use the following criteria’ if Pp> EP; Accept x7 as a cluster center and continue. else if Pr the first matrix on the right hand side of eq. 9 is constant, while the second matrix contains all the parameters to be optimized. To minimize the squared error be- tween the model output and that of the training data, we solve the linear least-squares estimation problem given by Eq. (9), replacing the matrix oon the left hand side by the actual output of the training data. Of course, implicit in the least- squares estimation problem is the assumption that the number of training data is greater than the number of parameters to be optimized. Using the standard notation adopted in most literature, the least-squares estimation problem of eq. (9) has the form 7 ast pea] Peada Pen || Gr rs (9) AX=B where B isa matrix of output values, A is a constant matrix, and X is @ matrix of parameters, to be estimated. The well-known pseudo-inverse solution that minimizes || AX — Bl" is given by X=(A'A) ‘AB However, computing (AA) is computationally expensive when (A'A) is a large matrix (ATA is e(N +1) x c(V+1)); numerical problems also arise when (ATA) is nearly singular. We useanother well-known method for solving for X, a procedure often referred to as recursive least- squares estimation (Astrom and Wittenmark, 1984; Strobach, 1990). This is a computationally efficient and well-behaved method that deter- mines X via the iterative formulae Xie =X F SoosQis(bhy~ BeX,) (10) SthaS, i+alSa,., =0.1, -1 ay where X, is the estimate of X at the ith iteration, Sis a c(N +1) x e(N +1) covariance matrix, a? is the ith row vector of A, and b' is the ith row vector of B. The least-squares estimate of X corresponds to the value of X,. The initial con- ditions for this iterative procedure are X,=0 and S, = y/, where / is an identity matrix and y is a large positive value. To summarize, the model identification meth- od consists of two distinct steps: (1) find cluster centers (0 establish the number of fuzzy rules and the rule premises, and (2) optimize the rule consequent. Neither of these steps involve non- linear optimization and both steps have well- bounded computation time. In step 1, the bulk of the computation time is consumed by evaluat- ing the initial potential of each data point. Each subsequent iteration to select a cluster center and subtract potential consumes the same amount of time as evaluating the potential of one data point. Assuming the number of cluster cen ters that will be obtained is much less than the total number of data points, we can accurately estimate the computation time of step 1 based on the number of data points alone. However, the number of cluster centers found in step 1 affects the computation time of step 2 linearly, because the number of parameters to be optimized grows linearly with the number of clusters. Hence, we can determine the computation time of step 2 only after step 1 is completed. Although the number of clusters (or rules) is automatically determined by this method, we should note that the user-specified parameter r, (ie. the radius of influence of a cluster center) strongly affects the number of clusters that will be generated. A large r, generally results in fewer clusters and hence a coarser model, while a small r, can produce excessive number of clusters and a model that does not generalize well (i.e., by overfitting the training dat ‘Therefore, we may regard 1, as an approximate specification of the desired resolution of the model, which can be adjusted based on the resultant complexity and generalization ability of the model. Resutts We will first apply the model identification meth- od to a simple 2-dimensional function-approxi- mation problem to illustrate some of its prop- erties. Next we will consider a benchmark prob- lem involving the prediction of a chaotic time s \d-compare the performance of this method with the published results of other meth- ‘ods. Lastly, we will show an application model- ing the relationship between the number of au- tomobile trips generated from an area and the demographics of the area FUNcTION ArPRoxiMATION For illustrative purposes, we consider the simple problem of modeling the nonlinear function For the range [~10, 10], we used equally spaced y values to generate 100 training datapoints Because the training data are normalized before clustering so that they are bounded by a hy- percube, we find it convenient to express r, as & fraction of the width of the hypercube; in this 0.25. Applying the clust estimation method to the training data, 7 cluster centers were found, Figure 1 shows the training data and the locations of the cluster centers. example we chose r, —— ning Figure 1. Comparison of training data with unoptim: ized (th order model output Fuzz Move. IwENTiFicaTion 271Figure 2. Degree of fulfillment of each rule as a function of input y Figure 1 also shows the output of a Oth order fuzzy model that uses constant 2* as given by the z coordinate of each cluster center. We see that the modeling error is quite large. Because the clusters are closely spaced with respect to the input dimension, there is significant overlap be~ tween the premise conditions of the rules. The degree of fulfillment « of each rule as a function of y (viz. eq. 4) is shown in Figure 2, where it is evident that several rules can have high firing strength simultaneously even when the input pre- cisely matches the premise condition of one rule ‘Therefore, the model output is typically a neu- tral point interpolated from the 2 coordinates of strongly competing cluster centers. One way to minimize competition among closely spaced cluster centers is to use the fuzzy C-means definition of 1, viz. eq. (1). When the input precisely matches the location of a cluster center, the FCM definition ensures that the input will have zero degree of membership in all other clusters. Using the FCM definition, the degree of fulfillment 1 as a function of y is shown in Figure 3 for a typical rule. We see that wis 1 when y is at the cluster center associated with the rule and = LW ATE Figure 3. Degree of fillment of a typical rule as a function of input y, based on the fuzzy C-means definition of pw. 272 cm ry WW Figure 4. Comparison of training data with unoptim- ized Oth order model output, for the case where inference computation is based on the fuzzy C-means definition of drops sharply to zero as_y approaches a neigh- boring cluster center. The output of the Oth ‘order fuzzy model when the FCM definition of adopted is shown in Figure 4, It is evident that the modeling accuracy has improved significant- Iy; in particular, the model output trajectory is now compelled to pass through the cluster centers, Although using the FCM definition of can improve the accuracy of unoptimized 0th order models, we note that the 4 function and the resultant model output trajectory have highly nonlinear “kinks” compared to that obtained with the exponential definition of .. These kinks tend to fimit the ultimately attainable accuracy when optimization is applied to the model We now apply least-squares optimization to 2}. For illustrative purposes, we consider the two cases: (1) 21 = h,, and (2) 2} = G,y +h; In the first case, we retain the assumption of a Oth order model, but now optimize the constant value assigned to z$. In the second case, we assume a Ist order model and optimize both G, and h,. Table I shows the root-mean-square (RMS) modeling error resulting from the differ- ent extents of optimization. Table I also shows the effects of using the exponential definition of versus the FCM definition. ‘Table L. Comparison of RMS Modeling Error for Different Extents of Optimization and for Different Definitions of RMSei Optimization RMS error using EXP using FCM ZFunoptimized — 0.180 0.119 Zeh 0.084 0.098 zt 0.010 0.015y iing a NG : SNM ma Figure 5. Comparison of training data with opti- mized Oth order model output ‘Using the exponential definition of z generally results in more accurate optimized models. In what follows, we will not draw any further com- parisons between using the exponential defini- tion versus using the FCM definition, but present only the results obtained from using the ex- ponential definition. The output of the optimized Oth order model is shown in Figure 5 and the output of the optimized Ist order model is shown, in Figure 6. The consequent function 2* = G,y + ‘h, of each rule is shown in Figure 7. Although our method can be used to approxi- mate a function from uniformly distributed data points as in the above example, the method is best suited for identifying models from repeti ous experimental data, where there are repeti: tive behavior patterns that create distinct clusters in the data space. In the next example, we will examine a chaotic time series that does create such a data set, Chaotic Time Senies PREDICTION ‘We now consider a benchmark problem in model idemtfication—that of predicting the time series o nig a model Figure 6. Comparison of training data with opti ‘mized Ist order model output, ie rol ouput Figure 7. Consequent funetions for the Ist order model generated by the chaotic Mackey—Glass differen- tial delay equation (Mackey and Glass, 1977) defined by ay = 0280 a +xGoa) = 0.1K) ‘The task is to use past values of x up to the time 1 to predict the value of x at some time f+ At in the future. The standard input for this type of prediction is N points in the time series spaced $ apart, i.c., the input vector is y = (x(t (N 1S)... x= 28), x= 8). x9}. To allow comparison with the published results of other methods, we use T= 17, N=4, S=6, AT=6. Therefore, each data point in the training set consists of x(0= 18), x(¢= 12), x(0 = 6), 2(0), (0+ 6)} where the first 4 elements correspond 0 the input variables and the last clement corresponds. to the output variable. We will compare the performance of the model identification method with the Adaptive-Network-Based Fuzzy Infer- ence System (ANFIS) proposed by Jang (1993) as well as other neural network-based and poly: nomialitting methods reported by Crowder (1990). The ANFIS algorithm also. produces fuzzy models consisting of the Takagi-Sugeno type rules. After specifying the number of mem- bership functions for each input variable, the ANFIS algorithm iteratively learns the parame- ters of the premise membership functions via backpropagation and optimizes the parameters fof the consequent equations via Tinear least- squares estimation. ANFIS has the advantage of being significantly faster and more accurate than many pure neural network-based methods. Fast ‘computation speed is attained by requiring much Furry Move. loewrinicarion 273Jess tunable parameters than traditional neural networks to achieve a highly nonlinear mapping, and is also attained by optimizing a large fraction ‘of the parameters via linear least-squares estima- tion, thus further reducing the use of back- propagation, Because ANFIS typically has much less tunable parameters than a traditional neural network, it can avoid the pitfall of over-fitting the training data, thereby achieving excellent generalization ability. Comparison of ANFIS with our model identification method is pa ticularly interesting because of the similarity in model structure. ‘The performance of ANFIS Provides a good indicator of the added benefit and computational burden that accompany non- linear optimization of the Takagi-Sugeno type rules. For the Mackey-Glass time series problem, we used the same data set as that used in Gang 1993), which consisted of 1000 data points ex- tracted from f= 118 to 1= 1117. The first 500 data points were used for training the model, and the last 500) data points were used for check- ing the generalization ability of the model. As mentioned previously, the cluster radius r,s an ‘approximate specification of the desired res- olution of the model, which can be adjusted based on the resultant complexity and generali- zation ability of the model. To illustrate this principle, we applied the cluster estimation method with different values of r, ranging from 0.15 10 0,5 (ie., the cluster radius was varied from 0.15 to 0.5 times the width of the data hypercube), This produced models of varying size, ranging from 69 rules to 9 rules. For each ‘model, the consequent equations for a Ist order ‘model were then optimized. Figure 8 shows the number of rules generated, as well as the model- ing errors with respect to the training data and checking data, as functions of the cluster radius. i lee 2 oon jad : nd Chena 5 Figure 8. Model size and prediction error as func~ tions of cluster radius, 274 eww Sample. Figure 9. Comparison of training data with model ‘output for Mackey-Glass time series, We see that the error with respect to the training data and crror with respect to the checking data begin to diverge when the cluster radius is less than 0.3, showing that the model is over-fitting the training data as the number of fitting param- ters becomes too large. We use the results at 7, =0.3 (a model with 25 rules) as the basis for comparison with other algorithms. Figures 9 and 10 show the model output evaluated with respect to training data and checking data, respectively. We see that the model output is indistinguishable from both training and checking data. The clus- ter centers and consequent equations for this model are listed in Appendix A. ‘The modeling error with respect to the check~ ing data is listed in Table II along with the results from the other methods as reported in Jang (1993) and Crowder (1990). The error index shown in Table IL is a non-dimensional error defined as the RMS error divided by the standard deviation of the actual time series Giang. 1993; Crowder, 1990). Comparison be- tween the various methods shows that the cluster estimation-based method can provide approx mately the same degree of accuracy as the more complex methods. Only ANFIS produced a re- ‘Samplene, Figure 10. Comparison of checking data with model ‘output for Mackey-Glass time series.Table I. Comparison of Results from Differ Methods of Model Identification” Method # Training Error Data Index Cluster Estimation-Based 500 ois ANFIS 500 0.007 ‘Auto-Regressive Model 500 0.19 Cascaded-Cortelation NN 500 0.06 Back-Prop NN 500 0.02, 6th-order polynomial 500 0.08 Linear Predictive Method 2000 0.55 = Rows 2 and 3 are from Jang (1995); the last 4 rows are from Crowder (1990), sult that is more accurate than the cluster estima- tion-based method. The ANFIS model refer enced here uses two membership functions for each input variable; for the 4-input Macke Glass problem, this leads to 2*= 16 fuzzy parti tions in the input space, and thus 16 rules. This ANFIS model has 24 parameters optimized via back-propagation (3 parameters for each premise membership function) and 80 parameters opt: mized via linear least-squares estimation (5 pa rameters for each consequent equation). The cluster estimation-based method docs not in- yolve any nonlinear optimization but the 1: sultant model has 125 consequent parameters optimized via linear least-squares estimation. Identifying the Mackey-Glass model via the ANFIS algorithm (coded in C) required 1.Shours on an Apollo 700 series workstation Gang 1993), while the cluster estimation-based algorithm (also coded in C) was able to idemtfy the model in 2 minutes on a Macintosh Centris, 660A (68040 processor running at 25 MHz) ‘Trip Genenarion MoDELine ‘We have applied the model identification method to estimate the number of automobile trips gen- erated from an area based on the demographics of the area, Five demographic factors were con- sidered: population, number of dwelling units, vehicle ownership, median household income. and total employment. Hence, the model has 5 input variables and 1 output variable. ‘Demographic and trip data for 100 traffic anal- ysis zones in New Castle County, Delaware, were used; this data was transcribed directly from (Kikuchi et al., 1994). Of the 100 data points, we randomly selected 75 as training data and 25 as checking data. Using r,=0.5, the Pili etiei tbe Eee Figure 11. Comparison of checking data and model ‘output for trip generation modeling. model identification algorithm generated a Ist order model comprised of 3 rules. The computa- tion time was 2 seconds on a Macintosh Centris 660AV. The cluster centers and consequent equa- tions for this model are listed in Appendix B along with the data set. A comparison of the model output with that of the checking data is shown in Figure 11. The average modeling error with respect to the training data was 0.34 and that with respect to the checking data was 0.37, indicating the model generalizes well. The fact that we can accurately cover a 6 dimensional data space with only 3 rules attests, to the particular advantages of using the Takagi- Sugeno type rules. ConcLusion We presented a cluster estimation method based fon computing @ measure of potential for each data point and iteratively reducing the potential of data points near new cluster centers. The computation grows only linearly with the dimen- sion of the problem and as the square of the number of data points. This method can be used to estimate the number of clusters and their locations for initializing iterative optimization- based clustering algorithms such as fuzzy C- ‘means, or it can be used as a stand-alone approx- imate clustering algorithm, Combining the cluster estimation method with a linear least-squares estimation procedure pro- vides a fast and robust algorithm for identifying fuzzy models from numerical data, Fast compu- tation and robustness with respect to initial pa- rameter values are achieved by avoiding any form of nonlinear optimization. Robustness with respect to noisy data is achieved by the data Furzy Moog loewriricarion 275,averaging that takes place in both the cluster estimation and least-squares estimation proce- dures. The cluster center selection criteria also avoid engendering a rule from a few erroneous outlying data points. Although there exist even simpler and faster model identification methods based on look-up tables (Wang and Mendel, 1992) and nearest neighbor clustering (Wang, 1993), they are sensitive to noisy data and prone to generating a rule from a single outlying data point. Compared with more complex. model identification methods, our method can provide similar degree of accuracy and robustness with respect to noisy data while significantly reducing computational complexity ‘The author thanks Jyh-Shing Jang at MathWorks Ine. for providing the data set for the Mackey-Glass benchmark problem. Appenpix A: MacKEY-Giass Exampie ‘The input coordinates of the 25 cluster centers: obtained in the Mackey-Glass example are yy = (0.9479 1.0754 1.1302 0.6686 1.0659 Lissa 1.1135 0.8264 0.6652 0.9671 0.7423, 0.8575, 0.9930 1.2001 1.2177 0.4807 0.4978 1.0240 0.7635 Lisi2 1.0840 0.6038 1.2985 0.8753 0.6581 0.8067 0.9713 0.6702 1.1165 1.1382 1.1514 1.0133 1.0728 0.8227 0.7594 0.6652 1.0007 0.9652 0.8946 1.2001 0.8575 0.5256 L1si2 0.5383 1.2551 0.7295 0.4473, 1.0902 0.6038 0.9713 0.7005 0.9814 0.6539 1.2329 1.1393}; 1.0481}: 0.7805}; 1.1912}; 1.0942}: 0.6748}; 0.8227); 0.9736); 1.1350}; 0.6462); 0.8946); 1.0007}; 0.9060) 1.2551}: 0.5456}; 1.2063}; 0.4978}; 0.7307}; 0.7687}; 0.4473}; 0.9814) 0.6865}; 1.0441}; 0.9690}; 13114}: The corresponding output equations are =| -08077 0.8586 1.0985 —Masedly + 1.6873; [1067 saat v6s09 —n6id3}y+ 2941 = [0.7946 1.0950 0.5198 0.1406} + EE =] 0308 W463 ~0.2007 O.tlonly + 0313903855 0.2478 Longs] + — LM 0.8059 —0.1730)y + 0124 O16 605} + 0.2308 03994 02059}y-4 02751, =N8965 OAD —OTRAS]y + 22601 HUIMS 0.0730 O.5958)y-+ 1.0357: S097 0.0372 WasoRhy +1305 Wabi —OX6L 6957} + —O.0812 4.7959 077 6178}y + 0.0061 “US888 OTHLE USOH|y + UL444y, 02107-00846 Hai]y + 0.9531 AT 0.0097 1.9832] + 0.0493; fn 4634 0.479 O.6150]y+ 1-202 =| 02107 7624 05270 UsaIa]y+ 0.050 =] O4us8s 0.71% Ha10GOTB]y= Lams, 0.3792 0889102783 Oat] y+ LSS Pap 027m 0.025 020% asi0l]y + 0.0530, PHA [14782 0386407371 —024O]y+ 2.958; 28 = [poss ons 0.1202 “0. 1231}y+ 1305 O.ATI2 R240 01140 117589] + -O.OTOL = M6685 1.9702 WOLLD—L3s74)y + 0.8855 where y is @ column vector of the input values: [x= 18), x= 12), x(¢= 6), x(0)} ApPeNpix B: TRIP GENERATION Mopvetine The input coordinates of the 3 cluster centers ‘obtained for the trip generation model are ¥} = (1.5070 0.6570 0.7060 15.7350 0.6360}; {3.1160 1.1930 1.4870 19.7330 0.6030}; {0.0440 0.0240 0.0210 9.3400. 0.8500): The corresponding output equations are [04337 sod —ust26 -uo106 orusy 1.1066; SE=[=0.7690 0.1857 2.1078 LT 1.4631)y + 8378; Ties “19202 7.5781 O8SS UsDTy +0519; where y is @ column vector of the input values: {population, number of dwelling units, vehicle ownership, median income, total employment] The trip generation data are given in Table B-L All numbers are expressed in units of a thousand.Table B. ‘Trip Generation Data ‘Table BA. (continued) Popula- Dwelling Vehicle Median Total No.of Popula- Dwelling Vehicle Median Total No. of tion Units Owner Income Employ- trips tion Units Owner Income Employ: trips ship ment ship ment “Training Data 0038 0032 GUT TTS 16.499 8460 2487 0788S 240] 1.399 3198 0.148 0.062 0.051 9.375 3.176 2.250 1.089 0.430 O.7BEARGTL OKT L.A O24 0.139 0.068 «6.473 0.193 0317 0.293 0.117 OBS 27.220 0.108 OSI, 0.222 0.204 0.064 7.738 3.450 1.907 -2.TOL_ 0.836 STB 1.686.322 2.10 0.023 0.013 0.039 10.125 7.064 4.746 1028 0453 0.657 15.944 1.812 4.431 OAL 0.192 0.093 7.824 0.179 O37S 3.562 1.36) LSM 20.083 2.000 5.049 0.132 0.040 0.017, «S.378 0.627 OSH 4.247 1.651 2.526 20.93 0.265 3.138 0.084 0.024 0.021 9.340 0.850 0.768 2674 1,022 L649 26.034 6.109 2.015 0.066 0.011 0.004 11.705 0.859 0.605 V381 OSIS 0.961 31.280 3.363 2.302 0.337 0.250 0.203 17.406 2.208 LA3l 0494 0731 188731286 1.714 1465 0471 0436 17.340 178 2.210 SOS 1.460 25.000 0.477 2.388 1357 0433 0.303 8.260 0843 0450 0.188 «0.212 13.251 VIB 0.951 407L LMS 1339 11.590 2.801 3.206 07S 1.567 17.099 0.710 2.373 2.277 0.78 0569 18-451 1239 3071 1.035 1.562 20.454 0.599 2.670 3837 1.395 1347.15 27596 2.358 0.732 1.285 4.159.592 1.749 1590 0.762 0.787 11.708 1,304 2507 0.857 L495 21.843 0.073 1.723, BMG L193 1A87 19.733 2385 GS7T 2.122 -3.710 26.349 -0.536 4.02 1368 0.517 0.708 25.681 1.883 3.007 1175-2014 25.889 0.684 3.151 0.937 0.488 0.639 21.691 1242 © XSSL L674 DIB 19.401 1.584 4.698 319 2.333 1.766 13.403 0.870 2.946 0.229 098.14 3HAS LOZ 5.127 1877 0.763 0917 18.750 1.565 2483-2731 OSL SAS MSDS DATS STS L872 O83) 1.065 27.661 0.667 L868 3.510 1.008 2.273 41.679 0.235 2.8od 3.107 13691561 16.843 L146 2.772 3.390 1.073, 2.329 48440 0.316 3.074 1535 0.605 0.784 16.544 0.399 1.313 2.246.809 LATT.31.303_ 0.111.805 2326 0.981 0.612 11454 0.573. 1.a72 ae 3300 0.925.775 11.560 0.942 1.852 Checking Data 2223 0.8590 0.970 16.400 2.528 3.177 te eee 3274 06s 1.008 13.193 0.750 2.037, Ay ue ee) es 1.490 0.592 0.685 13.796 1.797 2.037 a ee eae 3447 1.366 1.396 16461 0.705 2.349 MOBS O.DIS DIZ 12.624 3.2 3.035 1176 1.494 18.963 O47 2.169 SOB .290 0.523 28.835 DIL 0,003 0.008 6.002 6.036. 1.687 1.272 pee oe 1,600 0.506 0.260 5.369 L385 1.467 eo rd Bu ay 0.398 O.ISL 0.132 8.159 0.625 0.688, 3.005 1.098 2.288 eeu 239 0.901 0542 S603 1.27 Ta3s ALIS DOH OUTS oe 2.698 1.268 0.541 6,663 0.758 1.548 Ga) el ae 414 0.144 0080 3.119 0245, 0.387 oa es 0.037 G.015 0.005 9.659 1.277 0.966 pnp ell Lue 2.76) 1333-1200 13498 1.064 2.996 ae pee ce ee 4476 1.656 2.719 26.266 OSIT 3.631 ey ee i el 1507 0.657 0.706 15.735 0.636 1.446 oe need oie 1179 0.434 0.784 32.328 ae ee) es aed 1727 690 4.232 30.573 pie alas ee Be ee Hes 1.093, 0.095 019 0.860) aos iaTi 2iicelt nasal 1859 0738 0.396 1.346 Hee eG ae) 2.083 0292 0777 1.251 naa as ovale 20M 0.384 9083 9.401 20.529 fe rsy tose fovea asia) 202 0.759 LAT 258.619 1.259.911 Toe Gait. eae alsa 3219 1337 L782 15.466 0.795 3.221 oie latat | ise oral 2.280 M68 1391 27.008 OS212.327 iaiat one; yar) 0s) 0992 0296 032 WAN 11371468 Fuzzy Moos toenriricarion 277REFERENCES Astrom K, Wittenmark B (1988): Computer Cont- rolled Systems: Theory and Design. Englewood Ciitfs, NI: Prentice Hall Beadek J (1974); “Cluster validity with fuzzy sets.” J Cybernetics 33): 8-71 Bezdek J, Hathaway R, Sabin M, Tucker W (1987): “Convergence theory for fray e-means: Counterex- amples and repairs.” The Analvsis of Fuczy Ine formation, Bezdek J (ed). CRC Press, Vol. 3, Chap. 8 Crowder RS (1990): “Predicting the Mackey-Glass time series with cascade-correlation learning.” In Proc. 1990 Connectionist Models Summer School, Carnegie Melion University, pp. 117-123. Dunn J (1974): “A fuzzy relative of the ISODATA process and its use in detecting compact, well sepa- rated cluster” J. Cybernetics 33): 32-57 Jang JSR (1983): “ANFIS: Adaptive-network-based fuzzy inference system.” IEEE Trans. on Stems Man & Cybernetics 23(3): 665-685 Kikuchi $, Nanda R, Pesincherry V (1994): “Estima- tion of trip generation using the fuzzy regression method.” 1994 Annual Mecting of ‘Transportation Research Board, Washington, D.C Mackey M, Glass L (1977). “Oxilation and chaos in physiological control systems” Science 197: 287-289 Powell MID (1987): “Radial basis functions for mul \atiable interpolation: A review.” In Algorithms for Approsimaion, Mason J. Cox M (ed), Oxford: Clarendon Press, pp. 143-167 278 cw Strobach P (1990): Linear Prediction Theory: A Mathematical Basis for Adaptive Systems, New York: Springer-Verlag Sugeno M, Tanaka K (1991): “Successive identiiea- tion of a fuzzy model and its applications to predic- tion of a complex system.” Fuzzy Sets and Systems 42(3): 315-384, Takagi T. Sugeno M (1985): “Fuzzy identification of systems and its application to modeling and control.” fEEE Trans. on Systems, Man & Cyber: netics 18: HO= Wang LX, Mendel IM. (1992): “Generating fuzzy rules by learning from example." /EFE Trans. on Systems, Man & Cybemerics 22(6). Wang LX (1993): “Training of fuzzy logic systems using nearest neighborhood clustering.” Proc. 2nd IEEE Int | Conf. on Fuzzy Systems (FUZZ-IEEE), San Francisco, CA, pp. 13-17. ‘Yager RR, Filev DP (1992): “Approximate clustering via the ‘mountain method.” Tech. Report #MIT- 1305, Machine Intelligence Institute, Yona College, New Rochelle, NY, Also to appear in IEEE Trans. fon Systems, Man & Cybernetics Yager RR. File DP (1993): “Learning of fuzzy rules ‘by mountain clustering.” Proe. SPIE Conf. on Ap- plications of Fuzzy Logie Technology. Boston, MA. pp. 246-254, Received January 1994 Accepted June 1994