Beruflich Dokumente
Kultur Dokumente
377, 1995
Copyright~i 1995ElsevierScienceLtd
Pergamon 0360-8352(94)00041-7 Printed in Great Britain.All rights reserved
0360-8352/95 $9.50+ 0.00
B U D D H A D E V R O Y C H O U D H U R Y I and J O H N F. M U T H "
~College of Business, Management Department, Mankato State University, Mankato, MN 56002, and
-'Operations Management, Graduate School of Business, Indiana University, Tenth and Fee Lane,
Bloomington, IN 47405, U.S.A.
INTRODUCTION
SOLUTION METHODS
Walas and Askin [2] develop an iterative heuristic based algorithm involving a traveling salesman
problem and a quadratic assignment problem. The traveling salesman heuristic (nearest neighbor
rule) generates sequence of operations with possible constraints. Constraints for tools are
considered only at the next stage, when a quadratic assignment heuristic is used for allocation of
tools to the sockets in the turret holding the tools.
Chauney et ai. [3] also formulate sequencing of operations as a traveling salesman problem. Their
solution method is based on space filling curve techniques. The type of problems considered here
involves more time for tool changing operations compared to that for repositioning of jobs between
operations. These authors present two heuristics--a direct approach and a clustering approach. In
the direct approach, a tour is searched for all locations on a parallelopiped as a three dimensional
traveling salesman problem. In the clustering approach, jobs requiring the same tool are identified
and formed into clusters. Two traveling salesman heuristics are then applied--one at higher level
367
368 Buddhadev Roychoudhuryand John F. Muth
to determine the sequence of these clusters, and another, at a lower level, to determine the sequence
of jobs within a cluster.
Meenakshi Sundaram [4] discusses the sequencing problem associated with metal cutting
operations. She uses a technique based on graph theory for determining the order in which surfaces
are to be machined. Complex products with a combination of cylindrical and prismatic surfaces
can also be handled by this technique.
Ssemakula and Rangachar [5] develop an artificial intelligence program to generate a sequence
of manufacturing processes. This system handles a variety of processes not limited to punching,
and attempts to reduce job and tool handling times, and machining time.
From the earlier discussion it is clear that the problem can be formulated as a traveling salesman
problem, and the same solution approaches can also be adopted here with the underlying
assumptions that:
(a) the tool holder is capable of holding all the tools required by all operations on the sheet;
(b) each operation on the sheet requires a single tool.
It may be noted that in this research repositioning of workholders is not considered. In
some of the jobs there may be several repositionings involved. This repositioning activity
involves additional movements on x- and y-axes directions besides the ones required for the
positioning of the jobs. Lack of consideration of the repositioning of workholders, therefore,
may not have any contradictory effect on the conclusions of this research. In view of the facts
that the operations considered here involve only punching, and that tool changing times are not
too high compared to job repositioning times, it is felt proper to include only some of the
approaches discussed in the various research reported earlier in the current research. Here the
focus is on the tool path determination problem, and the problem of tool assignment is not
considered.
The computational effort required to solve a traveling salesman problem increases exponentially
as the size of the problem increases. Optimization is impractical even for relatively small problems
involving numerically controlled machine tools or flexible automation. The literature on operations
research holds a number of heuristics for approximate solutions to the traveling salesman problem.
As such, they differ in both solution quality and computational effort. Tour construction
procedures, one of the two broad categories of these heuristics, generate an approximately optimal
tour from the distance or cost matrix. Tour improvement procedures start with a given initial tour
and attempt to improve upon this initial tour.
Included among the tour construction procedures are the nearest neighbor (NNBR) [6] and a
variation of it, the nearest neighbor within tool. In this category, there are two other heuristics--dis-
tance saved sequential (DSS) and distance saved multiple (DSM)--which use a ranked list of
"savings" resulting from joining any pair of nodes from high to low (Hill [7]). The basic idea behind
the "savings" is the reduction in tour length by connecting two nodes instead of covering each of
them individually from the originating node. Besides these methods in this subcategory, Golden
et al. [6] discuss four insertion heuristics--cheapest insertion, farthest insertion (FARIN), nearest
insertion (NEARIN), and random insertion (RANDIN)--based on an idea very similar to the
distance savings. They look at the cost of inserting any node into any subtour between any pair
of nodes. Therefore, it is not assumed that any of the nodes are initially connected. Bartholdi and
Platzman [8] discuss a heuristic based on the space filling curve technique (SFILL). This technique
can be applied to solve the traveling salesman problem. Chauney et al. [3] use a three dimensional
version of the same technique in their research. The important feature of this heuristic, besides
being simple, is the possibility of arriving at a reasonably good initial solution within a short period
of time.
Among the tour improvement procedures, the 2-OPT and 3-OPT heuristics are developed by Lin
[9], and K-OPT procedure by Lin and Kernighan [10]. According to Lin, a tour is K-optimal if
no improvement can be made in a tour by replacing any set of K links with any other set of K
links. Starting from any arbitrary node, the K-OPT heuristic chooses the best K edges that can
be swapped for reduction of tour length, instead of determining a priori the number of edges for
which a swap will be attempted. Unless otherwise stated, random initial tours are used for applying
the OPT procedures in this research.
Tool path optimization procedures for machine tools 369
0.50
0.50
L/2 _1 -4- 12.00
9.00 ~ [
= 5o
-f-- " ~ - ~ 0.32 (8)
4.03
4 24.00
A _
I 1.00
3.00
O
3.54 B
10.50
6.38~...~
,J
L_l L. . . . . . . . . . . . . . . .
TEST P R O B L E M S
The performance of all of the heuristics have been compared on 42 problems related to
numerically controlled punching operations at a major metal fabricator. A drawing of one such
part is shown in Fig. 1. The maximum and the minimum values of four problem characteristics
across all 42 problems are presented in Table I.
Preprocessing program
Preprocessing of the programs from the fabricating firm is required to set the problem up for
the traveling salesman heuristics.
The numerically controlled turret punch press used for this study is produced by the Wiedemann
Division of the Warner and Swasey Company (Model W-4560). Up to 36 punch and die sets may
be accommodated in the standard turret, with automatic changes from one punch to another taking
an average of 2 s [11]. This punch press uses a proprietary NC part programming language called
WIEDEPOINT V. An excerpt of the NC part program of a typical part is shown in Fig. 2.
Most of the instructions in the NC program follow the format:
X (x - coordinate) Y ( y - coordinate)T (turret position).
Because these determine the location of the operations involved in a job, the X- and Y-values are
important from the point of view of this research.
The input to the preprocessing program is a form of the output file from running of the
Wiedepoint V NC program modified in the form of spatial coordinates. This file needs the following
modifications for running the preprocessor program:
(a) The x- and y-axis locations along with the tool position is required for each line, although
in the above file repeated information from the previous line is not shown again.
(b) The correct position of the decimal point in the x- and y-axis values has to be explicit.
X72 14 T 2 4 M O O
X01705 Y0511
X05615
X09525
X13435
X17345
X21255
X25165
X29075
X32985
......
X01705 0225 T08
0797
X0316 0361
Y0661
(c) The letters X, Y, T, etc. must be removed from the respective fields.
(d) In a particular record, the positions of the fields require us to always follow the order:
X-value, Y-value and tool position.
(e) The x- and y-axis values should always be absolute with reference to the (0, 0) point, and
not any other point.
After incorporation of the modifications, the lines shown in the above excerpt should appear as
in Fig. 3.
72. 14. 24
01.705 5.11 24
05.615 5.11 24
09.525 5.11 24
13.435 5.11 24
17.345 5.11 24
21.255 5.11 24
25.165 5.11 24
29.075 5.11 24
32.985 5.11 24
....... . ..... .
1.705 2.25 8
1.705 7.97 8
3.16 3.61 8
3.16 6.61 8
1/4 116 0.52 114 113 0.52 113 0.528 114 250 0.240 171 0.350
1/2 106 0.57 1/2 104 0.57 104 0.572 1/2 213 0.281 153 0.391
3/4 98 0.60 3/4 191 0.313 141 0.423
1 94 0.64 1 94 0.63 94 0.636 1 176 0.340 133 0.450
1-1/2 86 0.70 1-1/2 87 0.68 87 0.684 1-1/2 155 0.384 121 0.494
first, the table movement between the operations along the x-axis; second, a similar movement
along the ),-axis; and third, the rotation of the tool holder to switch the tools. For any pair of
operations, one has to examine all of the above mentioned times to determine the effective time
between the operations. The largest of the three for any pair is the effective time for that pair of
operations.
The modified file thus created is then used to determine the effective time of travel between any
pair of points. The three fields of this file represent the absolute values of the locations of the
operations along the x- and y-axis, and the tool position on the tool holder respectively. Besides
creating this input file for the heuristics, the preprocessor also generates the information about the
job characteristics. The output file created by the preprocessor program is the input file to the
programs written for the heuristics. A typical output file from the preprocessor program, illustrated
in Fig. 5, has three fields: the first two for the two nodes, and the third one for the distance between
these two nodes. This distance is expressed in terms of the largest of the three travel times with
1/100s as the unit.
1 2 333
1 3 317
1 4 301
i 5 285
1 6 269
1 7 253
1 8 237
1 9 221
1 10 205
O O O O . . . . O . . . . O e o e O . . . . Q
O Q O O O e Q O O . . . . O O Q O O O O O e e
due consideration to the comparability of the coding styles. IBM 4381 computing system has been
used for running all the heuristics for solving the problems.
EXPERIMENTAL DESIGN
Considering the various aspects of this research as discussed above, and the availability of the
data, the experiment is designed to include the various independent and dependent variables as
covered in the following subsections.
Independent variables
The suitability of the various heuristics to solve the sequencing problem may be determined by
some characteristics of the jobs. Investigation of matches between some job characteristics and the
heuristics is a part of the current investigation. One of the important characteristics is the number
of different tools used in a job. Hill [7] uses "clustering" and "scale" to measure these characteristics
in his research. Clustering is the coefficient of variation of the data in the distance or cost matrix.
Scale refers to the "average distance in the distance matrix, excluding the main diagonal".
In line with the preceding discussion, the following characteristics of the jobs are considered in
this study as the first set of independent variables:
(a) Number of nodes in the job.
(b) Average number of operations per tool.
(c) Clustering of the operations on the job.
(d) Scale of the job.
Heuristics tested
Effectiveness of the following heuristics has been investigated in this study:
(a) Tour construction methods:
(i) Nearest neighbor
(ii) Nearest neighbor within tool
(iii) Random or arbitrary insertion
(iv) Nearest insertion
(v) Farthest insertion
(b) Tour improvement methods:
(i) 2-OPT (using random rule as initial solution generator)
(ii) 2-OPT (using nearest neighbor rule as initial solution generator)
(iii) 2-OPT (using random insertion rule as initial solution generator)
Golden et al. [6] and Hill [7] report comparative studies of some of the heuristics for solving
traveling salesman problems. In our earlier comparative study [12] of different heuristics for
traveling salesman problems, we found that the computational effort required by distance saved
sequential, distance saved multiple and k-OPT procedures were too high relative to that of the
Tool path optimization procedures for machine tools 373
nearest neighbor within tool heuristics, which is used by the numerically controlled punch machine
currently in question. These three heuristics are therefore dropped from consideration in this study.
In spite of limited need of computational effort, the random rule and the spacefilling curve
heuristics were observed in the above study to generate solutions of very poor quality. These two
heuristics are also therefore not considered in this study. It is assumed that these five rules would
not compare favorably with the one currently in use.
Considering the potential of effectively improving the solution quality of the tool path
determination problem, application of 2-OPT tour improvement rule with initial solution generated
by tour generation rules other than random rule has been included in this study. Nearest neighbor
(NNBR2) rule, and random insertion (RANDIN2) rule have been chosen for this purpose.
The eight heuristics, the performance of which are compared here, are included as dummy binary
(0, 1) variables, which are independent variables.
Dependent variables
The following criteria are used for comparing the performances of the heuristics:
(a) Tour length--Tour length covers the time for both metal removal or formation, and job
or tool handling.
(b) Computation time.
It is assumed that the comparison of the computation time in terms of CPU time is a good
representation of the actual situation because of the uniformity of the computer system used for
the different problems and the heuristics. The steps involved in the random insertion heuristic,
chosen as an example of the rules used here, is described in Fig. 6.
The two performance criteria mentioned above are the two dependent variables involved in this
research.
RANDOM INSERTION
SUBROUTINE NEXTNODE
1. Find the costs of insertion of the Candidate (k) node between all possible pmrs of
nodes in the partial tour.
2. Select the pair of nodes (x,y) in the partial tour for minimum cost of insertion.
3. Perform Subroutine EXECFILL to insert Candidate (k) between nodes x and y in
the partial tour.
4. Return.
SUBROUTINE EXECFILL
I. Readjust the partial tour array to permit insertion of Candidate (k) node in
the partial tour.
2. Return.
Statistical analysis
In view of the nature of the possible relationships among the various independent and dependent
variables, and the availability of data, multiple regression analysis is considered the most
appropriate statistical method for analysis in this experiment. In the case of regression analysis,
the data, as available from the actual problems, can directly be used in the models without any
need of assignment of levels.
The names of the variables used in the analysis of the problems and heuristics are given in
Table 2.
EXPERIMENTAL RESULTS
180 p
170 -- n DSM
160
150
140
130
120
I10 NEARIN FARIN
rn
100
O OPTK
p. 90
8O
x.
70
60
5o
40
30
20
OPT2
10 - DSS SFILL RAND
0
I I~l"l I I~ I I J i ~,--n
60 100 140 180 220 260
80 120 160 200 240
Tour length ( s )
F i g . 7. Tour length and CPU time (42 problems) for all rules.
rule may be appropriate for geographic data when the points are more or less randomly scattered
on the plane, but that is definitely not the characteristic of the parts that we are dealing with here.
Another leg of the L is for heuristics that give reasonably good results but are extremely time
consuming. If they gave better results than the other heuristics then the computing time would be
a secondary consideration, but they do not. That would rule out OPTK, FARIN, NEARIN, and
DSM. Figure 8 shows an expanded view of the cluster of points at the corner of the two legs of
the previous chart. It shows the NNBR, NNBRT, NNBR2, RANDIN, and RANDIN2 heuristics.
The generally used heuristic, namely the nearest neighbor within the tool, is dominated by the
random insert, and the other three here are better in the tour length but take a somewhat longer
1.8
1.7 O NNBR2
1.6
1.5
1.4
1.3
1.2
u,: I.I
1.0 o R A N D ! N2
0.9
0.8 o NNBR
e~
0.7
0.6
0.5
0.4
o NNBRT
0.3
0.2
0.1
0 RANDIN
0
i I I 1 ~ I
88 89 90 91 92 93
computing time. However, the computing time is completely irrelevant as we are dealing with times
of 0-2 s. It may also be pointed out that the tour length differences are not great, although they
may approach economic significance considering repetitive manufacturing.
Regression analysis
In Table 4, the results of statistical analyses of the regression model are presented. These analyses
have been performed using the SPSS program.
Table 4 deals with tour length (TRLENG) as the dependent variable in a regression model with
part features and heuristics as independent variables. Its R-squared value of 0.946 indicates that
the linear fit between the dependent and the independent variables is quite close. The probability
(P > F) value of 0.0000 is another indication that the regression model fits the data well. In this
regression model, the significance of the binary variables representing the heuristics, and of the
problem characteristics is studied with respect to the nearest neighbor within tool (NNBRT) rule.
For the individual variables representing the problem characteristics, the significant variables
describing the problems are the average distance (AVDIS), the number of nodes (NNDS), and the
number of nodes within a tool (NNDST). These all relate to the size of the problem or the scale
of the data, so it is hardly surprising that these are highly significant. The only variable that is not
a scale variable, the coefficient variation (CVAR), is not significant. The difference is not observed
to be significant for any variables representing the heuristics.
The interaction effects between three problem characteristics (NNDS, AVDIS, NNDST) and the
heuristics are also tested. Only number of nodes (NNDS) seems to show some significant difference
when nearest neighbor, or random insertion, or 2-OPT with random insertion rule is used.
CONCLUSIONS
This research compares the effectiveness of heuristics for sequencing operations in a numerically
controlled punch press machine. Proper sequencing of operations in punch press operations, as in
the case of other metal removal or metal forming operations, is important from the point of view
of controlling the time and therefore the cost associated with the movement of jobs or tools. In
punching operations, the contact time of a job with a tool is small compared to other operations,
and determination of the proper sequence is therefore even more crucial. The current research also
investigated the possibilities of matches between the solution techniques and the problem
characteristics. In an environment of decreasing product life cycle, exploration of such matches is
perceived to have potential benefits in terms of meaningful exploitation of the flexibility available
in modern machines. Tour length, as determined by the heuristics, and CPU time required by such
a run, are used as the bases of comparison. Four job characteristics are used in the exploration
of a match with solutions.
The results of the study bring into focus some interesting facts. Heuristics like random insertion
and the nearest neighbor perform well enough to be comparable to the more widely used nearest
neighbor within tool. Many other rules are quite ineffective. It is more interesting to observe that
some tour improvement rules tend to be more effective than the rule currently in use if the initial
solution generator rule is carefully chosen.
REFERENCES
1. M. E. Merchant. World trends and prospects in manufacturing technology. Int. J. Vehicle Design 6, 121 138 ~1985).
2. R. A. Walas and R. G. Askin. An algorithm for NC turret punch press tool location and hit sequencing, liE Tran.~.
6, 280-287 (1984).
3. F. Chauney, A. Hourie, R. Loulou and E. Wagneur. Sequencing punch operations in an FMS: a three-dimensional
spacefilling curve approach. INFOR 25, 26-45 (1987).
4. R. Mcenakshi Sundaram. Process planning and machine sequence. Proc. 8th Ann. ConL Computers md Engng I 1,
184-188 (1986).
5. M. E. Ssemakula and R. M. Rangachar. The prospects of process sequence optimization in CAPP systems. Computers
ind. Engng 16, 161--170 (1989).
6. B. Golden, L. Bodin, T. Doyle and W. Stewart Jr Approximate traveling salesman algorithms. Ops Res. 28, 694 71 I
(1980).
7. A. V. Hill. An experimental comparison of human schedulers and heuristic algorithms for the traveling salesman
problem. J. ops Mgmt 2, 215-223 (1982).
8. J. J. Bartholdi III and L. K. Platzman. Heuristics based on spacefilling curves for combinatorial problems in euclidean
space. Mgrnt Sci. 34, 291-305 (1988).
9. S. Lin. Computer solutions of the traveling salesman problem. Bell Syst. tech. J. 44, 2245 2269 (1965).
10. S. Lin and B. Kernighan. An effective heuristic algorithm for the travelling salesman problems. Ops Res. 21,498 516
(1973).
I 1. F. Robert Jacobs, Kieran Mathieson, John F. Muth and Terence M. Hancock. A rule-based system to generate NC
programs from CAD exchange files. Computers ind. Engng 20, 167-176 (1991).
12. Buddhadev Roychoudhury and John F. Muth. The solution of travelling salesman problems based on industrial data.
Accepted for publication. J. Opl Res. Soc.