Mathl. Comput. Modelling Vol. 16, No. 1, pp.

57-74, 1992
Printed in Great Britain. All rights reserved
0895-7177/ 92 $5.00 + 0.69
Copyright@ 1992 Pergamon Press plc
COMPOSITE STOCK CUTTING
THROUGH SIMULATED ANNEALING
HANAN LUTFIYYA AND BRUCE MCMILLIN
Department of Computer Science, University of Missouri at Rolla
Rolla, Missouri 65401 U.S.A.
PIPATPONG POSHYANONDA AND CIHAN DAGLI
Department of Engineering Management, University of Missouri at Rolla
Rolla, Missouri 65401 U.S.A.
Abstract-This paper explores the use of Simulated Annealing as an optimization technique for the
problem of Composite Material Stock Cutting. The shapes are not constrained to be convex polygons
or even regular shapes. However, due to the composite nature of the material, the orientation of the
shapes on the stock is restricted. For placements of various shapes, we show how to determine a cost
function, anneahng parameters and performance.
1. INTRODUCTION
For many important practical or theoretical problems, the objective is to choose a best solution
out of a large number of candidate solutions, or the solution space. Such problems are typically
known as combinatorial optimization problems. A combinatorial optimization problem is formal-
ized as a pair (S, C), where 5 is the finite-or possibly countably infinite-set of configurations
(also called configuration or search space) and C a cost function, C : S -+ R, which assigns a
real number to each configuration. For convenience, it is assumed that C is defined such that
the lower the value of C, the better (with respect to the optimization criteria) the corresponding
configuration. The problem now is to find a configuration for which C takes its minimum value,
i.e., an (optimal) configuration ic satisfying
copt = C(i0) = @C(i),
where Co,, denotes the optimum (minimum) cost.
One common combinatorial optimization problem that arises frequently in applications is the
problem of allocating rectangular and/or irregular patterns onto a large stock sheet of finite di-
mensions in such a way that the resulting scrap will be minimized. This problem is common to
many applications in aerospace, shipbuilding, VLSI design, steel construction, shoe manufactur-
ing, clothing and furniture. This problem is commonly known as the stock cutting problem or
the 2D bin packing problem.
The stock cutting problem is an example of a large scale optimization problem. It is unlikely
to be solved by an algorithm whose run time is bounded by a polynomial function. This is
an example of an NP-hard problem [l]. Th is means that this problem requires a computing
effort that increases exponentially with the problem size. Since the stock cutting problem is of
practical importance, efficient approximation algorithms, namely, algorithms that do not produce
optimal but rather close-tooptimal solutions, have been developed. These include linear and
This work was supported in part by the National Science Foundation under Grant Numbers MIP-8969749 and
CDA-8820714, in part by the AMOCO Faculty Development Program, in part by the Manufacturing Research
and Training Center (MRTC), in part by the McDonneiI Douglas Corporation, and in part by the University of
Missouri Weldon Springs Fund.
We wish to acknowledge the progr amming efforts of Mr. Rob Zeigler and Mr. ChuI-Eui Hong.
Typeset by A&-T@
57
58 H. LUTFIYYA et al.
integer programming, dynamic programming, tree-search algorithms, and artificial intelligence
and operations research (AI/OR) integration [2]. These proposed approaches have restrictions
on the size and type of applications.
The paper is organized as follows: Section 2 discusses previous approaches to optimization of
stock cutting, Section 3 presents Simulated Annealing through the concept of statistical mechan-
ics. Section 4 describes the relevant parameters that control Simulated Annealing and Section 5
relates these parameters to the problem of stock cutting. Section 6 discusses the notion imple-
mentation of moving figures and Section 7 discusses the cooling schedule used. Experimental
results are reported in Section 8.
2. PREVIOUS METHODS
Linear programming methods have been extensively researched. This work has been done
by Gilmore and Gomory [3-51, Geoffrion and Marsten [6], Haessker [7] and Dyckhoff [8]. In
general this method involves a solution of the problem through development of mathematical
models consisting of an objective function that is to be minimized or maximized, and constraint
functions indicating the limitations on the allowed values of the variables of the objective function.
Both the objective and constraint functions are linear functions of independent variables. Any
model may be transferred to the following standard form:
minimize clrl + c2z2 + . . - + cn2,,
subject to
aliz + ~12x2 +. . . + alnzn = h,
a2121 + ~2222 +. . . + a2nz, = b2,
amlIl+Qm2~2++~mn2, =bm.
The stock cutting problem can result in a mathematical model consisting of hundreds, and
even thousands, of variables and constraints. The above mentioned work has concentrated on
finding special structural characteristics of the model and developing techniques that exploit
these structures. These methods have successfully been applied to a broad class of stock cutting
problems. There are, however, many real problems for which these methods are not appropriate
due to their structure or size. In many cases, this is caused by special restrictions. For such
problems, other methods-often heuristic ones-are used.
Dynamic programming is one heuristic often used. Dynamic programming is an algorithm
design method that takes a model of the problem and converts it into a series of single stage
problems. This transformation is intuitively based on the principle that an optimal set of de
cisions has the property that, whatever the first decision is, the remaining decisions must be
optimal with respect to the outcome which results from the first decision. The difficulty is in
determining quickly the optimal decisions. Otherwise, the problem degrades into enumeration
of the decisions and determining which is the best. This has exponential complexity. Studies of
dynamic programming approaches to the stock cutting problem have been done by Beasley [9,10]
and Sarker [ll].
Another class of heuristics often used are tree-search methods. This method enumerates all
possible solutions in a tree like organization. Many different tree organizations may exist for the
solution space. Heuristics exist for finding the solution to the problem by traversing the tree.
These heuristics will start out on one path and will terminate when either an optimal solution is
believed to have been found, or the path is known to result in an unsatisfactory solution. It is
difficult to determine which path to start on and, once on a particular path, determining whether
the path is worth traversing, i.e., lower costs possible, or whether to proceed on a different path.
This work has been done by Christofides and Whitlock [12], Hinxman [13] and Beasley [9,10].
Another heuristic used is the iterative improvement method. Application of the method of
iterative improvement requires the definition of a solution space, a cost function and a set of
moves that can be used to modify a solution. Define a solution xi = (zic, xii,. . . , Xim) on m
variables. A solution y is a neighbor (neighboring solution) of a solution x if y can be obtained
from x via one of the moves, In this method, one starts with an initial solution represented
Composite stock cutting 59
by xo = (~oo,~oI,...,
20~). At iteration i, if the current solution is xi, then its neighbors are
examined until a neighboring solution xi+1 is found with a new cost. In that case, x,+1 is the
new solution and the process is continued to examine the neighbors of the new solution. The
algorithm terminates when it arrives at a solution which has no neighboring solution with a
lower cost. This process tends to minimize the cost but can get trapped in a poor solution, i.e.,
it may be a local minimum, but not global. Figure 1 shows how this may happen. If xe is the
chosen initial configuration, then the iterative improvement method will choose configuration A
as the optimum. If xb is the chosen initial configuration, then the iterative improvement method
will choose configuration B as the optimum. Solution A is a local optimum. All neighboring
solutions have a higher cost than A, hence iterative improvement is trapped at A. However, B is
the configuration that minimizes the cost. It is the global solution. Thus iterative improvement
is sensitive to the choice of the initial configuration.
Cost R
I
Configuration S
Figure 1. Local minimum problem. The initial guess is important in iterative im-
provement. A poor initial guess xo leads to a local minimum and a sub-optimal
solution, A. A good initial guess, x0 leads to the optimal solution, B.
Because most of the above approaches either do not give an optimal or near optimal solution,
or are not applicable to a wide variety of applications, a new approach using simulated annealing
is considered. Simulated annealing can be used to give near optimal solutions and be used for all
applications.
3. SIMULATED ANNEALING
3.1. Motivation
Simulated annealing is a method of optimization that tries to avoid the pitfalls inherent in
other optimization methods, such as the iterative improvement approach; i.e., it seeks the global
or near global minimum of a function without getting trapped in a local minimum. Simulated
annealing is designed to optimize functions of several hundred variables or more, and is especially
attractive when the functions are not smooth, i.e., have many local minima. Simulated annealing
has been used in solving circuit routing, image processing and the traveling salesman problem
[14-191. Possible use of simulated annealing for solving the stock cutting problem is proposed by
Dagli [20] as a promising tool based on a small problem.
3.2. Statistical Mechanics
At the heart of the method of simulated annealing is an analogy in statistical mechanics.
Statistical mechanics is the central discipline of condensed matter physics, a body of methods for
analyzing aggregate properties of the large numbers of atoms, to be found in samples of liquid or
solid matter. Because the number of atoms is of order 1O23 per cubic centimeter, only the most
probable behavior of the system at a given temperature is observed in experiments. Since at a
60 H. LUTFI YYA ei al.
given temperature, the behavior may vary, determining the most probable behavior is done by
considering the average behavior of a collection of identical systems. This is called an ensemble
of systems. In this ensemble, each configuration, defined by the set of atomic positions, P = (ri),
of the system is weighted by its Boltzmann probability factor,
e-E(r,)/bT
I
where E(r) is the energy of the configuration, Icg is Boltzmanns constant and T is temperature.
A fundamental question in statistical mechanics concerns what happens to the system in the
limit of low temperature T. For example, whether the particles (could be referring to atoms or
molecules) remain fluid or solidify, and if they solidify, whether they form a crystalline solid or a
glass. Ground states, i.e., states with lowest energy, and configurations close to them in energy
are extremely rare among all configurations, yet they dominate at low temperatures because as T
is lowered the Boltzmann distribution collapses into the lowest energy state or states.
Consider a two-dimensional network of points arranged in a square lattice with each point,
labeled by its integer coordinates i,j, connected to its four nearest neighbors, where two points
ii, jr and iz, j, are nearest neighbors if
Suppose we place in each point a particle with a magnetic moment (spin), and that each particle
can be in one of two states, conventionally labeled
T = -1 and P = +l,
or called spin down or spin up, respectively. Suppose that each particle interacts with its
nearest neighbors. Then the energy of the system can be written as
where Ciljli3ja is a connection matrix such that
Ciljli2j2 =
{
1, iflil-izI<andJjl-jzlI1,
0, otherwise.
This is a very simplified but physically reasonable model, which was introduced by Ising [21], for
a two-dimensional substance exhibiting ferromagnetic behavior. J (a positive constant) is the
contribution to the total energy given by a pair of adjacent spins, the sign of the contribution
being negative if the two adjacent spins are aligned (i.e., both up or both down) and positive
otherwise. The system, in order to minimize its energy, tends to align all its spins in the same
direction. Therefore, the ground state configuration (state with minimal energy) is very simple,
with all the spins pointing in one direction. In practice, the magnetization is rarely perfect, i.e.,
the spins are slightly shifted towards the interior of the (-1, 1) segment i.e.,
ri = 1 - El or pi = -1 + Ez,
where 0 < ~1.2 << 1. This is corrected by replacing
ri -t sign Ti.
The system is crystallized when E is minimal. The Boltzmanns distribution implies that E is
minimal at a low temperature.
In practical contexts, low temperature is not a sufficient condition for finding ground states of
systems. Experiments that determine the low temperature state of a system that gives us minimal
energy are done by a method called annealing. First, the system is raised to a high temperature,
Composite stock cutting 61
then the temperature is lowered slowly and the system spends a long time at temperatures in the
vicinity of the freezing point. Enough time must be spent at each temperature for each of the
spins to reach a steady state. In other words, each spin must be given enough time to polarize
towards 1 or -1. This is called thermal equilibrium or quasi equilibrium. If the temperature
is lowered too quickly, the system does not have sufficient time to achieve equilibrium and the
resulting state might have high energy.
As mentioned before, studies have shown that the distribution of atoms in a system, at a
temperature T, satisfies the Boltzmanns distribution. This probability expresses the idea that
a system in thermal equilibrium at temperature T, has its energy probabilistically distributed
among all different energy states. Even at low temperatures, there is a chance, albeit very small,
of a system being in a high energy state. Therefore, there is a corresponding chance of a system
to get out of a local energy minimum in favor of finding a more global minimum.
The spin variables are discrete and updated asynchronously, i.e., spin variables do not polarize
all at once. The annealing process of the system described by Ising can then be represented
formally as follows:
ri(t + E) = sign (Ti(t) Ai( (1 - eBA*q) - ri(t) eeASq,
where pi(t), ri(t + E) represent the value of the spin variable ri at time t and t + E, respectively,
and Ai = ri(t + E) - q(t).
For large temperatures, the second term of the above equation dominates. However, as the
temperature decreases, the first term will dominate.
3.3. Metropolis Algorithm
It is sometimes difficult to work with ensembles of physical systems. Metropolis [22] introduced
a simple algorithm to simulate the annealing process. For a given temperature T, the Metropolis
method is a way to sample states of the physical system with the Boltzmann distribution. A
state xc is randomly chosen. At each iteration, the algorithm then chooses (again randomly)
a small perturbation Azi in the system and calculates the energy change AE caused by the
perturbation
AE = E(xi + AZ~) - E(x~).
I f AE < 0, then the perturbation is accepted, for it means that is energetically favorable for
the system; otherwise, it is accepted with probability
e-AEI kBT
When the perturbation is accepted, one continues the process with the perturbed state Xi +
Axi replacing the old one; otherwise, a new perturbation Azi is attempted. This process is
continued until thermal equilibrium is reached, i.e., until the probability distribution of the
configurations approaches the Boltzmann distribution. It can be shown that the sequence of
states obtained in this way is distributed according to the Boltzmanns distribution. Note that
the xis are the encoding of the spin variables. The Metropolis method is widely used in physics
to study numerically the thermodynamical properties of large systems than cannot be treated
with analytical methods. The procedure is shown in Figure 2.
The Metropolis method can be used to simulate the annealing process of the substance given in
the previous section, the Ising model, as follows. Let ro represent the values of all spin variables.
At each iteration, a spin variable pi is chosen to have its value changed to -pi. This is the small
perturbation of the system. The new energy is calculated according to Isings equation for energy.
AE is computed. If AE < 0, then this new state is accepted. The Boltzmann distribution implies
that it is possible for a system to be at a higher energy state. This is simulated by the Metropolis
procedure by computing a probability of accepting the new state. Thermal equilibrium is reached
as the new states generated approach the Boltzmann distribution.
62 H. LUTFIYYA et al.
Metropolis Procedure
Begin
Choose some random initial configuration S;
Repeat
5 := Some random neighboring configuration;
A := E(S) - E(S);
Prob := min(I, esAlkBT);
if random (0,l) 5 Prob then S := S' ;
until false;
End
Figure 2. The Metropolis Procedure probabilistically samples states
uration.
Physical systems Optimization problems
State Configuration
Energy Cost function
Ground state Optimal solution
Quick cooling Iterative improvement
Careful annealing Simulated annealing
of the config-
Figure 3. Analogy between physical systems and optimization problems.
3.4. Simulated Annealing Technique
Simulated annealing was proposed independently by Kirkpatrick [23] and by Cerny [24], as a
method for minimizing functions of many variables. The idea was derived from the algorithm
by Metropolis [22]. There is an analogy between a combinatorial optimization problem and the
problem of determining the lowest-energy ground state of a physical system. This analogy is
summarized in Figure 3.
The states of a system correspond exactly to the configuration of the combinatorial optimization
problem. The ground state corresponds to the optimal configuration, i.e., a configuration that
minimize the cost function. The finding of the optimal configuration corresponds to determining
the ground state. This process is called simulated annealing. The temperature becomes a control
parameter.
With simulated annealing, a high value of T is started with, so that the probability of the
system being in a given state is independent of the cost of that state. T is slowly reduced, by
making sure that at each new value of T enough steps of the Metropolis procedure are made to
guarantee that thermal equilibrium has been reached. The procedure is continued until T = 0. I f
the cooling has been slow enough, the final state reached is the ground state of the combinatorial
optimization problem being considered; i.e., the values of xi so obtained realize the absolute
minimum of the cost function. In practice, in many cases one is not really interested in finding
the absolute minimum. Rather, in many interesting situations the minimum configuration is
highly degenerate. In other words, there are many minima with values very close to the absolute
minimum value. A general simulated annealing algorithm is given in Figure 4.
Initially, the control parameter, T, is given a high value and the system is assumed to start
at some state S. To simulate reaching thermal equilibrium at a temperature T, the generic
simulated annealing algorithm generates a sequence of configurations of the cost function. A
generation mechanism is defined, so that given a configuration S, another configuration S can
be obtained randomly from the neighborhood of S. A is defined as the difference in the costs
of each configuration, i.e., A = C(S) - C(S). The probability for a state S to be the next
configuration is determined as follows:
Prob (new state = S) =
I
I,
ifAs0,
e-A/T
, ifA>O.
Composite stock cutting
Generic Simulated Annealing Algorithm
begin
S := Initial solution So;
T := Initial temperature To
while (stopping criterion is not satisfied) do
begin
while (not yet in equilibrium) do
begin
S := Some random neighbor of S;
A := C(S) - C(S);
Prob := min( 1, e-*iT);
if random (0,l) 5 Prob then S := S;
end ;
Update T;
end ;
Output best solution;
end ;
Figure 4. The simulated annealing algorithm makes probabilistic decisions based on
a control parameter T, c&led temperature.
63
The acceptance criterion is implemented by drawing random numbers from a uniform distribution
on [ 0, 1). Thus, there is non-zero probability of continuing with a configuration with higher cost
than the current configuration. This is continued until equilibrium is reached. The control
parameter, T, is then lowered in steps, with the system being allowed to approach equilibrium
for each step. The algorithm is terminated for some small value of T. The final configuration is
taken as the solution of the problem at hand.
Determination of the initial temperature T, the decrement factor of T at each step (i.e., how T is
updated) and the determining of thermal equilibrium is called the cooling schedule. Convergence
with various cooling schedules has been proven [25]. Convergence at each temperature, or control
parameter, means that the inner loop has reached thermal equilibrium. The convergence of the
entire algorithm means that an optimum configuration, or cost value, has been reached. An
exponentially long time may be required for convergence, depending on the size of the problem.
Various heuristics, in addition to the physical analogy, must therefore be used in determining the
cooling schedule.
4. SIMULATED ANNEALING PARAMETERS
The crucial parts of the algorithm are the definition of move, or the way in which the
configuration is perturbed and the cooling schedule.
4.1. Moves
An important parameter in simulated annealing is the definition of move, or the way a
configuration is perturbed to get a new configuration. Most researchers agree that at high tem-
peratures, large moves should be made, but as the temperature is lowered, smaller moves should
be made. A large move is defined as a move that causes a large difference in the value of the cost
function. A small move is defined as a move that causes a small difference. Large moves at high
temperatures allows the algorithm to avoid local minima.
4.2. Cooling Schedule
A summary of some effective cooling schedules is presented in [25]. The initial value of the
control parameter T should be large enough to allow the algorithm to accept any new config-
urations regardless of the change in cost. Let the notation Ti refer to the temperature at the
jth iteration of the outer loop. The updating of the temperature usually takes the form of the
decrement rule T~+I = f(Tk). The decrement rule controls the outer loop and should cool the
system slowly enough to provide an accurate solution without running an excessively long time.
64 H. LUTFIYYA et al.
The value of the final temperature determines when the system should halt processing. This is
usually when there is little improvement in the optimization. The quantity /iMe must be specified
by the user and must be determined empirically such that solutions are accurate and terminate
in a reasonable amount of time.
The time required for the inner loop to reach quasi-equilibrium is another parameter of the
cooling schedule. There is a trade off between time and accuracy. The theory [25] says that for
convergence, every state must be visited infinitely often. Obviously, in practical uses, this is not
feasible. The length of the inner loop should be long enough to provide a good solution, but not
so long as to provide little extra information for the time spent. Therefore, some heuristic must
be used to determine inner loop iterations.
One common cooling schedule is described by van Laarhoven and Aarts [25]. Let the nota-
tion ACij refer to the difference in the cost of the system when the system goes from configuration
i to configuration j.
DEFINITION 4.1 (ACCEPTANCE RATIO). The acceptance ratio x is the ratio between the number
of configurations accepted by the algorithm and the totai number of configurations generated by
the algorithm for a given temperature.
As pointed out before, the value of Tc should be high enough such that virtually all generated
configurations are accepted. This corresponds to x close to one. The value of TO can be obtained
by monitoring the evolution of the system during a number of runs of the algorithm before the
actual optimization process starts, and adjusting the cooling parameter in such a way that a
constant value of the acceptance ratio is maintained. The value of TO is then given by the final
value T obtained by updating To m times according to the expression:
T = A??-
>
-1
In
m2
m2x-(l-x)9
where ml and rn2 are the number of generated configurations with AC, 5 0 and ACij > 0
(ml + rn2 = mo), respectively, and Ac(- the average value of those ACij for which ACij > 0.
Determination of the start value of the cooling parameter in the way described above can only
be done reliably in those cases where the values of the cost function for different configurations
are sufficiently uniformly distributed. Otherwise, the above will result in a value of TO which is
too small and, therefore, will cause the algorithm to get stuck in a local optimum. In this case
one is forced to determine TO differently.
The notation Ci(Tk) refers to the cost of the configuration when the temperature or control
parameter is Tk. The term ??(Tk) approximates the statistical expectation of cost at tempera-
ture Tk, and is the average cost over n accepted moves, achieved after the inner loop has reached
equilibrium, i.e.,
c(Tk) = ; f: cj(Tk).
a=1
Similarly, c2(Tk) approximates the second moment of cost
c2(Tk) = ; $ cf(Tk).
r=l
The sample variance of cost is defined as
u2(Tk) = c2(Tk) - [c(Tk)12.
The decrement rule f is established by the following equation:
Tktl k
=T l+ln(l+b)Tk -
I
3a(Tk) *
Composite stock cutting 65
The smaller u, the slower the system cools. Also note the dependence on a(Tb); the larger the
variance of costs for the inner loop, the slower the system cools. The final value of the control
parameter, ZJ, is taken to be the first value of the control parameter, that satisfies
u(Tk)
for some small value of E. The smaller E, the longer the algorithm runs.
Theoretically, the thermal equilibrium is reached after each configuration is visited infinitely
often. However, for practical purposes an approximation of when the algorithm has reached ther-
mal equilibrium must be used. However, there is no practical, theoretical method of determining
thermal equilibrium. In this paper, the number of iterations of the inner loop is dependent on
the number of moves that were accepted. Hence, as an annealing run approaches a solution, the
longer the inner loop will run.
5. DISCUSSION OF THE COST FUNCTION
The composite, or oriented 2D bin packing, stock cutting problem is the problem of cutting
patterns from a stock sheet such that the scrap area is minimized, and stock orientation is
important. It is important to note that most applications do not allow the patterns to be
rotated any random angle. In other words, patterns may rotate for a limited number of rotation
angles. Various cost functions based on different parameters can be used. Research has led us to
believe that a cost function based on a weighted sum of distances between patterns within the
configuration represents a computationally feasible function that allows us to minimize the scrap
area.
DEFINITION 5.1 (SHAPE SET). Let the run of shapes to be cut be s = (si), for i = 1,. . . ,S. s
is the shape set.
Let hsi = (h,,j,, hsijY) where j = 1,. . . , Hi. (hd,jl, h,,j,) is the one end of a line segment
and (h,,(j+l),,h ,
d (J+l)y) is the other end of a line segment of shape si. We will denote this line
segment by e,,j,j+l.
DEFINITION 5.3 (ROTATIONAL ANGLE). 6k is the ICth rotational angle of an object.
DEFINITION 5.4 (OBJ ECT). For each si and ok, let
[h rrkj~h~akjyl = [h~sj+h~ijyl
COS ek sin en:
_ sin ok
1
CO&k
where rik is the kth rotation of an object si.
An object is h,, for some shape si of the shape set s, or is the kth rotation of shape si.
5.1. Finding Minimal Area
The cost function used is a weighted sum of distances between patterns within the configuration.
An index, which we call the affinity relation and is denoted by aij, represents the tendency of
object i to attract object j. The affinity relation between pattern i and j is the weight associated
with the distance between objects i and j.
DEFINITION 5.5 (EDGEWISE ADJ ACENT). Define 6ijklmn between each pair of objects i,l, for
each jth and mth rotational angles of objects i and 1, respectively, and for each pair of line
St?grIWI~S G,jk,k+l and er,,n,ntl, such that
Sijklmn =
1, if
hrnjkV-h,*j(k+l)v hrr,nv-krj,(n+l)v < E
,ijkr-.,j(k+l)r - hlmr-lrn(+l)= -
1

0, otherwise.
6 ijklmn iS one if the slopes of the line segments er,jk,k+l and er,,,,,,,+l differ no more than E,
where & is a very small number. 6ijklmn having a value of one implies that the jth rotation of
MCM 16:1-F
66 H. LUTFIYYA et al.
pattern i and the mth rotation of the pattern 1 may be placed adjacent to each other by placing
the line segments e,,,k,k+l and e rr,,,nn,n+l adjacent to each other. The corresponding objects i
and j are sai d to be edgewise adjacent.
EXAMPLE 5.1. In Figure 5, assume that both objects are at their Oth rotational angle and no
other rotational angle is permissible,
hi = (0,8),(0,0),(4,0),
hz = (13,13), (12, lo), (10,14).
The slopes of edges are computed as follows:
8-O
O-O=0
q12 : O_O + 00, q23 : 0_4 ,
8-O 2
e1,31 : 0_4 = - ,
13 - 10
: - = 3, e2,2s
10 - 14
e2,12 13 _ 12
: - = -2,
14- 13 1
12 - 10
e2,31 : - = --.
10-13 3
Edges ei,si and e2,2s have equivalent slopes. Hence, the two objects may be placed adjacent to
each other by placing edges ei,si and e2,2s adjacent to each other. Figure 6 shows this.
hllO,h118
h2320,h2314
13
h120,h120 cl,23 h134,h130
Figure 5. Shape representation. Figure 6. Maximize edgewise
adjacency.
DEFINITION 5.6 (LENGTH OF EDGEWISE ADJACENCY).
aijklmn = mh ([(r,jkr - r,j(k+l)t)2 + (hijky - hrsj(k+l)y)2] i.iklmnj
[(h
r1lnnz - b,,,,(n+~)r)~ + (hl,ny - hrl,,,(n+~)y)~l hklmn) .
DEFINITION 5.7. The affinity relation for object i and 1 i s the following:
ail = max(aijkhn).
THEOREM 5.1. For each pair of objects i, j, the affinity relation represents the positioning of the
object that maximizes the length of edgewi se adj acency.
PROOF. aijkl,,,,, is the length of the edgewise adjacency between the line segments erijk,k+l
and e
ri,n,n+1*
Note that if 6ijklmn is zero, then eijklrnn is also zero. This indicates that there is
no edgewise adjacency between the line segments er,jk,k+l and err,,,nn,n+i. If, on the other hand,
5 ijklmn iS equal t0 one, then Uijklmn is equal to the length of the shorter segment.
I
Composite stock cutting
5.2. Cost Function
The cost function is the following:
67
- (Y
(
c
ij
$aij+PC $
ij
ij 1
1
9
where (Y, p are positive real numbers that indicate the contribution of each of the components
in the cost function. Choosing CY too small will result in configurations positioned close to the
origin, but far from an optimal placement. The cost function for the simulated algorithm consists
of two independent terms.
Table 1.
Table 2.
affinily relation
F+l
LIH IL I
distance
J-x--fq
aflinily relation
L H
L L H
distance
H
LL L
5.2.1. First Term of Cost Function
Consider the first term of the cost function is the following:
c
dij aij ,
ij
This term is minimized as the higher the affinity relation between two objects the smaller the
distance between the two objects. Since the affinity relation represents the positioning of the
corresponding objects such that the box surrounding the two shapes is small, then the higher
the affinity relation, the greater possibility of saving scrap area. Hence, minimizing the distance
between two objects that have a high affinity relation should bring the objects closer the position
that maximizes their edgewise adjacency. Table 1 illustrates the status of the cost relative to the
status of the affinity relation and distance values.
There is one problem associated with this function. If the adjacency between two patterns is
zero, then the cost associated with those two patterns is 0, irrespective of the distance between
them. Our solution to this problem was to maximize this term:
c
1
ij
dijj aij3
or equivalently, minimize this term
-T $aij.
G
Table 2 illustrates the status of the cost relative to the status of the affinity relation and distance
values. Note that the cost associated with two patterns is maximized when their adjacency is
high and distance between the patterns is low. If the adjacency is 0, it doesnt affect the result.
It is important to note that the cost function used in this paper does not always find the
optimal solution. The affinity relation represents the positioning of the corresponding objects,
such that the box surrounding the two shapes is small. However, this box is not minimal.
68 H. LUTFIYYA et al.
THEOREM 5.2. The affinity relation does not guarantee that the positioning of the corresponding
objects is such that the bounding box is minimal.
PROOF. See Figures 8, 9 and 10. It is easy to compute the affinity relation for shapes 1 and 2.
Edgewise adjacency is maximized in both Figures 9 and 10, however the bounding box in Figure 9
is larger than the bounding box in Figure 10.
I
xl
hl43J l41 h&h131 k?47&41 h2311&231
I]
Figure 8. Candidate figures are placed to maximize edgewise adjacency and minimize
the bounding box.
s2
I
c
sl
1
Figure 9. Maximize edgewise adjacency. Figure 10. Minimal bounding box.
5.2.2. Second Term of the Cost Function
The cost function will also try to minimize the following:
where diO represents the distance of pattern i from the origin. This forces the clustering of objects
into a small bounding box.
5.8. Overlap
One component of the the cost function should include a penalty function on the area of
overlapping patterns. It is easy to formulate this component if the objects are rectangular in
nature [26]. However, we would like to be able to work with more irregular patterns. Formulating
algebraic constraints that prevent the overlapping is very difficult.
Compoeite stock cutting 69
Our method of handling overlaps uses a three dimensional grid of indicators where the output
of indicator Vijk is as follows:
Kj k =
1, if object i occupies grid jk,
0, otherwise.
Vijk = 1 implies that object i occupies grid point jk.
DEFINITION 5.8 (OVERLAPPING). A configuration overlaps when there exists the following sit-
uation:
c
%jk > 1,
i
for a fixed j and fixed k.
6. MOVES
6.1. Types of Moves
Three types of moves are allowed in order to be able to change the arrangement of the given
object. These three types of moves, Cp, are the following:
Qi: Displace an object to a new location.
Cpz: Interchange two objects.
as: Change the orientation of an object.
6.1.1. Description of @I
An object i, where hai = (h,,jZ,h,,jv), i = 1,2,. . . , Hi, is the set of vertices, is randomly
selected for displacement to a randomly selected new location. The new location is determined
by randomly selecting a displacement in the horizontal direction, represented by z_disp, and by
randomly selecting a displacement in the vertical direction, represented by y_disp. The new set
of vertices can be computed as follows:
h:, = (h,,jc + z_disp, h,,jy + y_disp).
Large-distance moves usually imply large values of AC. At low temperatures, only moves
which approximately satisfy AC 5 0 have a reasonable chance of being accepted. Hence, at low
temperatures, the large distance moves are almost invariably rejected. In order to generate moves
which have a reasonable probability of acceptance, these large distance moves are prohibited by
the use of a range-limiter window [27]. When an object is selected for displacement, such as
shape 1 in Figure 5, the range-limiter window is centered at (te, ys), corresponding to the center
of object i. The randomly selected new location for object i must lie within the range-limiter
window. At the beginning of the annealing process, the window size is set to be large enough to
contain all the objects, and it shrinks slowly as the temperature decreases. In fact, the height
and width of the window are proportional to the logarithm of the temperature:
r_disp=x_dim{F},
y_disp=y_dim{T}.
x-dim represents the window span in the x-direction and y-dim represents the window span in
the y-direction at the initial temperature (To). The value of X was chosen such that for the initial
temperature, the term in the braces in the right hand side of the above equations is normalized
to one. That is,
70 H. LUTFIYYA et al.
6.1.2. Description of @z
@2 will cause an interchange of objects i and Ic, where
hi = (hs,js, ha,jy),
j = 1,2 )..., Hi,
h6 = (&,jz, hsbjy),
j= 1,2 ,...) Hk.
Let ci = (xi, yi) and CL = (xk, yk) denote the centers of objects i and h, respectively. Compute
the following:
xi_disp =
xi - xk, if xi > xk,
xk_disp =
Xk
-Xi, if Xk > Xi,
Xk -
Xi, otherwise, Xi - Xk, otherwise,
yi_disp =
Yi -Yk,ifYi > Yk,
yk_disp =
Yk-Yi,iifYk >Yi,
yk - pi, otherwise, Iji - yk, otherwise.
The displacements can be added to the chosen objects in a similar manner to the displacements
displayed in @I.
6.1.3. Description of QP,
An object i is randomly selected to have its orientation changed. A rotational angle 6k is
randomly chosen. hPlk is the set of vertices after object i has been rotated. It is computed as
follows:
P r,kjzl hr,kjyl = [h~lj~~ hsdYl
Co.5 t?k Sin ek
_ .& ek
1
COSek
where rik is the kth rotation of object Si. Hence,
h
s, = hr,,.
6.2. Selection of Moves
Moves are selected by randomly picking an integer number on the interval [1,3], where 1
corresponds to choosing ai, 2 corresponds to choosing $ and 3 corresponds to choosing ip3. We
experimented with several discrete probability distribution functions.
We first experimented with a discrete uniform distribution on the integers 1,2,3, where the
probability distribution function f(x) is the following:
f(x) = f,
where x = 1,2,3
In subsequent experiments the probability distribution function was the following:
f(z) = ::
i
ifx= 1,
if x = 2,
l-p-q, ifx=3,
where p + 1 < 1 and p, q > 0. The choice of p and q can have an important
solution.
Objects of moves, the displacements and rotational angles were chosen from
distribution.
effect on the final
a discrete uniform
Composite stock cutting 71
7. COOLING SCHEDULE
Using the method described in Section 4 for determining the initial value of the control param-
eter To, resulted in a value of Ts that was too small. This result made it difficult to determine the
decrement value and the stopping criteria described in Section 4. Hence, the following heuristic
was used in determining values for the simulated annealing parameters.
A fixed number of runs of the algorithm, denoted by N, was assumed. The decrement function
was assumed to be of the form:
Tk+l = a Tk-1,
where/c = 1,2,3 ,...,
where is 0.85 < (Y < 0.95.
This is justified, because the decrement rule described in Section 4 would allow only small
changes in the value of the control parameter. Several researchers have used this decrement rule
and it has been shown to be successful [16,28].
We also chose a final value of the control parameter, TJ , to be a small enough number such
that the acceptance ratio was very small. The initial value of To was computed as follows:
The length of the inner loop, L, was based on the intuitive argument that for each value of
Tk, a minimum amount of moves should be accepted. However, as Tk approaches 0, moves are
accepted with decreasing probability, and thus L becomes very long. Hence, the length of the
inner loop is ceiled by some constant. Other researchers [16,28] have used this rule.
8. EXPERIMENTAL RESULTS
We implemented the procedure described in this paper on placement of regular and irregular
polygons described through line-segment geometry. Discretization of this geometry was accom-
plished through application of a scan algorithm to fill each polygon with Kjt. The pairwise
affinity relation eij was computed offline from the main annealing procedure for fixed rotations
of 0, 90, 180 and 270 degrees.
Calibration of Procedure
To calibrate the annealing procedure, regular convex polygons were placed with the result
shown in Figure 11. While, by Theorem 5.2, the method is not guaranteed to minimize the area
of the bounding box, these results are clearly good.
Figure 12 depicts the performance of the annealing algorithm as the temperature control pa-
rameter in the annealing procedure decreases. Notice that the cost function does not decrease
monotonically as would be the case with straight iterative improvement.
Figures 13 and 14 depict a typical irregular shape packing. Notice that placing the regular
figures took only a small initial temperature while placing complex figures takes a large initial
temperature and a small decrement. These results, however, are coarse approximations and
better annealing schedules are expected to exist which will decrease the run time significantly.
9. SUMMARY
There is not yet enough practical experience with the method of simulated annealing to say
definitively that it will realize its current promise. The method has several extremely attractive
features, rather unique when compared with other optimization techniques. First, it is not
greedy, in the sense that it is not easily fooled by the quick payoff achieved by falling into
unfavorable local minima. Second, changes which cause the greatest cost difference are sifted
over when the control parameter T is large. These decisions become more permanent as T is
lowered, and attention then shifts more to smaller refinements in the solution.
72 H. LUTFI YYA et al.
Figure 11. Placement of five regular polygons. To = 1500. Decrement= 0.95. Length
of Markov chain = 40. Number of chains = 1000.
4e+OE
-I
3c+O8
cost
2e+O8
lc+O8
I
I I I I I I I I I
0.1 1 10 100 1000 loo00 I O0000 le+06 le+07 le+08
Temp
Figure 12. Results of annealing schedule for complicated figures.
10. FUTURE RESEARCH
Simulated Annealing is clearly a time-consuming procedure. Two possible methods exist for
speeding up the process. The first is based on the granularity of the discretization and the second
appeals to parallel processing.
The simulated annealing algorithm is fast for coarse granularities since the moves, 0, are
less computationally expensive than for finer granularities. However, the finer granularities are
necessary to prevent overlapping and to achieve a usable result. Annealing at several different
granularities can utilize a coarse approximation at the high T, and utilize a fine granularity at
lower values of 7. This structure yields a multigrid approach to the problem such as exploited
in [29]. The problem structure is formulated as a pyramid as shown in Figure 15.
The second alternative is to appeal to parallel processing. The main difficulty is the mainte-
nance of a global state S of the annealing process. This is further complicated by the desirability
of using distributed memory multiprocessors or multicomputers, [30] which have no single global
picture of the state. We are currently working on multigrid methods for this problem. In an
73
Figure 13. Complicated oriented bin
packing results under variable width
constraints. To = 106. Too = 0.38.
Decrement= 0.99. Length of Markov
chain= 2000. Packing density = 81.7%.
l=O
I=1
Figure 14. Complicated oriented bm pack-
ing results under fixed width constraints.
To = lo*. Too = 0.25. Decrement = 0.99.
Length of Markov chain = 2050. Packing
density = 81.4%.
Composite stock cutting
Figure 15. The pyramid architecture for a multigrid solution. At each level, the
granularity increases by 4.
asynchronous multigrid, both the number of processors and the granularity change simultane-
ously during the problem solution. At the start of the annealing process, at high T, large moves,
and most moves are accepted. To minimize the amount of state information that must be com-
municated, the number of processors involved is kept low. As the temperature decreases, more
rejected moves are made and accepted moves are smaller. Thus, less state information needs
to be communicated between processors, and more processors may be employed. Thus, in Fig-
ure 15, execution and processor utilization moves down through the pyramid hierarchy as the
temperature decreases.
1.
2.
3.
4.
5.
6.
REFERENCES
MR. Garey and D.S. J ohnson, Computers and I ntractability: A Guide to the Theory of NP-Completeness,
W.H. Freeman and Company, San Francisco, (1979).
C.H. Dagli, Knowledge-bases systems for the cutting stock problems, European Journal of Operational
Research 44, 160-166 (1990).
P.C. Gihnore and R.E. Gomory, A linear progr amming approach to the cutting-stock problem, Operations
Reseaxh 9, 849-859 (1961).
P.C. Gilmore and R.E. Gomory, A linear programmi ng approach to the cutting-stock problem, Operations
Research 11, 863-888 (1963).
P.C. Gihuore and R.E. Gomory, Multistage cutting-stock problems of two and more dimensions, Operations
Research 13, 94-120 (1965).
A.M. Geoffrion and R.E. Marsten, Integer progr amming algorithms: A framework and state-of-the-art
survey, Management Science 18 (lS), 465-491 (1972).
74 H. LUTFIYYA et al.
7. R.W. Haessler, A note on computational modifications to the Gihnore-Gomory cutting stock aIgorithm,
Operations Research 28 (4), 1001-1005 (1980).
8. H. Dyckhoff, A new linear programmi ng approach to the cutting-stock problem (1981).
9. J .E. Beasley, Algorithms for unconstrained tw-dimensional guillotine cutting, J . Opl. Res. Sot. 36 (4),
297-306 (1985).
10. J .E. Beasley, An exact twedimensional non-guillotine cutting tree search procedure, Operations Research
33 (l), 49-64 (1985).
11. B.R. Sarker, An optimum solution for one-dimensional slitting problems: A dynamic programnu ng approach,
J . Opl. Res. Sot. 39 (S), 749-755 (1988).
12. N. Christofides and C. Whitlock, An algorithm for two-dimensional cutting problems, Operations Research
25, 3044 (1977).
13. A.I. Hinxman, The trim-loss and assortment problems: A survey, Ezlropean Journal of Operational Research
5, 8-18 (1980).
14. C. Sechen and Sangiovanni-VinceneIIi, The Timberwolfplacement and routing package, I EEE J . Solid State
Circuits SC-20, 510-522 (1985).
15. C. Sechen and Sangiovzmni-Vincenelli, TimberWoIf3.2: A new standard cell pliicement and global routing
package, Proceedings 23rd Des. Automation Conf., Las Vegas, pp. 432-439, (J une, 1986).
16. C. Secben and Sangiovanni-Vincenel, Placement and global routing of integrated circuits using the simulated
annealing algorithm, Ph.D. Dissertation, University of California at Berkeley, (1986).
17. E. Bonomi and J . Lutton, The N-city travelling salesman problem: Statistical mechanics and the Metropolis
Algorithm, SI AM Rev. 26, 551-568 (1984).
18. S. Kirkpatrick, C.D. Gelatt, J r. and M.P. Vecchi, Optimization by simulated annealing, IBM Research
Report RC 9355, (1982).
19. P. CarnevaIIi, L. Coletti and S. PaterneIIo, Image processing by simulated annealing, IBM J . Res. Develop.
29, 569-579 (1985).
20. C.H. DagIi, Neural networks in manufacturing: Possible impacts on cutting stock problems, Proceedings of
Rensselaers Second International Conference on Computer I ntegrated Manufacturing, pp. 531-537, IEEE
Computer Society Press, Los Alamos, CA, (1990).
21. Ising, A contribution to the theory of ferromagnetism, 2. Phys 31, 253 (1925).
22. N. Metropolis, A. Rosenbluth, M. Rosenbluth, A. Teller and E. Teller, Equation of state calculations by fast
computing machines, J . of Chem. Physics 21, 1087-1092 (1953).
23. S. Kirkpatrick, C.D. Gelatt, J r. and M.P. Vecchi, Optimization by simulated annealing, Science 220,671-680
(1983).
24. V. Cerny, Thermodynamical approach to the traveIling salesman problem: An efficient simulation algorithm,
J . Opt. Theory Appl. 45, 41-51 (1985).
25. L. Aarts and P. van Laarhoven, Statistical cooling: A general approach to combinatorial optimization
problems, Philips J . Res. 40, 193-226 (1985).
26. D.D. Caviglia, G.M. Bisio and F. Curate& Neural algorithms for cell placement in VLSI design, Presented
at the International Conference on Neural Networks (1989).
27. C. Sechen, Chip-p1 arming, placement, and global routing of macro/custom cell integrated circuits using
simulated annealing, 25th ACM/I EEE Design Automation Conference, (1988).
28. D. J ohnson, C. Aragon, L. McGeoch and C. Schevon, Optimization by simulated annealing: An experimental
evalzlation, Parts I and I I , AT&T Bell Laboratories, (1987).
29. S. Franks, R. Khanna and B. McMiIIin, PAFMV-Pairwise asynchronous multigrid, Proceedings 1990 In-
ternational Conference on Parallel Processing, August, 1990, pp. 1-388-I-392; Operations Research (29)
1981, 1092-1104.
30. W. Athas and C. Seitz, Multicomputers: Message-passing concurrent computers, CompzLter, 9-25 (August,
1988).