You are on page 1of 9

488 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO.

3, JUNE 1999

Heuristics for a Tool Provisioning


Problem in a Flexible Manufacturing
System with an Automatic Tool Transporter
Hong-Bae Jun, Yeong-Dae Kim, Member, IEEE, and Hyo-Won Suh

Abstract—This paper considers a tool provisioning problem in systems, once a part is loaded on a machine, the part stays there
a flexible manufacturing system (FMS) with an automatic tool until all operations of the part to be processed in the system
transporter. The problem considered here is that of determining are completed. That is, each part visits only one machine for
the number of copies of each tool type for a limited budget with
the objective of minimizing makespan. This problem should be its entire processing, and tools which are not loaded on the
solved to decide the (additional) purchase of required tools when machine but required to produce the part are borrowed from
the FMS should be reconfigured due to change of part mix. To other machines or the tool storage area by an automatic tool
solve the problem, two heuristic algorithms are proposed. One transporter at the time of requirement.
is a composite search algorithm based on two greedy search In tool movement systems, the part scheduling problem
methods, and the other is a search algorithm in which numbers
of tool copies are determined based on tool groupings. In both is simpler than in part movement systems. However, tool
algorithms, simulation results are used to find search directions. flow control becomes more difficult since tools are moved
Computational experiments show that the latter search algorithm frequently. Frequent tool movements cause tool delay time,
gives better results. which is composed of tool transportation time and tool waiting
Index Terms—Flexible manufacturing systems, tool delay time, time. Tool transportation time is the time during which a
tool movement systems, tool provisioning problem. machine waits until a tool transporter brings a necessary tool
from another machine or the tool storage area. Tool waiting
time is the delay time during which a machine waits until
I. INTRODUCTION
required tools that are currently being used by other machines

I N this paper, we consider the tool provisioning problem in


a flexible manufacturing system (FMS), i.e., the problem
of determining the number of copies of each tool type with
become available. There will be positive tool waiting time
when tool conflict occurs, that is, when more machines require
the same tool at the same time than there are copies of this
the objective of minimizing makespan for a given set of part tool type in the system.
types. An FMS typically consists of numerically controlled In general, performance of a system is significantly affected
machines, a material handling system and an on-line computer by tool delay times since approximately 16% of scheduled
system to manage and control all operations. Each machine is production time is lost due to such delay [1]. Of the two
equipped with a tool magazine and an automatic tool changing types of tool delay times, tool waiting times affect the system
device that can switch tools in seconds. Some more advanced performance more significantly since tool transportation times
material handling systems include one or more automatic are often negligible compared to operation processing times
tool transporters as well as part transporters. Automatic tool in most systems. Therefore, it is necessary to minimize tool
transporters can transfer a tool from the tool magazine of one waiting time to improve the performance of a tool movement
machine to that of another, or between a tool magazine of system. In general, the length of tool waiting time depends
a machine and a tool storage area while the system is in on the tool availability, i.e., tool copy configuration, which is
operation. defined by the number of tool copies for each tool type. The
FMS’s can be classified into two groups according to the problem of determining the tool copy configuration is referred
material handling system: part movement systems and tool to as the tool provisioning problem, which will be considered
movement systems. In part movement systems, parts are moved in this paper.
between the machines according to the process plans of the To reduce tool waiting time, one can duplicate critical
parts. Once a tool is loaded onto a machine, the tool is not tools which cause tool conflicts. However, Gray et al. [2]
removed until the next tool loading time, as long as it is not and Tomek [3] report that tooling comprises 20–30% of the
broken or worn out. On the other hand, in tool movement cost of installing a new manufacturing system. As the budget
Manuscript received October 1, 1997; revised August 10, 1998. This paper for tool purchases is limited in general, it is necessary to
was recommended for publication by Associate Editor R. Uzsoy and Editor determine a tool copy configuration that can give the best
P. B. Luh upon evaluation of the reviewers’ comments. system performance for a given limited tool budget.
The authors are with the Department of Industrial Engineering, Korea
Advanced Institute of Science and Technology, Taejon 305-701, Korea. There have been a number of studies of the tool provisioning
Publisher Item Identifier S 1042-296X(99)03916-6. problem. Wang et al. [4] consider a tool provisioning problem
1042–296X/99$10.00  1999 IEEE
JUN et al.: HEURISTICS FOR A TOOL PROVISIONING PROBLEM IN A FLEXIBLE MANUFACTURING SYSTEM 489

magazine, it can borrow them from other machines and/or the


tool storage area. A machine can start processing a part only
after all tools required for the part are loaded on the machine.
If the tool magazine of a machine cannot accommodate all
required tools for parts assigned to the machine, only a subset
of the tools is loaded on the tool magazine and the rest are
stored at the tool storage area.
Each part requires a set of operations whose sequence is
known in advance, and each operation requires a set of tools.
We assume that each tool type has a tool life that can be
Fig. 1. Schematic view of the FMS in this research.
estimated (considering tool regrinding), and a tool is replaced
before the time duration the tool has been used exceeds its
in stochastic job shops, and Koo et al. [5] solve a tool tool life. For example, if the estimated tool life of a tool is
provisioning problem in a single-stage multi-machine system 1000 min, it is not usable after it has been used for 1000 min.
by analyzing tradeoffs between the number of tool copies and Therefore, at least a certain number of tool copies are needed
tool waiting time with an open queueing network model in for each tool type for processing a set of parts. If there is a
which tool life and tool transportation time are considered. For worn-out tool in the tool magazine of a machine, a usable tool
part movement FMS’s, Graver and McGinnis [6] introduce of the same tool type is transported to the machine from the
a tool provisioning problem and present a static estimation tool storage area or another machine and is exchanged with
model and a simulation approach for decisions regarding the the worn-out tool.
number of tool copies for a particular application. Chung [7] In this paper, it is assumed that
proposes mathematical models to solve a tool provisioning 1) machines do not fail;
problem, and Atan and Pandit [8] give solution approaches to 2) each tool has a deterministic tool life;
minimize the total number of tools, or to reduce unnecessary 3) no more than one copy of a tool type can be assigned
tool duplications. to a machine;
In tool movement systems, however, research on the tool 4) each tool occupies only one tool slot in the tool mag-
provisioning problem is very rare, if any, although there azine; and
are a few research results on operational problems [9]–[14]. 5) each machine does not start processing a part until all
This paper focuses on a tool provisioning problem in a tool tools required for all operations of the part are loaded
movement system with an automatic tool transporter. The tool on the machine.
provisioning problem considered here is that of determining
To describe the problem more clearly, a mathematical
a tool copy configuration that minimizes the makespan of a
formulation is given below using the following notation:
given set of part types under a tool budget constraint. The
makespan is used as the performance measure in this study
since it is closely related to production rate, throughput, and Indexes
production capacity, which are common long-term or steady-
index for part types
state performance measures for many manufacturing systems.
index for operations
In the next section, the problem considered is described
index for machines
in more detail. Section III presents heuristic algorithms for
index for tool types
the tool provisioning problem. Performance of the proposed
algorithms is evaluated through computational experiments
and results are given in Section IV. Section V concludes the Parameters
paper with recommendations for further research. budget for tools;
purchase cost of one copy of tool type ;
II. PROBLEM STATEMENT demand of part type ;
In this paper, we consider a class of tool movement FMS’s set of operations that are needed for part type ;
that consist of several identical machines, each of which has tool magazine capacity of machine ;
a tool magazine of a limited capacity, a tool storage area that processing time for operation of part type ;
stores tools, and a tool transporter (tool transport robot) that if tool type is used for operation of part type
transports tools between machines or between a machine and and 0 otherwise;
the tool storage area. An example of such an FMS is depicted lower bound on the number of tool copies of tool
in Fig. 1. Each machine can process any operation of any part type ;
type to be processed if the required tools are provided. A upper bound on the number of tool copies of tool
workpiece loaded on a pallet at a load/unload station stays at type ;
a pallet stocker until it can be transported to a machine by a (estimated) tool delay time for which machine
stacker crane. Once a part is loaded on a machine, it stays there waits for a type- tool that is not loaded on
until processing of the part is completed. If a machine does machine under given tool copy configuration
not have all the tools needed for processing a part on its tool .
490 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 3, JUNE 1999

Decision variables methods and the other is a search algorithm using the results
number of tool copies of tool type ; of tool grouping. In both these algorithms, an initial tool copy
if tool type is assigned to machine and 0 configuration is improved by modifying the configuration.
otherwise; Results from simulation of the FMS under a given tool
if part type is assigned to machine and 0 copy configuration are used to generate a new tool copy
otherwise. configuration. This process is repeated until the solution cannot
be improved further under a given tool budget.
Using the above notation, the tool provisioning problem can
In the search procedures of both algorithms, tool waiting
be formulated as the following nonlinear integer program:
time of each tool type (obtained from simulation) is used
for a search direction. Tool waiting times for individual tool
types, which reflect the impacts of the tool types on the system
subject to
performance, are used as a criterion for selecting a tool type
whose number of copies should be increased or decreased. The
for all
number of tool copies of a tool type with a longer tool waiting
(1) time is increased ahead of others, since system performance
for all (2) can be more easily improved by reducing the tool waiting
time when it is larger. Similarly, the number of tool copies
for all (3) of a tool type with a shorter tool waiting time is decreased
ahead of others.
To reduce the search space in the algorithms, bounds on
for all (4)
the number of tool copies of each tool type are used. Since
it is assumed that the tool life of each tool type is known
(5) and constant, one can calculate a lower bound and an upper
bound on the number of tool copies for each tool type. A
and integer for all (6)
lower bound for tool type can be calculated simply as
(7) where is the workload of tool type i.e.,
is the tool life of tool type and
Constraint (1) defines the makespan, i.e., the maximum
denotes the smallest integer not less than Similarly, an
completion time at the machines. The first term on the left
upper bound for tool type can be calculated as the sum
hand side of (1) denotes the workloads of the machines that
of upper bounds (for tool type over the machines. Since
are affected by assignments of parts to machines, while the
is an upper bound for tool type
second term denotes tool delay times that are affected by
at machine the number of tool copies of tool type does
not only and but also the tool copy configuration
not have to be greater than
and scheduling rules or control policies
Since
of parts and tools. Constraints (2) and (3) ensure that each
part type is processed by only one machine and that tool slots
needed for tool types assigned to a machine do not exceed
the tool magazine capacity of the machine, respectively. Con-
straint (4) models the restriction that the total number of tool
copies assigned to machines should not exceed the number of we let in this research. Note that if the number
purchased tool copies for each tool type, while (5) denotes the of tool copies for a tool type is equal to its upper bound,
restriction on the budget for tool purchasing. tool waiting time due to lack of tool copies becomes zero
To solve the problem, and the scheduling rules theoretically. In other words, tool conflicts which cause tool
and control policies of parts and tools should be determined. waiting time do not occur because there are enough tool copies
Since scheduling rules and control policies are not our major of the tool type for each machine to have a copy of its own
concern in this paper, it is assumed that these decisions are installed on its magazine.
made by a certain method such as those suggested in the
literature [15], [16]. Also, we do not consider tool waiting A. Composite Search Algorithm (CSA)
time incurred by replacing worn-out tools, which depends on
tool control policies. This study focuses on the determination In this algorithm, the final solution is obtained by selecting
of values of decision variable i.e., a tool copy configuration a better solution between those from two simulation based
to minimize the second term of the left hand side of constraint greedy search methods, namely, CSA-L and CSA-U, which
(1). start from different initial tool copy configurations. CSA-L
and CSA-U start from lower bounds and upper bounds on the
numbers of tool copies of the tool types, respectively.
III. HEURISTIC ALGORITHMS In CSA-L, the lower bound tool copy configuration, i.e., the
In this paper, two simulation based heuristic algorithms tool copy configuration in which the number of tool copies is
are presented to solve the tool provisioning problem. One equal to be lower bound for all tool types, is used for an initial
is a composite search algorithm based on two greedy search tool copy configuration. Then, the numbers of tool copies of
JUN et al.: HEURISTICS FOR A TOOL PROVISIONING PROBLEM IN A FLEXIBLE MANUFACTURING SYSTEM 491

(more critical) tool types with longer tool waiting times are The basic idea of the above two algorithms is similar to
increased to improve the system performance until the tool those of heuristics for integer programming problems proposed
budget constraint is violated or the performance cannot be by Senju and Toyoda [17] and Kochenberger et al. [18]
improved any further. Note that this initial solution may be
quite poor since tool waiting occurs most frequently under the
lower bound tool copy configuration. However, this solution B. Tool Group Based Search Algorithm (TGSA)
does not violate the tool budget constraint. In this algorithm, tool types are classified into three groups
The procedure of CSA-L is given below, where denotes (A, B, C) based on the impact on the system performance.
the tool waiting time of tool type (estimated by a simulation In general, a tool type with a longer tool waiting time
experiment) and denotes the number of elements in set has a greater (negative) impact on the system performance.
Algorithm CSA-L: Therefore, for the classification of tools, tool waiting time that
Step 1) Let for all is estimated through a simulation experiment with a lower
Step 2) Perform a simulation experiment using the current bound tool copy configuration is used as a measure of the
tool copy configuration, to obtain impact. In the algorithm, tool types are classified into three
an estimate for the makespan, and calculate for equal sized (in terms of the number of tool types in a group)
all groups, A, B, and C, in a nonincreasing order of tool waiting
Step 3) Sort the tool types in in a nonin- time. Tool types with greater impacts are included in group
creasing order of to make a sequence of the tool A. We classified the tool types into three groups (somewhat
type in i.e., where denotes arbitrarily) as in the ABC classification scheme for inventory
the index of the tool type with the th longest tool management, in which items are classified into three unequal-
waiting time. sized classes according to their importance.
In TGSA, lower bounds on the number of tool copies
Step 4) For to do: if there is enough tool budget
are used for an initial tool copy configuration as in CSA-
left and let
L. The initial solution is improved by increasing the number
Step 5) Perform a simulation experiment using to obtain
of tool copies of tool types in groups A, B, and C, in this
an estimate for the makespan. If the makespan does
order, without considering the tool budget constraint until the
not decrease any further or if there is not enough
solution cannot be improved any further. At each time, the
tool budget left, stop. Otherwise, go to Step 3.
same number of tool copies are increased for all tool types
In CSA-U, on the other hand, upper bounds on the number in a group. Then, if the resulting tool copy configuration
of tool copies for the tool types are used for an initial tool does not satisfy the budget constraint, the numbers of tool
copy configuration. Therefore, an initial solution gives the best copies of tool types included in groups C, B, and A, in this
system performance since the tool waiting time does not occur order, are decreased until a feasible solution is found. At this
(theoretically) under such tool copy configuration. However, time, the numbers of copies of tool types are decreased one
the tool budget constraint may not be satisfied. Hence, in the tool type after another in the order of (negative) impact on
algorithm, to make infeasible solutions feasible without having the system performance, which is estimated by a simulation
much (negative) impact on the system performance, the initial experiment as in the other algorithms, until the solution
solution is modified by decreasing the number of copies of becomes feasible.
less critical tool types with a smaller workload until the tool As the number of tool copies in a group is decreased, its
budget constraint becomes satisfied. The procedure of CSA-U (negative) impact on system performance increases, which
is summarized below. may cause rapid deterioration of the solution quality. There-
Algorithm CSA-U: fore, it may be better to reduce the numbers of tool copies
Step 1) Let for all of tool types in less significant groups. To determine when
Step 2) Perform a simulation experiment using the current the number of tool copies of tool types in less significant
tool copy configuration, and groups are reduced, we consider the difference, called the gap,
calculate for all between makespan of the current solution (with reduced tool
Step 3) Sort tool types in in a nondecreas- copies in the current group) and that of the best solution found
ing order of to make a sequence of tool types so far. When finding a feasible solution by decreasing the
in i.e., where denotes the number of tool copies in a group, if the gap is greater than
index of the tool type with the th lowest workload. a predetermined limit, gap the numbers of tool copies in the
Step 4) For to do: if the tool budget constraint current group are not reduced any further. Instead, the next less
is still violated and let significant group (C, B, and A, in this order) is considered for
Step 5) Perform a simulation experiment using and cal- decrease in the numbers of tool copies of tools in the group.
culate If a feasible solution cannot still be found after reduction of
Step 6) For to do: if let tool copies in group A, group C is considered again with an
Step 7) If the tool budget constraint is satisfied or if the increased value for gap A detailed description of TGSA is
number of tool copies for none of the tool types given below.
in was decreased during Step 6, stop. Otherwise, Algorithm TGSA:
let and go to Step 2. Step 1) Let for all and gap
492 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 3, JUNE 1999

Step 2) Perform a simulation experiment using the current handled by a parameterization of the simulation model. A
tool copy configuration, and single replication of simulation is done for each tool copy
calculate s, tool waiting time, and classify tool configuration since there is no variability involved in the
types into three equal sized groups using the values model. What needs to be clarified more are how the parts and
of s. tools are assigned to the machines, and scheduling and control
Step 3) Without considering the tool budget constraint, polices of parts and tools. Assignments of parts and tools must
Step 3-1) Repeat letting for all tools in be done as a setup process before production begins, while part
group A until performance (estimated by a scheduling and tool control are done in real time.
simulation experiment) is not improved any In the simulation model, parts are assigned to machines
further. with the objective of balancing the machine workloads, which
Step 3-2) Repeat letting for all tools in is known to be closely related to minimizing makespan (if
group B until performance is not improved tool waiting time is not considered). To balance machine
any further. workloads, the LPT (longest processing time) rule is used for
Step 3-3) Repeat letting for all tools in part assignment. In this algorithm, part types are first sorted
group C until performance is not improved in nonincreasing order of workloads and then, in this order,
any further. each part type is assigned to the machine with the minimum
Step 4) If the tool budget constraint is satisfied, stop. Other- workload allocated to it so far.
wise, calculate and estimate system performance For tool assignments, a tool of a tool type that is most
through a simulation experiment using frequently used among those that are not yet assigned is
Step 5) Classify tool types into three equal sized groups assigned to the machine which will use the tool for the longest
using the values of s and sort the tool types time and has enough tool slots available in its tool magazine.
included in each tool group in a nondecreasing This rule is repeatedly applied until there are no more tool
order of to make a sequence. Let copies to be assigned. This rule is based on the presumption
and be the indices of the tool types in the that a more frequently used tool is more likely to cause
th position in the sequence for groups A, B, and tool conflicts. Therefore, when selecting a tool to assign to
C, respectively. a machine, it is reasonable to give a higher priority to a tool
Step 6) For 1 to do: that is more frequently used. Also, it is reasonable to assign
If let per- the selected tool to a machine that will use the tool for the
form a simulation experiment using and calculate longest time since tool movements may occur less frequently
gap. If a feasible solution is found, stop. If gap by doing so. (See Amoako-Gyampah and Meredith [19] for
gap , let and go to Step 7. tool assignment procedures in part movement systems.)
Otherwise, proceed for the next Since it is difficult to obtain an optimal schedule for parts
Step 7) For to do: processing (and it is not a major concern in this paper), the
If let per- following dispatching rules are tested and a rule that gives the
form a simulation experiment using and calculate best results is selected and implemented for part scheduling
gap. If a feasible solution is found, stop. If gap in the simulation model.
gap let and go to Step 8. SPT Select a part with the shortest processing time.
Otherwise, proceed. LPT Select a part with the longest processing time.
Step 8) For to do: MTM Select a part that incurs tool movements least
If let per- frequently.
form a simulation experiment using and calculate The tool control policy defines how and when to borrow
gap. If a feasible solution is found, stop. If gap required tools from another machine or the tool storage area
gap let and go to Step 9.
and how and when to return the borrowed tools. Han et al. [11]
Otherwise, proceed.
analyze through a simulation study two tool return policies,
Step 9) Increase gap by 5 percent of makespan of the best
immediate tool return (ITR) policy and no tool return (NTR)
solution found so far and go to Step 5.
policy. Recently, Park and Kim [16] propose another policy,
called the partial tool return (PTR) policy, which gives better
C. Simulation performance than the other two in certain circumstances. They
In this subsection, the simulation model used in the al- report that the PTR policy does not work well when the tool
gorithm is briefly described. The simulation model is deter- magazine capacity constraint is tight, which is a characteristic
ministic, in that nondeterministic factors, such as machine of the FMS considered here. Therefore, in this research, the
breakdowns, tool breakages, and the variability of processing NTR and ITR policies are included in preliminary tests and
times are not considered. Parameters or data needed for the the better will be used in the simulation model.
model, such as the numbers of machines and parts, and In the NTR policy, tools that are borrowed from other
processing times of the operations, are obtained from the data machines are not returned to the machines from which the tools
of the tool provisioning problem. In the suggested algorithms, were borrowed until these machines require them. However, in
changes of these parameters from one problem to another are the ITR, the tools are returned to these machines immediately
JUN et al.: HEURISTICS FOR A TOOL PROVISIONING PROBLEM IN A FLEXIBLE MANUFACTURING SYSTEM 493

TABLE I
COMPARISON OF PART DISPATCHING RULES AND TOOL CONTROL POLICIES

after they are used. If a machine needs to borrow a tool from 9) The tool transportation time from one machine to an
another machine which is using the tool, it must wait until adjacent machine was set to 0.2 time unit and that from
after the tool is used at this machine. In the simulation model, the tool storage area to the nearest machine was set to
when a machine needs to borrow a tool of which multiple 0.6 time unit.
copies are loaded on other machines and/or the tool storage All the algorithms were coded in C, and computational
area, the machine selects and borrows a tool in such a way experiments were done on a personal computer with a Pentium
that the waiting time can be minimized. If the machine cannot processor operating at 200 MHz clock speed. Since there
accommodate a borrowed tool because of the space limit in the does not exist any efficient optimal solution algorithm for the
tool magazine, it temporarily sends tools that are not required problem considered here and it is impossible to find an optimal
for the part to be immediately processed to the tool storage solution with a full enumeration method in a reasonable
area. amount of time, the suggested algorithms were compared with
each other. For the comparison, we use a relative performance
measure called the relative performance ratio (RPR), which is
IV. COMPUTATIONAL EXPERIMENTS defined as for algorithm for each problem,
To compare performance of the proposed heuristic algo- where is the solution value (makespan) obtained from
rithms, computational experiments were done on a number of algorithm is the best solution value obtained by the
test problems. All data used for experiments were generated tested algorithms for the problem. This index reflects how well
based on real data in an FMS of a manufacturing company a particular heuristic algorithm performs relative to others.
in Korea. For the experiments, ten problems were generated As mentioned earlier, the best combination of part dis-
randomly for each set of all (36) combinations of two levels patching rules and tool control policies are to be used for
for the number of part types (20 and 25), two levels for the comparison of the algorithms for the tool provisioning
the number of operations (15 and 18), three levels for the problem. To select the best combination, all combinations of
number of machines (4, 6, and 8), and three levels for the the three dispatching rules (SPT, LPT, MTM) and the two
number of tool types (140, 160, and 180). Other data used for policies (NTR, ITR) were tested on a set of 36 problems
the problems were generated from probability distributions as generated randomly as described above. All the suggested
follows. DU(a, b) denotes the discrete uniform distribution algorithms were used in this series of tests. Results of the
with a range from a to b. tests are shown in Table I, which gives the mean and standard
1) The demand quantity (requirement) for each part type deviation of RPR’s. The combination of MTM and ITR gave
was generated from DU(15, 40); the best results for all the tool provisioning algorithms. The
2) The number of operations for each part type was gen- effects of the tool provisioning algorithms on the system per-
erated from DU(1, 6); formance were more significant than those of part dispatching
3) Processing time for each operation was generated from rules and tool control policies, as depicted in Fig. 2. Note
DU(20, 420); that the differences of RPR’s of the best and the second best
4) The number of tool types needed for each operation was rules/policies/algorithms were larger for the tool provisioning
generated from DU(25, 40); algorithms than for the part dispatching rules and tool control
5) Tool life of each tool type was generated from DU(3000, policies.
8000); Since the combination of MTR and ITR worked better
6) Cost of a tool was generated from DU(1, 10); than others, they were selected for a main experiment to
7) The budget for purchasing additional tools (besides those compare the tool provisioning algorithms. Table II shows the
for the lower bound tool copy configuration) was set to mean and standard deviation of the RPR, the number of
be one tenth of the sum of unit costs needed to purchase problems for which each algorithm found the best solutions,
one copy for each of all tool types; and computation time. Overall, TGSA outperformed CSA
8) The tool magazine capacity was set to 80 for each in terms of all the measures, RPR, the number of best
machine; solutions and computation time. Although CSA selects the
494 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 3, JUNE 1999

TABLE II
PERFORMANCE OF THE HEURISTICS (RESULTS FROM PROBLEMS OF PRACTICAL SIZES)

(a)
Fig. 3. RPR’s for different numbers of machines.

performance. Both algorithms require CPU times of less than


an hour on a personal computer for solution of a problem. (A
simulation run required less than 8 s.) From a computational
perspective, these CPU times seem reasonable since the tool
provisioning problem would not be solved very often.
As the number of machines increases, the superiority of
TGSA over CSA becomes clearer (see Fig. 3). This may be
explained as follows. As the number of machines increases, the
solution space becomes larger due to the increase in the upper
(b) bounds on the numbers of tool copies. Also, tool conflicts
Fig. 2. Comparison of the heuristics (practical sizes). (a) Overall av- occur more often since it is more likely that more than one
erage RPR’s. (b) Average RPR’s given the best combination for each
rule/policy/algorithm.
machine requires the same tool at the same time when there
are more machines in the FMS. The better algorithm (TGSA)
can search large solution space more efficiently and resolve
better solution from the two greedy search methods (there was tool conflicts more effectively and the performance difference
no significant difference in performance of the two greedy becomes clearer in such cases.
search methods), it did not perform an well as TGSA. This To see the absolute (not relative) quality of solutions from
shows the effectiveness of tool grouping in TGSA, which the suggested algorithms, heuristic solutions were compared
classifies tool types according to the impact on the system with optimal solutions. For the purpose, small sized test
JUN et al.: HEURISTICS FOR A TOOL PROVISIONING PROBLEM IN A FLEXIBLE MANUFACTURING SYSTEM 495

TABLE III
PERFORMANCE OF THE HEURISTICS (RESULTS FROM SMALL SIZED PROBLEMS)

problems were generated and optimal solutions were obtained customized through a slight modification to deal with special
by a full enumeration method, in which all tool copy con- characteristics of the systems.
figurations were considered. MTM and ITR were used for This research can be extended in several directions. For
part scheduling and tool control, respectively. For this test, 80 example, one may develop a procedure which simultaneously
problems were randomly generated, 10 problems for each of 8 determines a tool copy configuration, assignment of parts
sets (see Table III for a description of these sets). The results and tools, and scheduling or control policies of parts and
are shown in Table III, which gives the number of optimal tools to maximize the overall system performance under a
solutions each algorithm found, average and standard deviation given budget constraint. Also, other statistics (rather than tool
of percentage errors. For each problem, the percentage error waiting time) can be used to estimate the impact of tools on the
of algorithm is defined as where is system performance such as frequency of tool movements for
the solution value obtained using algorithm and is the each tool type and the frequency of tool conflicts. Of course,
optimal solution value. In this case too, TGSA works better a more improved and efficient search method may be devised
than CSA in terms of the solution quality. Both algorithms and used in the suggested algorithms. On the other hand, other
found optimal solutions in many cases (in about 70% of the objectives such as those related to flow time or due dates may
problems). Even when they did not find optimal solutions, they be considered. For such objectives, one may use the approach
gave good solutions which are close to optimal. suggested in this paper if the measure used for simulation in
the suggested algorithms is changed to the one related to the
objective of the problem.
V. CONCLUSION
We considered a tool provisioning problem to find a tool REFERENCES
copy configuration which minimizes the makespan for a given
set of part types under a tool budget constraint in a class of [1] D. Veeramani, D. M. Upton, and M. M. Barash, “Cutting-tool manage-
ment in computer-integrated manufacturing,” Int. J. Flexible Manufact.
FMS’s. The FMS’s considered in this paper are tool movement Syst., vol. 4, no. 3, pp. 237–265, 1992.
systems, in which each part visits only one of the machines [2] A. E. Gray, A. Seidmann, and K. E. Stecke, “A synthesis of decision
for its entire processing and tools which are not loaded on the models for tool management in automated manufacturing,” Manag. Sci.,
vol. 39, no. 5, pp. 549–567, 1993.
machine but required to produce the parts are borrowed from [3] P. Tomek, “Tooling strategies related to FMS management,” FMS Mag.,
other machines or the tool storage area by a tool transporter. vol. 4, pp. 102–107, 1986.
To solve the problem, we suggested two heuristic algorithms [4] J. Wang, J. Yang, and V. B. Gargeya, “Tool requirement planning in
stochastic job shops: A simulated annealing approach,” Comput. Indust.
in which a good tool copy configuration is found by using Eng., vol. 24, no. 2, pp. 249–265, 1993.
tool waiting time for each tool type that is estimated by a [5] P.-H. Koo, J. M. A. Tanchoco, and J. J. Talavage, “Tool requirements in
simulation experiment. Test results showed that the tool group manufacturing systems under dynamic tool sharing,” in Proc. 20th Int.
Conf. Comput. Industr. Eng., 1996, pp. 1271–1274.
based search algorithm (TGSA), which finds the best tool [6] T. W. Graver and L. F. McGinnis, “A tool provisioning problem in an
copy configuration after grouping tool types according to their FMS,” Int. J. Flexible Manufact. Syst., vol. 1, pp. 239–254, 1989.
[7] C. H. Chung, “Planning tool requirements for flexible manufacturing
impacts on the system performance, gave the better results systems,” J. Manufact. Syst., vol. 10, no. 6, pp. 476–483, 1991.
than the composite search algorithm (CSA), which consists of [8] T. S. Atan and R. Pandit, “Auxiliary tool allocation in FMS,” Euro. J.
two algorithms that determine directly the number of copies Oper. Res., vol. 89, no. 3, pp. 642–659, 1996.
[9] A. Agnetis, M. Dror, A. J. Vakharia, and F. Rossi, “Tool handling and
for each tool type. The approach used in this study can be scheduling in a two-machine flexible manufacturing cell,” IIE Trans.,
applied to most tool movement FMS’s if the heuristics are vol. 28, pp. 425–437, 1996.
496 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 3, JUNE 1999

[10] H. S. Hahn and J. L. Sanders, “Performance analysis of a LIM- Yeong-Dae Kim (M’98) received the B.S. degree in
based high-speed tool delivery system for machining,” Int. J. Flexible industrial engineering from Seoul National Univer-
Manufact. Syst., vol. 32, no. 1, pp. 179–207, 1994. sity, Korea, the M.S. degree in industrial engineer-
[11] M.-H. Han, Y. K. Na, and G. L. Hogg, “Real-time tool control and job ing from the Korea Advanced Institute of Science
dispatching in flexible manufacturing systems,” Int. J. Prod. Res., vol. and Technology (KAIST), Taejon, and the Ph.D.
27, no. 8, pp. 1257–1267, 1989. degree in industrial and operations engineering from
[12] H.-K. Roh and Y.-D. Kim, “Due-date based loading and scheduling the University of Michigan, Ann Arbor.
methods for a flexible Manufacturing system with an automatic tool He is a Professor in the Department of Industrial
transporter,” Int. J. Prod. Res., vol. 35, no. 11, pp. 2989–3003, 1997. Engineering, KAIST. His research areas include
[13] C.-Y. Song, H. Hwang, and Y.-D. Kim, “Heuristic algorithm for the design and operation of manufacturing systems,
tool movement policy in flexible manufacturing systems,” J. Manufact. operations scheduling, and production and inventory
Syst., vol. 14, no. 3, pp. 160–168, 1995. management.
[14] O. S. Susan and C. Jeya, “Evaluation of alternative tool combinations Dr. Kim is a senior member of IIE, a member of INFORMS, and the
in a flexible manufacturing system,” Comput. Indust. Eng., vol. 26, no. Operational Research Society.
4, pp. 633–645, 1994.
[15] Y.-D. Kim and C. A. Yano, “Heuristic approaches for loading problems
in flexible manufacturing systems,” IIE Trans., vol. 25, no. 1, pp. 26–39,
1993.
[16] S. S. Park and Y.-D. Kim, “A heuristic algorithm for tool loading and Hyo-won Suh received the M.S. degree in mechan-
scheduling in a flexible manufacturing system with an automatic tool ical engineering from the Korea Advanced Institute
transporter,” J. Korean Inst. Indust. Eng., vol. 21, no. 1, pp. 119–135, of Science and Technology (KAIST), Taejon, and
1995. the Ph.D. degree in industrial engineering from West
[17] S. Senju and Y. Toyoda, “An approach to linear programming with 0–1 Virginia University, Morgantown.
variables,” Manag. Sci., vol. 15, no. 4, pp. 196–207, 1968. He is an Assistant Professor at the Department
[18] G. A. Kochenberger, B. A. McCarl, and F. P. Wyman, “A heuristic for of Industrial Engineering, KAIST. His research ar-
general integer programming,” Decision Sci., vol. 5, no. 1, pp. 36–44, eas include concurrent engineering and computer
1974. integrated manufacturing system.
[19] K. Amoako-Gyampah and J. R. Meredith, “A simulation study of FMS
tool allocation procedures,” J. Manufact. Syst., vol. 15, no. 6, pp.
419–431, 1996.
[20] G. L. Nemhauser and L. A. Wolsey, Integer and Combinatorial Opti-
mization. New York: Wiley, 1988.
[21] R. Macchiaroli and S. Riemma, “Design of a tool management system
in a flexible cell,” Int. J. Prod. Res., vol. 34, no. 3, pp. 767–784, 1996.

Hong-Bae Jun received the B.A. degree in eco-


nomics from Yonsei University, Korea, and the M.S.
degree in industrial engineering from the Korea Ad-
vanced Institute of Science and Technology, Taejon,
where he is currently pursuing the Ph.D. degree.
His research areas include production planning
and operations scheduling in flexible manufactur-
ing systems, concurrent engineering, and computer
integrated manufacturing system.

You might also like