Introduction
In nonlinear optimization problems, the design variables are to incorrect results. Fu, Fenton, and Cleghorn (1989) have
generally assumed to be continuous quantities. In many cases, given an example as to why this "common sense" approach
this is a valid assumption. However, very often integer vari- cannot be followed. In Fig. 1, the contour map of an uncon-
ables and discrete variables occur in engineering design. Ex- strained optimization problem is given, showing the continuous
amples of these types of variables are the number of gear teeth, and discrete optimum values. The continuous optimum so-
the value of a gear module, the specifications of rolled steel lution, indicated by point "A", is at (1.55, 1.4) with an ob-
members, the number of bars in a truss, nominal sizes of shafts jective function value of 0.0. The optimum solution to the
and other parts, the number of coils of a spring, the number discrete problem with discrete values as multiples of 0.5 is
of screws or rivets in a structure, etc. The presence of discrete indicated by point "B" at (0.5, 0.5) where the value of the
and integer variables along with continuous variables adds to objective function is 18.0. If the continuous optimum were
the complexity of the optimization problem. Since such situ- rounded to (1.5, 1.5), the objective function value would be
ations often occur in engineering design, an efficient method 40.0 and the amount of error would be quite high.
to solve this type of problems is of considerable use in optimal In the field of nonlinear mixed discrete integer continuous
mechanical design. programming (MDICP), a limited amount of research has been
Also, conceptual design optimization problems can be solved reported. Gisvold and Moe (1972) developed a method for
using a type of integer variables called "zero-one variables" nonlinear mixed integer programming problems by modifying
(commonly called "binary" variables) which can take on values the penalty function approach. Here, discrete requirements are
of either zero or one as the name suggests. Whenever there is
a choice to be made from a list of alternatives, zero-one vari-
ables can be used to choose the best alternative. More details
on the use of zero-one variables are discussed later.
In the linear optimization domain, solution methods have
been developed for integer and discrete problems. However,
in nonlinear optimization problems, no such parallel can be
drawn. An initial approach is to assume all variables to be
continuous and then round the result up or down to the nearest
integer or discrete value. However, this approach often leads
This problem was solved using the augmented Lagrange mul- 8PnwxD5N
g7(X)=-//+ -+ {P,n i)/K
tiplier algorithm combined with Powell's method. The starting Gd"
values for the variables and the optimum solution obtained as -1.05(yV+2)tf<0 (43)
well as the results obtained by Sandgren using Branch and
Bound method are shown in Table 3. gi(X) = {Pmax-Pload)/K-bw<Q (44)
The starting value for rp used in the problem was 0.0001
and the multiplication factor used was 1.8. The starting values In the above equations,
of Lagrange multipliers used were zero. S is the maximum allowable shear stress,
In this problem, the starting values used by Sandgren were
very close to the optimum and these values were used for Pmax is the maximum design load,
comparison purposes. The active constraint was found to be
gi(X) as opposed to gi(X) as found by Sandgren. Even still, dmin is the minimum wire diameter,
the method proposed here produced a better value of objective
function. In addition, seven more starting values were tried Dmm is the maximum outer diameter,
considerably farther from the optimum point and the same
result was obtained. The starting values tried were: (2.5, 2.5, 5pm is the maximum deflection under preload,
50.0, 120.0), (0.0,0.0,50.0, 120.0), (0.0,0.0, 100.0, 120.0), K is the spring correction factor
(5.0,0.0, 100.0, 120.0), (5.0,0.0,200.0, 120.0), (10.0, 1.0, = {AC- 1)/(4C-4) + .615/C, and (45)
200.0,120.0) and (0.0,0.0,200.0,200.0). In this formulation,
the inequality constraints are satisfied when they are negative.
In Sandgren's work, they were satisfied when they were po- C is the spring index = D/d. (46)
sitive.
The values of all the parameters can be found in Sandgren
Case Study 3: Spring Design Problem (Siddall, 1982; Sand- (1988) and the derivation of the equations used in formulating
gren, 1988). This is a problem involving discrete, integer, and ' the constraints can be found in Siddall (1982). The starting
continuous variables. The objective is to design a helical point and the final optimal solution along with comparison to
compression spring with the least volume. The spring is to be Sandgren's result are shown in Table 5. The starting value for
manufactured from music wire spring steel ASTM A228. This rp used in the problem was 0.0001 and the multiplication factor
means that the wire diameter can assume only discrete values used was 1.1. The starting values of Lagrange multipliers used
shown below in Table 4. The design variables are D (winding were zero.
diameter), d (wire diameter), and N (number of spring coils). As seen from Table 5, a better result has been obtained using
D is a continuous variable, cf is a discrete variable, and N is the Augmented Lagrange Multiplier algorithm compared to
an integer variable. Sandgren's Branch and Bound algorithm.