The Vehicle Road Load Problem - An Approach by
            Non-Linear Modeling
             Glenn D. Thompson
        United States Environmental
              Protection Agency

-------
                   ABSTRACT

     When vehicle exhaust emission tests or vehicle
fuel consumption measurements are performed on a chassis
dynamometer, the dynamometer is usually adjusted to simulate
the road experience of the vehicle.  Therefore, a method
to accurately characterize the road experience of the vehicle
is a prerequisite for accurate dynamometer testing of this
nature.  Commonly used methods for determining the road
characteristics of the vehicle are manifold vacuum measure-
ments, drive line torque measurements, and the coast down
technique.
     In the usual coast down technique, the vehicle road load
force is calculated from a numerical derivative of the speed
versus time measurements.  This study significantly improves
the coast down technique by eliminating the differentiation
of the speed-time data while maintaining the ability to treat
the forces on the vehicle as a general quadratic function of
vehicle speed.  A model equation for the forces on a freely
decelerating vehicle is constructed;  this equation is then
analytically transformed into an expression for the speed of
the freely decelerating vehicle as a function of time.  This
intrinsically non-linear equation is then computer fitted to
the speed versus time data recorded from the vehicle.
     The results of road load measurements by this method are
compared with road load measurements on the same vehicle by
the drive shaft torque method.

-------
Introduction






     Vehicle exhaust emission measurments are usually



performed on a dynamometer so that the vehicle engine may



be exercised in the laboratory at diverse power outputs.



The power demand of the engine is usually determined by



first adjusting the dynamometer power absorber to simulate



the power versus speed or force versus speed characteristics



of the vehicle;  and then the vehicle is operated at the



desired simulated speeds.  A method to accurately char-



acterize the road experience of the vehicle is therefore



a prerequisite for accurate dynamometer testing of this



nature.



     Common methods for determining the road characteristics



of a vehicle are manifold pressure measurements, drive line



torque measurements and the coast down technique.  Manifold



pressure measurements are the easiest to conduct, but various



engine emission control techniques, notably exhaust gas



recirculation and fuel injection have made this approach



unsatisfactory for many vehicles.  Drive line torque measure-



ments are quite satisfactory, but require placing strain



gages or torque meters in the drive line of each test vehicle.



The deceleration or coast down technique is well suited for



measurements on diverse vehicles since only instrumentation



to record the vehicle speed as a function of time is required.

-------
                        -2-
The Deceleration Technique









     The deceleration technique is perhaps the oldest



method of determining the forces acting on an automobile



as a function of the vehicle speed.*•*•*  The concept



is to determine the rate of deceleration of the vehicle



and then, knowing the mass of the vehicle, the road



load force may be calculated by Newton's, second law.



     Experimentally it is not practical to measure the



deceleration as a function of velocity directly, however,



it is easy to record speed as a function of time.  It



is possible to differentiate the recorded velocity versus



time data to obtain acceleration versus time, and then



to use the velocity versus time data to map the accelera-



tion as a function of velocity.  This is the most common



approach but it is theoretically undesirable for two



reasons.  The non-analytical differentiation process is



inherently noise sensitive.  This can be a problem when



attempting a least squares fit to the differentiated data,



and can lead to a falsely large linear term in the least



squares fitted curve.  Also since the acceleration must



be derived from the velocity, the .initially random errors



may no longer have a normal distribution.  This makes many

-------
                       -3-
of the usual statistical tests invalid or of question-



able validity.



     Theoretically it is a better approach to assume a



model or form of the acceleration versus speed equation



and then perform analytical operations on this equation



to convert it to the form of a speed versus time equation.



This expression may then be directly fitted to the



velocity versus time data.  This approach has been used



to investigate road load force equations of the usual



constant plus a velocity squared term form.[2]  However,



tire datat3} indicate that terms linear in speed should be



included in the force equation and this method may be



generalized to include a linear term.








The System Energy







     An expression for the system energy is constructed



to develop a generalized expression for the force acting



on the vehicle.  The total energy of the decelerating



vehicle is the sum of the translational kinetic energy



and the rotational kinetic energy of any vehicle compo-



nents in rotational motion.  For all mechanical components



of the wheels and drive train, the rotational velocity is



proportional to the vehicle velocity, arid the energy of

-------
                      -4-
the system is
     E = 1/2 mv  + 1/2 vli «i                     (1)
where:
     m  = the vehicle mass
     v  = the vehicle speed
     Ii = the rotational inertia of the ith rotating
          component
     oti = the proportionality constant between the
          rotational velocity of the i*1*1 rotating
          component and the vehicle speed
Differentiating equation  (1) with respect to time; and

comparing the resulting power expression with the similar

time derivative of a purely translational system; the

generalized force on the  system may be expressed as:
     f = (m + Ameq)  v                               (2)
                    dt
The term Ameg is equal to  irfl.«i2 and is the mass equiva-

lence of the inertia of the rotating components.



The Model Equation for the Vehicle Deceleration



     For ideal wind free conditions the acceleration of

-------
                        -5-
a freely coasting vehicle is assumed to be a polynomial


function of the velocity:
     dv                          ~
     — = A(v) = -(a0 + aiv + a2v^)                   (3)
     The ag term is usually identified as representing™]


tire losses, hpwever, both tire losses and mechanical drive


train losses probably appear in both the ag and a-^v term.


The a2y2 term of (3) is identified as the aerodynamic


term and the velocity in this term must be the vehicle


air speed.  Therefore, if any ambient wind exists,


equation  (3) must be written as:
     A =  [aQ t a^v +  a2(v - W-S)2]                   (4)
where:




     S = a unit vector in the direction of vehicle travel




     w= the wind velocity vector.




This analysis only considers the effect of the component


of wind in the direction of vehicle travel.  Also equation

-------
                         -6-
(4) is only valid when  (v - w-8) £0.  If the component of

wind speed in the direction of vehicle travel exceeds the

vehicle speed, then the aerodynamic drag term must be

replaced by an aerodynamic driving term by changing the

sign of a2-  The effect of cross winds can be treated in

this analysis if sufficient aerodynamic informatipn is

available about the vehicle to allow the cross wind

effect to be approximated as a simple polynomial

expression of the vehicle speed.

     A term must be added to equation (4) to describe

the effects of grade.  The grade term is equal to g SinV

where 7 is the angle between the test surface and the

horizontal and g is the gravitational acceleration.

Since y is a very small angle the approximation of using

the grade is quite accurate.  Inserting this grade term,

equation  (4) becomes:
     A = -[an + gs + anv + a~(v - w-S)2]             (5)
where:
     g = a vector in the track direction of increasing
         grade with a magnitude equal to the product of
         the test surface grade and the gravitational
         acceleration.

-------
                       -7-
Expanding and regrouping, equation  (5) becomes:
     A = - ) [a0 + a2(ws )2 +  g-s] +


             [a-L - 2a2(W-S )]v  + a2v2 }              (6)
Integration and Inversion
     Using A = dv/dt arid separating variables  of
equation  (6):


                     dv
     _dt =
            [a0 + a2(W-S)2 + g-S] +  [ai  -  2a2(w-S)]v + a2v2   (7)
Equation  (7) can be integrated  in  closed  form.


                         V

     The  integrals of equation  (2) will depend  on the


relative  magnitude and  sign of  ag, a^, a2, w-S  and g-s.


The appropriate choice  for the  form  of the  integrals is


more easily seen if the variable transformation:
         2u -  [a-, -  2a0( w-s )]
     v=	—	±	_              -      (8)
                 2a2
and



          4a2[ a0 + a2(ws)2  +  (g-s)]  -  [a-^  - 2a2(W-S)]2   (9)
     B2 = 	

-------
                       -8-
are applied to equation  (7).  This yields:








     dt = du/(B2 + u2)                 ,             (10)








Because of the constraint  [v -  (W-S) ] > 0, u will always



be real and non-negative.  The B is not so strongly



restrained, B2 will always be real but it may be either



negative, positive or zero.








Case A;  B2< 0



     If B2 is negative the transformation:








     B2 = -D2                                      (11)








yields:








     dt = -du/(u2 - D2)                 ,          (12)








Now both u2 and D2 are real and positive, hence both u



and D are real.








     The integral of equation D-12 can take either of



two forms:

-------
                       -9-
         1       ,
     t = -[tanh  -"-(u/D)]* C,                        (13)
         D                 -1
or
                _i
     t = -tcoth  •L(u/D)]+ C2                        (14)
         D
The terms GI and C2 are constants of integration, deter-


mined by the initial conditions.




     The choice of equation D-13 or D-14 to represent  the


coast down data must depend on the physical possibilities


of the motion of the system.  If equation D-13  is used,


the magnitude of u cannot exceed the magnitude  of D  since


the argument of the tanh "•*- function is bounded between


+1 and -1.  But, if u is less than D, then equation  D-12


shows dt/du is positive.  Hence, considering the transfor-


mations 8 and 9, dv/dt is positive and the vehicle must be


accelerating.  Consequently equation D-14 describes  a


"coast-up" which can only occur if there is a grade  or


wind driving term of sufficient magnitude that  the vehicle


freely accelerates from its initial speed to some higher


steady state speed.




     If equation 14 is to be used to describe the coast


down data,u must be greater than D since the magnitude

-------
                       -10-
of the -argument of the coth "•*• function must exceed 1.



The physical interpretation of this can be seen by re-



writing equation D-12 as an expression for du/dt.








     du/dt = -(u2 - D2)                            (15)








If the vehrcle is initially at some velocity:where u > D,



the rate of change of u is negative hence the vehicle



will decelerate until u = 0 or u = D.  If the u = D



condition "occurs, du/dt vanishes and the vehicle will



remain at the velocity where u = D.  Physically this means



there must be a grade or wind .effect acting as a driving



term.  The vehicle will decelerate until the drag forces



are balanced by this driving term,, and the vehicle will



remain in motion at a constant speed.







Case B;  B = 0







     The case B = 0 is a special case that can theore-



tically occur, however in practice it will probably .never



be observed because of .experimental error.  It is pr»e-   '



sented since it is the transition between two classes of



motion, and numerical analysis difficulties may Occur in



this area.



If:  B = 0

-------
                        -11-
     dt = -du/u2                                   (16)








and:








     t = 1/u + C3                                  (17)








where C-j is the constant of integration  to  be  determined



by the initial conditions of the  system.








Case C; B2 > 0








     In the case B2 >  0 equation  D-9 may be integrated



     in the form:








     -t =  (1/B)[tan -1 (u/B)] + C4                 (18)








or:








     t =  (1/B)[cot ~1  (u/B)] +  C5                 (19)








Again C4 and C$ are constants  of  integration.








Equations  18 and 19 are equivalent.  This may  be shown



by  inverting equation  18 to yield:

-------
                        -12-
     u = B tan t-B(C4 + t) ]                        (20)



Using the trignometric identity relating the tangent

and cotangent functions, equation 20 may be written as:



     u = B cot (Bt + V2 + BC4)                    (21)



Inverting equation 19 yields:



     u = B cot [B  (t-C5)]                          (22)



Comparing equations 21 and 22 these equations differ only

by the sign of the constants of integration, and by the

phase angle of »/2.  Either equation may be used to

describe the vehicle coast down.  However, since the B2< 0

case requires the coth -1 solution, equation 22 will be

used for the case B^ > 0 because of the convenient parallel

nature of the cot and coth functions.


                                                  »
     Applying the original transformations 8 to 22 and

solving for v yields:
         1    '           '•'•'-          	
     v = - [B cot (Bt - C5) - ax/2 + a2(w-5)]      (23)
         3. o            •         .      *•-."•

-------
                      -13-
where:







     C5 = - cot -1  [ (2a2v0 +







The velocity VQ is  the initial vehicle  speed  at  t = 0.







Applying the same operations  to  equation  D-13 yields:





          1

     v = —   [D coth  (Dt  - C2) - a,/2 + a2(ws)]   (24)
         •a                  *•     J-
         a2
and
     C2 =  -  coth   1  [(2a2vQ  +  a.|J
      Either  equation  23  or  24  must be fitted to the speed



 time  data  by the  method  of  least squares.   The values of



 the coefficients  a^,  a-.,  and a~ will determine which



 expression,  23  or 24  is  the appropriate form to describe



 the data.





      *

 The^fcitting  Technique







      The  least  squares method requires the sum of squares:
      S = £[V;L ~ V(ti)]                            (25)

           i

-------
                        -14-
be minimized with respect to the fitted parameters.  To



simplify the fitting process, the change of variables:
     b0 = a0 + a2(w-S )2 +  g
     b2 = a2                                       (26)



     b3 = the constant of integration







is made in equations 23 and 24.  The equation for the



vehicle velocity may now be written as







     v = v(t,b)                                    (27)







where b designates the four parameters of the fitting



process, bQ, b, , b2, and b.,.







     Equations 23 and 24 are transcendental and intrin-



sically non-linear in b; hence, an iterative approxima-



tion method must be used somewhere in the fitting process,



The equations expressing the necessary conditions for a



minimum of equation 25 may be constructed analytically,



but then this set of equations would have to be solved



numerically by some iterative approximation method.  An



alternate approach, which will be used, is to approximate

-------
                       -15-
v(ti,b) by a Taylor series in b, about some trial point,


and then iterate the entire fitting process until the


system converges on a value for b.






     Expanding v about the trial point b^; and retaining


only the terms through the first derivative:
     v(tfb) = v(t,bu) +
                        k=0
         avitd
                                                   (28)
Substituting equation 28 into equation  25
     S(b) =
v(ti|b°) -E
   1      k=0
$
                                                            (29)
If S is to be a minimum,  the partial derivative  of  S  with


respect to each of  the  fitting parameter must  vanish.


Applying  this condition to  equation  29;
                        K—U
                                                dv(t.
                                                           = 0   (30)
                                                         b°
Equation  30  represents  a  system of  4  equations  linear


in  the  four  unknowns  b .   This  system can  be  solved  for


the b, which  are  then  used  as  a new  trial band the  process


repeated  until convergence is obtained.

-------
                     -16-
  In the analysis of vehicle coast down data, the



non-linear cotangent and hyperbolic cotangent models



gave rise to a system which exhibited convergence



difficulties.  Convergence sometimes occured slowly,



requiring many iterations, each one decreasing S(b), the



sum of squares, but the solution vector  would not  stabi-



lize.  Sometimes the system would pscillate, each new



correction  complementing the correction of the previous



step, while S(b) both increased and decreased.  Finally,



divergence sometimes occurred with S(b) increasing



with each iteration.



     Several techniques have been devised to circumvent



these problems.  A "relaxation technique,"which amends



the correction vector  by  halving it whenever the



correction increases S(b) and doubling the correction



vector when it reduces S(b) is well known    .  Another



technique    computes a weighted average of the correction



vector found from the linearization technique and the



correction vector found from the steepest descent.



Reparameterization of the model is yet another technique



which often helps, if a convenient reparameterization can



be found.  Unfortunately, however, there is as yet  no



general convenient way to determine a priori whether a

-------
                       -17-
proposed reparameterization will improve the situation.



Although any of these techniques may provide the additional



help needed to make a balky system converge, if they fail



to help; no further insight about the problem or its



possible solution is provided.



     The approach used in this analysis attacks one of



the fundamental problems and provides insight into the



fitting of non-linear models.  Since the theoretical con-



vergence of these systems can be proven, failure to con-



verge in practice must partially arise from a deficiency



in the tools used to compute the solution.  Specifically,



'computers use only a small subset of the rational numbers,



and this must be considered when any algorithm-computer



combination is used to approximate the solution of a



difficult mathematical problem.



     In virtually all cases where convergence difficulties



are encountered, the linear system of equations which must



be solved at each step in the iteration is  "poorly" conditioned,



Simply  stated, a poorly conditioned linear  system is one



in which propagated round-off error dominates the computed



solution.



     A  precise definition of the condition  of a system of



linear  equations and an explanation of how  this directly



affects the convergence of the  system can be expressed

-------
                        -18-
in terms of the norms of the system vectors  and matrices.
The norm of a vector is defined as
                                                   (31)
for p = 2 this norm is merely the  length  in  the  Euclidian
sense, when p = oo  equation  31 designates  the  infinity
norm, which is equivalent to the magnitude of  the  largest
component of the vector.  Analogous  to  Lite norm  of a
vector, the norm of a matrix is define! as:

          UAH = max || A x II
where                                              (32)
          llxll = 1
This norm can be viewed  as  the longest  image of  the unit
sphere under the mapping associated  with  A.  Norms defined
in this manner" have the  properties.
          llcMH = | cNlMil
          ||M  N || < IIMII HN||                     (33)
     Consider a system of  linear  equations  represented
byAx=r where A is a non-singular matrix.   If  the  components
of rare not known precisely, we would  like  to know  the
effect of  such uncertainty on  the solution  vector  X .
Let 5 r represent the uncertainty in r.   Since;

-------
                      -19-
                                               (34)


            Ax = r




then:




           A (x +  »x)  =  r + sr                 (35)




and consequently:






            *x =  A^r                         (36)







To establish a bound of the relative errors 2X/X, the norms



of ix and r are constructed from equations 34, 35, 36, and



the property ,of norms 33.  After rearranging terms:
             nsx n     ,,   ,  , r1  "&r"
             -i.   < HAlMiA || 	            (37)
             nx ii                I, r n
The quantity  II A II-HAH   is defined as the condition



of the matrix A and is designated by COND(A).



     A system of equations with a large condition number



for the coefficient matrix is said to be poorly conditioned.



     Such poorly conditioned systems are difficult to solve.



The propagation of round-off error in the many multiplications



and additions involved in computing the solution vector can



lead to large inaccuracies.  These problems are minimized



by using Gaussian elimination with pivoting and scaling



 [ 8 ] , augmented by iterative improvement. *•  *

-------
                        20-
These algorithms are capable of solving systems of linear



equations to machine precision accuracy.  However,



equation 37 demonstrates such techniques are' not suffi-



cient to produce accurate solutions if the right hand



side vector of the system is uncertain.



     Considering equation 30, the effect of the system



condition on the analysis of the coast down data may be



directly understood by defining:
     r . =
           o


     bk "  bk
a .,  —
dv(ti/ b)
                         dv(t±,b
                                                (38)
the system of equations depicted by 30 may then be written



in the same form as equation 34:




                  Ax= r                       <39>




where the .components of r,x and A are the r.  x ,  .and a.,
                                            3    K      JK


respectively.  As the iterative solution proceeds,



the components of  r tend .to vanish since X  must tend to



zero if the system is converging.  However the indivi-



dual elements which comprise the sums of each component of



r remain about the same size.  Thus the precision of



each component r. is given by the accuracy of the largest



term of the sum.  If the calculations are performed on

-------
                        -21-
a t-digit machine using a base number  system  N,  the

precision of r- may be expressed as:
       ~  -t
   *r. = N   max
       
-------
                        -22-
makes excellent initial progress toward convergence, then



suddenly slows down- and ceases to make additional progress in



reducing the sum of the squares of the errors.



     Considering the right aide of equation 41;  the



condition of.A is controlled by the choice of the model and



its parameterization, as is ||D||.   The "noise" of the data



is represented by IIv. - v(t.,t>.) ll   and can only be reduced



by improving the data collection instrumentation.  The only



computational hope for improving convergence is to decrease



N   by increasing the precision in computing  r.



     In analyzing the vehicle coast down data obtained from



a Ford station wagon, the system of equations resulted in a


                                                          + 6
coefficient matrix with a condition number of the order 10



This poor condition caused severe convergence difficulties.


Since the solution has been programmed in FORTRAN on an



IBM 370/168 computer the precision of  r could be increased



by using double precision variables in the computation.



This immediately improved the percentage of data sets which



converged, however, results were still unsatisfactory.  This


computer had an additional mode of arithmetic^ called extended



precision, which was then used to compute  r  to a precision



of 28 hexadecimal digits.  This resulted in a dramatic



improvement, and successful convergence was obtained for



almost all data sets.

-------
                        -23-
Data
     A 1971 Ford station wagon was used as the test



vehicle.  This vehicle was equipped with an incremental



digital magnetic tape recorder to record vehicle speed,



drive shaft torque and time, all at one second intervals.



The vehicle speed signal was generated by a rotary shaft



encoder in the vehicle speedometer cable.  The torque



transducer was a rotary transformer torque meter installed



in the vehicle drive shaft.  The time signal was generated



by a quartz crystal clock.



     The skid pad of the Transportation Research Center



of Ohio, in East Liberty Ohio, USA, was used for the test



track.  This facility is a multilane concrete straight



track with large  turnaround loops at each end.  Approx-



imately 1 kilometer of this straight track has a constant



grade of 0.5% and this section was used for all measure-



ments.



     The test vehicle and tires were warmed up for 30



minutes at a steady speed of 80 km/hr.  Steady state



speed and torque data were then collected for 1 kilometer



in each direction of travel on the track at approximately



9, 13,  18, 22 and 26 m/sec.  Following these measurements,



speed versus time data were recorded for ten coast downs,



five  in each direction of the track.  The initial speed

-------
                        -24-
of the coast down was approximately 26 m/sec., and the



terminal speed was 9 m/sec.  The vehicle was then weighed



with all instrumentation and test personnel still on board.



     Wind velocity and direction were recorded manually



during the test period.  A mean value for the component of



wind velocity in the direction of the test track was



calculated for the time the steady state torque data were



collected and for the time required for the coast downs.



The mean values were 0.64 m/sec. during the steady state



torque measurements and 0.85 m/sec. during the coast



downs.



     Each coast down was analyzed by the non-linear curve



fitting process to yield the four  b coefficients of



equations 26 .  An estimate of the standard error of each



coefficient was also calculated.  Equations  26  were then



used to calculate the  a coefficients from the  b , the mean



wind velocity and test track grade.  A weighted mean



was then calculated for aQ, a, and a.  using the reciprocal



of the square of the estimate of the standard error as a



weighting factor.  The resulting set of acceleration coef-



ficients were then multiplied by the total vehicle mass to



yield a set of coefficients for predicting the force on



the vehicle by the equation:





          f = f0 + fxv + f2v2                  (45)

-------
                         -25-
The force coefficients calculated by the coast down method

were:

         ffl =  4.4 x 102    nt

         f, =  -9.3         nt/(m/sec.)

         f2 =  9.9 x ICf1   ,nt/(m/sec.)2



     The data from each set of steady  state torque measure-

ments were converted to vehicle driving force measurements

using the rear axle ratio of the vehicle and the measured

rolling radius of the tire.  A regression analysis was

then used to remove any acceleration effects and to yield

an improved estimate of the vehicle driving force at  the

mean speed of each data set.  The estimate of the mean

force and the mean speeds were then fitted to the equation:
         f =  fQ + h  S +  fjv +  f2(v - W-S)            (46)
where
        h = a vector  in the  track direction of  increasing
            grade with a magnitude equal  to the product
           • pf  the  test track  grade,  the  gravitational
            acceleration,  and  the mass of the vehicle
      Equation    46  is  analogous  to  equation   5   .   The  f

 coefficients describe  the  vehicle road  load  on  a level

 track with  no wind.  The values  of  these  coefficients

 calculated  from the drive  shaft  torque  meter data are:

-------
                          -26-
       fQ = 3.6 xlO2 nt



       f-  = -2.0  nt/ (m/sec.)
                      nt/ (m/sec* ) 2
     The force versus speed, curves using the coefficients



calculated by each method are plotted in figure 1.








Conclusions








     The drive shaft torque measurements indicated slightly



higher calculated road load forces, than did the coast down




method^although the shape of the curves are in very good agree-



ment.  The difference in the predicted road load forces near the



center of the measured speed range is less than 4%.  There are



several reasons why the road load calculated by the steady



state drive shaft torque method might be expected to be



slightly higher than the road load calculated by the coast



down method.  For example, the energy dissipated in tires



transmitting power is greater than energy dissipation in



freely rolling tires.  Because of fuel consumption, the



measured vehicle mass used in the coast down calculation was



slightly less than the true mass during each coast down.  Also



the vehicle mass during the drive shaft torque measurements



would be greater than the vehicle mass during the coast



downs.  However, the standard estimate of error for the



coefficients calculated by each method would indicate that the



4% difference is not statistically significant.

-------
                      -27-
     The coast down technique can yield road load results



very similar to drive shaft torque meter measurements



when the added flexibility of a linear term is included.



The results cannot be expected to be identical since



the deceleration technique is intrinsically a transient



experiment, however good precision may be obtained with



either method.  Since coast down measurements can be con-



ducted with less instrumentation of the test vehicle than



drive shaft torque meter measurements, the coast down



technique is particularly useful for measurements on a



diverse class of vehicles.

-------
                                  VEHICLE ROAD LOAD

                               1971 FORD STATION WAGON
Force
 (nt)
      900
      800
      700
      600
      500
      400
                           coast down method

                           drive shaft torque meter method
                   10
12
 i
14
16      18      20      22
24
26
                                            SPEED  (m/sec.)

                                               FIGURE  1

-------
                    Bibliography
 1.   Hoerner,  Sighard,  F.   Aerodynamic  Drag,  The Otterbein
     Press,  Dayton,  Ohio,  1951.

 2.   White,  R.A.  and Korst,  H.H.,  "The  Determination of
     Vehicle Drag Contribution from Coast-Down Tests",
     Society of Automobile Engineers 720099,  New York,
     N.Y.,  1972.

 3.   Curtiss,  W.W.,  "Low Power Loss Tire",  Society of
     Automobile Engineers 690108,  New York,  N.Y., 1969.

 4.   Marks,  L.S.   Mechanical Engineers'  Handbook, McGraw-
     Hill Book Company, New York,  N.Y.,  1941.

 5.   Draper, N.R. and H. Smith  Applied Regression Analysis,
     John Wiley and Sons,  New York, N.Y.,  1966.

 6.   Marquardt, D.W.  "An Algorithm for Least Squares
     Estimation of Non-Linear Parameters"   Journal of the
     Society for  Industrial and Applied Mathematics, Vol.2,
     1963,  pp. 431-441.

 7.   Guttman,  I.  and Meeter, D.A.   "Use of Transformations
     on Parameters in Non-linear Theory, I.   Transformation
     to Accelerate Convergence in Non-Linear Least Squares",
     Technical Report No.  37, Department of Statistics,
     University of Wisconsin, Madison,  Wisconsin, 1964.

 8.   Forsyth,  G.E.  and C.B.  Moler  Computer Solution of
     Linear Algebraic Systems, Prentice Hall, Englewood Cliffs,
     New Jersey,  1967.

 9.   Ibid.

10.   Scarborough, I.E.   Numerical Mathematical Analysis
     The Johns Hopkins Press, Baltimore Maryland, 1966.

-------