You are on page 1of 14

0

Journal of Manufacturing Systems


M Vol. 2olNo. 4
~ 2001

Modeling and Analysis of Tandem AGV Systems


Using Generalized Stochastic Petri Nets
lgnacio Castillo, Dept. of Finance and Management Science, University of Alberta School of Business,
Edmonton, Alberta, Canada. E-mail: ignacio.castilloQualberta.ca
Sergio A. Reyes and Brett A. Peters, Dept. of Industrial Engineering, Texas A&M University,
College Station, Texas, USA. E-mail: sergiora, bpetersQtamu.edu

Abstract ify and validate. Moreover, in terms of model build-


This paper presents a Petri net modeling approach for ing and interpretation of results, discrete event sim-
analyzing tandem AGV systems. The approach allows the
ulation requires specialized training and experience.
consideration of several critical design issues such as multi-
load vehicles, multi-machine processing workstations, finite On the other hand, several researchers have con-
queue capacities, and constant work-in-process operating centrated their efforts on developing basic academic
policies. The proposed approach is capable of providing ana- research on AGV systems. This type of research
lytical as well as simulative results and allows tandem AGV
assumes that there exists aggregated information
systems to be fully explored within a single modeling frame-
work. Analytical analysis can be performed if the interarrival available and uses analytical tools to analyze AGV
and service times are assumed to be exponentially distrib- systems. Analytical tools allow the computation of a
uted. Simulative results do not require this restriction, and limited number of performance metrics and often
thus general distributions can be used.
apply only under certain modeling assumptions and
constraints. The use of these tools when the assump-
Keywords: Material Handling Systems, Tandem AGV
Systems, Petri Net Modeling and Analysis, Analytical and tions and constraints do not hold may lead to impre-
Simulative Methods cise results and questionable conclusions.
In this paper, a Petri net approach for modeling
and analyzing the class of AGV systems introduced
Introduction by Bozer and Srinivasan (199 1) is presented. The
Automated guided vehicle (AGV) systems pro- proposed approach provides analytical as well as
vide effective and efficient asynchronous movement simulative results with much of the modeling flexi-
of material in low to medium-flow manufacturing bility of discrete event simulation. The approach
systems. The advantages of AGV systems relative to allows the consideration of critical design issues
other types of material handling systems include: such as multi-load vehicles, multi-machine process-
reliable and automatic operation, improved position- ing workstations, finite queue capacities, and con-
ing accuracy, rapid and flexible response to changes stant work-in-process (CONWIP) operating poli-
in the material handling requirements, and easily cies. The analytical analysis alternative requires the
expandable layout and system capacity, among oth- assumption of exponentially distributed interarrival
ers (Peters, Smith, and Venkatesh 1996). and service times. While this assumption is very
The modeling and further performance analysis limited, it is needed for analytical tractability.
of AGV systems, nonetheless, can be very compli- Complexity in the computation of the model solu-
cated due to the number of issues that must be con- tion stems from the very large state spaces typical of
sidered, such as the guidepath layout, the number real-world Petri net models. The analytical analysis
and type of vehicles, the buffer capacities, and the alternative permits the computation of exact, yet
operating policies. Consequently, many vendors, conservative, results when the requirements of CPU
users, and researchers employ discrete event simula- time and main memory are met. Whenever the
tion as the primary tool in analyzing AGV systems. requirement of main memory is exceeded the simu-
Discrete event simulation allows the computation of lative analysis alternative permits the computation
numerous performance metrics quite easily. of point estimates. The simulative analysis alterna-
Detailed discrete event simulation models, however, tive also permits the use of general distributions for
are complex computer programs that are hard to ver- both interarrival and service times, thus allowing

236
Journal of Manufacturing Systems
Vol. 2a/No. 4
2001

realistic tandem AGV systems to be modeled and Interface station7 AGV loop--,
analyzed. Additionally, the use of Petri nets, in gen-
eral, allows the representation of concurrency, syn-
chronization, and resource sharing-properties that
cannot be expressed easily in traditional formalisms
(e.g., queuing networks).

-P/D 2 2

Background

AGV Systems: Conceptual Issues


AGVs are self-propelled vehicles that follow a
path determined in one of two ways: static a priori
determination or dynamic real-time determination.
Traditionally, the design and performance analysis Workstationas interface station

of AGV systems assumes that every vehicle is Figure 1


Tandem AGV System
allowed to visit any pickup/deposit (P/D) point in
the system. These conventional AGV systems ty. The modular configuration allows the addi-
require extensive control systems for vehicle dis- tion of loops with minimal disruptions to the
patching, routing, and traffic management. existing system.
Moreover, conventional systems are prone to con-
gestion when planning is poorly conducted and/or Tandem AGV systems, however, also present cer-
when material flow requirements increase. tain drawbacks that must be addressed. For instance,
In Bozer and Srinivasan (1991), an alternative tandem AGV systems will show less tolerance
configuration for the design and performance analy- toward vehicle breakdowns, thus requiring one or
sis of AGV systems is presented. The authors base more stand-by vehicles. Also, loads may have to be
their approach on partitioning all the workstations handled by more than one vehicle before they reach
into non-overlapping, single-vehicle zones with their destination. This requires the determination of
additional P/D points provided as an interface the shortest travel time route between the origin and
between zones or the external environment. Note the destination. In Lin and Dgen (1994), a task-list
that a single vehicle operating in a zone can be con- time-window algorithm to address this problem
sidered a single-server polling system, where the based on the current status of the system is present-
server polls the queues in a nondeterministic order. ed. Moreover, the efficient operation of tandem sys-
Thus, a zone may be also considered as a closed tems requires balancing the workload among loops.
loop. In this paper, single-vehicle zones are referred The issue of system layout and partitioning has been
to as single-vehicle closed loops (see Figure I). The a matter of research in Bozer and Park (1992), Bozer
advantages of this tandem configuration relative to and Srinivasan (1992), Hsieh and Sha (1996), and
the conventional design are: Hsieh and Sha (1997). In Hsieh and Sha (1996) and
Hsieh and Sha (1997), an iterative procedure to
l Congestion due to zone blocking and conflicts at determine the workstation partition and layout for
intersections and P/D points are completely single-load tandem AGV systems is presented. The
eliminated. procedure starts with two closed loops and iterates,
l The control system for vehicle dispatching increasing the number of loops by one, until the
and traffic management is significantly less available time of the vehicle exceeds the loaded trav-
complicated. el time of each vehicle within a loop.
l Because each closed loop is a module of the sys- Other design issues, such as multi-load vehicles,
tem, the implementation of hierarchical control, multi-machine processing workstations, finite queue
where independent loop controllers communicate capacities, and constant work-in-process (CON-
with a centralized control system, is facilitated. WIP) operating policies, have not been considered
l Tandem AGV systems also offer more flexibili- for tandem AGV systems. The proposed approach

237
Journal of~anufac~~~rjng Systems
Vol.20Mo.4
2001

allows the consideration of such issues by providing cal Petri net approach to analyze the perfo~ance of
analytical as well as simulative analysis alternatives tandem AGV systems is presented in Wang and
within a single robust modeling approach. Hafeez (1994). In this study, the authors consider a
simplified system in which each loop contains only
AGV Systems: Analysis Issues one workstation and a single-load vehicle. The
The perfo~ance analysis of conventional AGV authors, however, do not indicate how the Petri net
systems has motivated several simulation-based model can be modified to incorporate other critical
studies to evaluate the effects of different design design issues such as multiple workstation loops and
issues, such as the number of AGVs, dispatching multi-load vehicles.
policies, and queue capacities (Egbelu and Tanchoco The potential to improve vehicle availability for
1984, 1986; Ozden 1988; and Schroer and Tseng load transport requests and the further reduction of
1988). In Vosniakos and Mamalis (1990), issues deadhead or unproductive time of vehicles when
such as zone blocking, loading and unloading, and incorporating multi-load carrying capacity has been
traffic control are also addressed. Simulation-based acknowledged (Bilge and Tanchoco 1997, Nayyar
Petri net approaches for conventional AGV systems and Khator 1993). These and other design issues
have been presented in Alanche et al. (1984) and such as finite queue capacities and CONWIP oper-
Davis, Jackson, and Jones (1989). More recently, in ating policies, however, have not been previously
Venkatesh et al. (1996) and Yim and Linn (1993), addressed in the literature for the case of tandem
push and pull policies in flexible manufacturing sys- AGV systems. Thus, there is an obvious need for
tems that include AGVs using the simulation-based further analysis to fully explore tandem AGV sys-
Petri net approach are investigated. The times asso- tems; only then can fair and quantitative compar-
ciated with transportation delays cannot be properly isons with conventional AGV systems be conducted,
addressed in the aforementioned simulation-based
Petri net approaches due to the dynamic nature of
AGV assignments in conventional AGV systems. In Petri Nets: Basic definitions
Raju and Chetty (1993), a simulation-based Petri net Petri nets make it possible to model discrete event
approach that addresses this issue by extending the systems. Generalized stochastic Petri nets (GSPNs)
Petri net formalism is presented. The extension are a performance analysis tool based on the graph-
relates transportation delays with a data structure ical representation of Petri nets in which some tran-
that records the state and location of the AGVs. sitions are timed while others are immediate (Chiola
Analytical approaches for designing conventional et al. 1993, David and Alla 1992, Murata 1989,
AGV systems include Egbelu (1987), Mahadevan Peterson 198 1). Timed transitions are associated
and Narendran (1990), and Malmborg (1990). In with exponentially distributed firing times, whereas
Maxwell and Muckstadt (1982), a general design the firing of immediate transitions takes place in
methodology for single-load vehicles operating on a zero time, with precedence over timed transitions.
unidirectional path is presented. Bartholdi and Each transition is associated with a priority level.
Platzman (1989) present a simple and efficient dis- The default priority value for immediate transitions
patching policy, First-Encountered-First-Served is 1; timed transitions have priority 0. Higher priori-
(FEFS). This dispatching policy can be applied to ty levels take precedence over lower priority levels.
the cases of single and multiple vehicles operating in GSPNs are widely accepted as high-level for-
a closed loop. The authors present results for the malisms to define large and complex continuous-
number of traversals of the loop and derive a condi- time Markov chains (Ajmone Marsan et al. 1995).
tion under which the system will meet the required GSPN models, in general, are graphically represent-
throughput for a set of load arrival patterns. The spe- ed as directed bipartite graphs that consist of two
cific case for analyzing tandem AGV system con- types of nodes: places, which are drawn as circles,
figurations is studied in Bozer and Srinivasan and immediate (timed) transitions, which are drawn
(1991). This work is concerned with the particular as bars (boxes). Places typically model conditions or
situation of a single-load vehicle operating in a resources; transitions model events or activities.
closed loop. The authors develop an analytical Directed arcs connect places to transitions and tran-
model to study the throughput capacity. An analyti- sitions to places. Places may contain tokens, which

238
Journal of Manufacturing Systems
Vol. zo/No. 4
2001

are drawn as black dots. The state of the Petri net is allows the coordination of processes (e.g.,
characterized by the number of tokens in each place. pickup/deposit operations and input/output queue
The state is usually called marking. The initial mark- inspections); and resource sharing allows modeling
ing determines the initial state of the Petri net. In the of products competing for capacitated resources.
graphical representation, the number of tokens in In the development of the model, a set of consid-
places might be replaced by the name of a parameter. erations has been taken into account. The manufac-
The structure of a Petri net is formally described turing system consists of workstations, which repre-
by the input, output, and inhibitor functions. These sent one machine or a group of machines. Each
functions map transitions onto places. In the graphi- workstation has two queues: input and output. Loads
cal formalism, the input and output functions are rep- that are destined to a workstation are dropped at its
resented by directed arcs from places to transitions input queue, while the loads that are leaving the
and from transitions to places, respectively. The workstation are picked up from its output queue. The
inhibitor function is represented by circle-headed input queue precedes the output queue from the
arcs connecting places to transitions, and it is used to AGV traveling direction standpoint. These queues
inhibit the firing of transitions given a marking. may have a limited capacity.
Also, a function defined on the transitions maps There are two types of P/D points: those relat-
transitions into positive real numbers. This function ed to the workstation queues and those related to
is used to represent the rate of a transition if it is the input/output (I/O) stations (that is, P/D points
timed and the weight of a transition if it is n-imme- that serve as interface between adjacent loops or
diate, where n is the priority level. The association of the external environment). Thus, loads from adja-
firing rates with timed transitions implies that an cent loops or the external environment arrive at
exponentially distributed firing time with a given the output queue of an I/O station where they wait
rate is associated with the transition. The association to be picked up. Likewise, loads that require no
of weights with n-immediate transitions is necessary further processing within a loop are delivered to
to determine which of the enabled transitions will the input queue of an I/O station where they leave
actually fire in a marking that enables more than one the loop or system.
conflicting n-immediate transitions. It is assumed that the workstations in the system
A Petri net executes by firing transitions. The fir- have been partitioned in advance into non-overlap-
ing of a transition yields a new marking such that ping, single-vehicle closed loops. In each closed
tokens are removed from the input places and creat- loop, a multi-load AGV travels from one P/D point
ed in the output places. A marking is said to be to another. The AGV is assumed to follow the FEFS
reachable from another marking if there exists a vehicle dispatching policy. Under this policy, if there
sequence of transitions such that the firing of such is capacity available in the vehicle, the AGV will
sequence yields the desired marking. Readers are travel to inspect the output queues of each worksta-
referred to Chiola et al. (1993), David and Alla tion and I/O station in the loop to locate the next
(1992), Murata (1989), and Peterson (1981) for load requiring transport. Note that, under this policy,
more detailed information on Petri nets. the AGV is always traveling either loaded, partially
loaded or unloaded.
Also, it is assumed that the production plan and
Tandem AGV System Model the production routing are known. Finally, flow con-
servation is not required at the workstations or I/O
Considerations and Assumptions stations; that is, the number of loads entering a sta-
Petri nets provide a powerful ability to represent tion is not necessarily equal to the number of loads
concurrency, synchronization, and resource sharing. leaving it.
In this paper, these properties are exploited in a
modular and flexible modeling approach to analyze Modeling Approach
tandem AGV systems. In the context of tandem The Petri net modeling approach for tandem AGV
AGV systems, concurrency implies that the system systems consists of three basic modules. The mod-
has the potential to execute workstation processing ules are used to build the logic of the dynamics of
and vehicle movement in parallel; synchronization the tandem AGV system. Thus, by combining them

239
Journal of Manufacturing Systems
Vol. 2omo. 4
2001

Table I
Legend for Petri Net move Module

Place Description
pin AGV positioned at input queue
(deposits loads if necessary)
pout AGV positioned at output queue
(inspects and picks up loads if necessary)
pTrave1 AGV traveling to next P/D point
Timed Transition Description
tTrave1 Travel completed
Immediate Transition Description
tin Deposit completed (if necessary)
tout Inspection and pickup completed
(if necessary)
tLogic Logical transition to resolve case of
no capacity available in AGV
Figure 2
Tandem AGV Petri Net move Module ready to be picked up. The logic associated with
the presence of a load and the delay incurred in
any particular tandem system can be modeled. The the loading are modeled in the transfer module.
modules are: move, transfer, and process. l Place pTrave1 represents the traveling of the
AGV to the next P/D point in the loop. The trav-
The move Module el time is accounted by timed transition tTrave1.
The move module is related to the AGV traversal The firing of this transition represents the arrival
of the loop. The graphical Petri net representation of of the AGV to the input queue of the next P/D
the move module is shown in Figure 2. Table I gives point in the loop. This firing results in the cre-
the description of the module’s places and transi- ation of a token in place pin of the next P/D
tions. This module is the building block of each tan- point to be visited. If the loaded and unloaded
dem loop. Thus, the module is used to represent the speeds of the AGV are equal, transition tTrave1
physical guidepath traversed and the P/D points vis- is assigned a real number as the firing rate. If the
ited by the AGY The functionality of the module is loaded and unloaded speeds are different, how-
described next. ever, a marking dependent rate definition must
be defined.
l Place pin models the presence of the AGV at the l A sequence of concatenated move modules mod-
input queue of a P/D point (either workstation or els a loop, its P/D points, and the movement of the
I/O station). Assume that a token has been creat- AGV within it. Thus, the logic of the preceding
ed in this place. If the AGV is carrying a load list is repeated for each P/D point. The concate-
destined to that point, the load is deposited in the nation of move modules is called loop subnet.
queue. Once the load has been deposited, transi-
tion tin fires. The delay incurred in the unload- The transfer Module
ing is modeled in the transfer module. The transfer module is related with the logic asso-
l After the firing of transition tin, a token is cre- ciated with each product type and their interaction
ated in place pout. This place models the pres- with the P/D points. This module is the core of the
ence of the AGV at the output queue of a P/D modeling approach. It is the building block to model
point. At this point, the AGV inspects the queue. the dynamics of the FEFS dispatching policy, the
If there is a load waiting to be picked up and capacity available in the multi-load AGV, the pro-
capacity available in the AGV, the AGV picks up duction routing of each product type, the presence of
the load and transition tout fires. If there is no loads in the input and output queues of P/D points,
capacity available in the AGV, the AGV contin- the loading and unloading of loads, and the traveling
ues traveling and transition tlogic, instead of of the loads from one P/D point to the next accord-
transition tout, fires. That is, transition tLogic ing to their respective routings. Moreover, most of
allows the AGV to continue traveling to the next the enhancements presented in the sequel are
P/D point, disregarding the presence of a load achieved by extending this module. The graphical

240
Journal of Manufacturing Systems
Vol. 20Mo. 4
2001

Table 2
Legend for Petri Net transfer Module
Place Description
pAGV Capacity available in AGV
POutQ Load present at output queue of station (i)
@NOI pToInQ Load traveling to next P/D point
@Q Load present at input queue of station (i)+l
tOut[(i)] Timed Transition Description
tPre Load deposited at output queue of station (i)
6- plnKO+ll tLoad Loading completed
- tLogic((i)]
tUnload Unloading completed
tln[(i)+l] Immediate Transition Description
tpost Load transferred to workstation, or
load left the loop or system

Figure 3
Tandem AGV Petri Net transfer Module The firing of transition tUnload will create a
token in place pAGV, thus restoring the capacity
Petri net representation of the transfer module is of the AGV by one.
shown in Figure 3. In this figure, places pOut[(i)] The firing of transition tUnload will also create
and pIn[(i)+l], and transitions tOut[(i)], tLogic[(i)], a token in place pInQ. This token represents the
and tIn[(i)+l] correspond to the move module asso- presence of a load at the input queue of station
ciated with P/D point (i) or (i)+l. Table 2 gives the (i)+l in the routing of a particular product type.
description of the module’s places and transitions. The firing of transition tPost occurs either if the
The functionality of the module is described next. load is transferred to a workstation or if the load
leaves the loop or system.
l Place pAGV models the capacity available in the A sequence of concatenated transfer modules, with
AGV. That is, pAGV will have as many tokens as aprocess module in between, models each product
loads the AGV can carry. The existence of type, its production routing, and its interaction with
tokens in this place contributes to enabling tran- the P/D points and the AGV under the FEFS dis-
sition tLoad and inhibits the firing of transition patching policy. The concatenation of transfer and
tLogic of station (i), in the move module, when process modules is called product subnet.
there is capacity available.
l Assume that transition tPre has fired and a token The process Module
has been created in place pOutQ. This token rep- Theprocess module is related to the workstations.
resents the presence of a load at the output queue Because this module models the processing of loads,
of station (i) in the routing of a certain product it will always be located between transfer modules.
type. The existence of tokens in this place The graphical Petri net representation of the process
inhibits the firing of transition tout in the move module is shown in Figure 4. Table 3 gives the
module. Under this condition, if the AGV hap- description of the module’s places and transitions.
pens to be inspecting the output queue of station The functionality of the module is described next.
(i) and it has capacity available, the AGV will
pick up the load and timed transition tLoad will l Place pWS models the capacity available in the
fire after the loading time has elapsed. workstation. The existence of a token in this place
l After the firing of transition tload, a token is corresponds to an idle workstation. This condition
created in place pToInQ. This place models the contributes to the enabling of transition tPost. If
traveling of the load from station (i) to station there is a token in place pInQ of the transfer mod-
(i)+l in the routing of a certain product type. ule, transition tPost will fire. Thus, a load is trans-
The existence of tokens in this place inhibits the ferred to the workstation to be processed.
firing of transition tin in the move module. l Place pProcess models the processing of a load.
Under this condition, if the AGV happens to be The firing of timed transition tPre after the pro-
at the input queue of station (i)+ 1, the AGV will cessing time has elapsed restores the capacity of
deposit the load and timed transition tUnload the workstation and creates a token in place
will fire after the unloading time has elapsed. pOutQ of the transfer module.

241
Journal of Manufacturing Systems
Vol. 20Mo. 4
2001

1 Table 3
Legend for Petri Net processModule
Place Description
PWS Capacity &ail&e in workstation
pProcess Workstation processing a load
Timed Transition DescriGtion
tPre Load depositeh at &put queue of station (i)
pProcess
Immediate Transition _ -- Description
tPost Load transferred to workstation, or load
left the loop or system
tPre

Finite Queue Capacities. Finite queue capacities can


be modeled if it is assumed that a transfer from one
Figure 4
Tandem AGV Petri Net press Module
P/D point to the next in the production routing of a
certain product type can only be performed if capac-
Modeling Enhancements ity in the destination P/D point is reserved. This pol-
The modeling enhancements are achieved by icy prevents the system from getting into a dead-
extending or modifying the three basic modules locked state. The extension requires the addition of
described above. The enhancements considered are: a place that models the queue capacity (i.e., place
pQ), an input arc from place pQ to transition tload,
No Conservation of Flow. The condition of no con- and an output arc from transition KInload back to
servation of flow is achieved by modifying the input place pQ.
and output functions of the Petri net model. In par-
ticular, the modification involves the multiplicity of ~~direct~ona~ Loops with Shortcuts. For the case of
input arcs from pInQ places to tPost transitions and single-load AGVs, bidirectional loops with shortcuts
the multiplicity of output arcs from tPre transitions can be modeled. The modification requires the trav-
to pOutQ places. el time to be modeled in the transfer module rather
than in the move module because the travel time will
CONWIP Operating Policies. CONWIP operating depend on whether the AGV is loaded or unloaded.
policies are achieved by adding one place for each
product type or for each tandem loop under consider-
ation. Let such a place for a certain product type be The proposed Petri net modeling approach and
called pWIF? The number of tokens in this place rep- the proposed enhancements allow the consideration
resents the level of work in process set for that partic- of several critical design issues for modeling and
ular product type. The logic of the CONWIP policy is analyzing tandem AGV systems. These issues have
introduced by adding an input arc from place pWIP to not been addressed by any analytical approach
the first transition of the respective product subnet, developed earlier, As an example, consider the Petri
and an output arc from the last transition of the respec- net shown in Figure 5. In this figure, a simple tan-
tive product subnet back to place pWIF! In the case of dem AGV system is modeled. The system has one
CONWIP operating policies for a loop instead of a loop. The loop is composed of two workstations
product, place pWIP becomes a colon input place W;Sl and WS4 modeled by places pWS1 and pWS4,
for the first transition of all product subnets and a respectively, and one I/O station, namely I/05. The
common output place for the last transition of all prod- capacity of workstation Ws4 is 1, while the capacity
uct subnets back to place pWII? Similarly, CONWIP of workstation WSl is 2. The physical configuration
operating policies for the entire system are possible by of the loop, with respect to the traveling direction of
extending the aforementioned logic. the single-load AGV, is ( WSl, WS4,UOS).
The single product subnet models the production
Multi-Machine Workstations. A multi-machine routing of product B. The production routing of
workstation is simply modeled by setting the num- product B is (I/05, WS4, WSI). That is, loads enter
ber of tokens in pWS equal to the number of the loop through I/O station 1/05 and are processed
machines in the workstation. by workstation Ws4 followed by workstation WSi.

242
Journal qf Manufacturing Systems
Vol. 2a/No. 4
2001

K=5 Petri net is defined as a labeled directed graph,


-,
/
whose set of nodes is the reachability set and whose
set of arcs represents all possible transition firing
tLoadB5 relations between pairs of markings.
Numerical algorithms for the solution or simula-
pTolnQB54
tion of GSPN models require the structure and the
tUnloadB4 state space of the net to satisfy certain properties. In
plnPB4 the case of the analytical alternative, the net model
must be bounded (have a finite state space) and must
tPoetB4
be live (no transitions can become unfirable for all
pProcessB4
reachable markings). A necessary and sufficient
tPreB4 condition for the net model to be bounded is that all
pOutQB4
places are covered by some P-invariant. A necessary
condition for the net to be live is that all transitions
tLoadB4
are covered by T-invariants (see Martinez and Silva
pTolnQB41 198 1 for a complete treatment of P- and T-invari-
tUnloadB1
ants). To discover if the Petri net model is actually
live, the structure of the reachability graph must be
pInaB
considered. If the reachability graph is strongly con-
tPoetB1 nected that is, if the initial marking is a home state,
pProcessB1 then the model is actually live. The properties of
liveness and home state are a sufficient condition for
tPreB1
the GSPN model and its underlying continuous time
\
Markov chain (CTMC) to be ergodic.
In addition, the construction of the CTMC is
based on the assumption that the GSPN model
Figure 5
belongs to the class of structurally confusion-free
Petri Net Model of a Simple Tandem AGV System nets. A confusion-free net is such that the firing of
simultaneously enabled transitions is immaterial for
the evolution of the net. Models comprising confu-
The loads leave the loop through the last processing sion are usually considered to be semantically
workstation. Product B works under a CONWIP wrong. A confusion-free net is generally achieved by
operating policy. The WIP level is represented by the assigning higher priority levels to transitions that
parameter K. Notice that, in this single-product cannot fire concurrently even if they are not in con-
example, place pWIP is not introduced to avoid flict (Chiola et al. 1993).
redundancy with place pOutQB5. That is, the CON- In the case of the simulative alternative, the
WIP operating policy is simply modeled by con- GSPN model must also satisfy the above properties,
necting transition tPreB 1 to place pOutQB5. except that it is not required to be bounded.
The Petri net modeling approach for tandem AGV
Model Validation Considerations systems presented in this paper has been conceived
The state space of a Petri net model is called the taking proper care of the aforementioned require-
reachability set and is characterized by the initial ments. The net in Figure 5 defines a reachability set
marking and the structure of the net. The reachabil- comprising 642 markings, 448 of which are tangible
ity set is defined as the set comprising all markings markings and 194 of which are vanishing markings.
that can be reached from the initial marking by fir- Each place in the GSPN model is covered by some
ing any legal transition sequence. This set is com- P-invariant. The computation of T-invariants reveals
posed of tangible and vanishing markings. Markings that all the transitions in the model are covered by T-
that enable timed transitions are called tangible, invariants. The initial marking is a home state; the
while markings that enable immediate transitions net is live and thus, ergodic. The net, moreover, is
are called vanishing. The reachability graph of a confusion free.

243
Journal of’A4amfircturing Systems
Vol. 2oiNo. 4
200 1

The proposed modeling approach for tandem To show the kind of performance metrics that can
AGV systems allows the consideration of several be obtained using a GSPN model, the tandem AGV
detailed design issues, as discussed in previous sec- model in Figure 5 is considered. Say that the expect-
tions. In certain cir~~stan~es where this detail is ed utilization of the vehicle, the expected utilization
not necessary, a considerably simplified net can be of the workstations, and the number of parts that are
developed. For example, if the load/unload rates are processed per time unit are required. These parame-
not important and can be assumed to be zero, such ters can be derived from the steady-state probability
as when comparing different systems rather than distribution on markings.
analyzing a single system, transitions tLoad and Place pWS4 is l-bounded, and when there is no
tUnload of the product subnets become 2-immediate token in place pWS4, the workstation is busy.
instead of timed. The simplification, moreover, Therefore, the expected utilization of workstation
reduces the modeling overhead by not considering WS4 is given by 1 - ApWS4, where ApWS4is the
any of the i~ibitors arcs from product subnets to expected number of tokens in place pWS4. The
loop subnets, nor the transitions tLogic in the move same holds for the AGV, modeled by place pAGV
modules. This modeling simplification preserves the The expected number of products of type B
properties previously mentioned and helps reduce processed per time unit is given by the inverse of the
the number of tangible markings by a considerable frequency of firing of transition tPreB1. The fre-
amount. For example, with this simplification, the quency of firing of tPreB1 is
net in Figure 5 would have 308 tangible markings
instead of 448. f tPreB1 =
M:M[tPreBI>

Solution and Performance Analysis where &real is the firing rate of transition tPreB1
of GSPNs andprM is the steady-state probability of marking Ad.
Because of the memoryless property of the expo- The summation is carried over all the markings
nential dist~bution, a GSPN model is recognized as where ~ansition tPreB1 is fired. Alternatively, the
a CTMC defined over a set of states isomorphic to number of products of type B processed per time
the set of tangible markings of the net. The analytic unit can be computed as
analysis alternative solves this underlying CTMC.
The analytic solution allows the computation of t&z= %PreBl
P {#pProcessB 1> 0]
exact solutions; that is, as opposed to the simulative
alternative, the outcome does not depend on the gen- that is, the probability that place pProcessB1 is not
eration of pseudo-random numbers. The analytical empty (P{ #pProcessBl > O>), times the firing rate
solution of a GSPN, however, is limited by the state- of transition tPreB 1, which represents the number of
space explosion problem. Even if the GSPN is com- products processed per time unit when processing is
pact, the underlying CTMC can easily have an enor- being performed. In addition, because place
mous set of states. Thus, even if the state space is pOutQB4 is K-bounded, the expected number of
finite, it cannot be too large relative to the main products in this place can be computed as
memory.
The analytical alternative assumes that the GSPN A pOutQB4 = biP (#pout QB4 = i)
and the underlying CTMC are ergodic. A sufficient
condition for a CTMC to be ergodic is that the net is On the other hand, the simulative analysis alter-
live and that the initial marking is a home state. native does not require the generation of the state
Moreover, the construction of the CTMC is based on space of the Petri net and therefore, does not require
the assumption that the GSPN is confusion free a finite state space. This alternative allows the com-
(Chiola et al. 1993), as previously discussed. The putation of point estimates and their respective con-
steady-state solution of the CTMC can be obtained fidence intervals. Moreover, the underlying stochas-
by solving the ~o~esponding linear system of dif- tic process is not restricted to be M~kovian; that is,
ferential equations using an iterative method based transitions can have generally distributed firing
on the Gauss-Seidel algorithm. times. This is important because with no congestion

244
Journal of Manufacturing Systems
Vol. 2olNo. 4
2001

Table 4
Workstations as interface stations AGV Travel Distances Between Stations
n
I I I Loop I
..______.
P/D 1 P/D 4
4
From-To Distance (feet)
1
7Eix 90
WS4 - WS5 135
WS5 - WS3 225
ws3-wS2 405
ws2 - WSl 225
- P/D 2 LOODII
From-To Distance (feet)
ws3 - ws5 225
WS5 - WS6 540
WS6 - WS3 315

c Table 5
; Routing Data by Product Type
Product Type Production Routing Loop
Figure 6
Experimental Tandem AGV System A WSl - WS3-- WS6 Loop I - Loop II
B ws4 - WSl Loop I
c WS6 - WS5 - WS2 Loop II - Loop I
D WLQ-wSbws5 Loop I
and blocking, AGV travel times in a tandem AGV
system are unlikely to follow the exponential distri-
Table 6
bution. Complicated models, as in discrete-event Exuerimental Set of Svstem Parameters
simulation, demand long simulation times to Case WIP Level per AGVs Capacity
achieve the desired accuracy. Product Type
The simulation of a Petri net is performed by exe- 1 1 1
2 2 1
cuting the following cycle:
3 3 1
4 4 1
l Find enabled transitions. 5 5 1
l If immediate transitions are enabled, choose one 6 6 1
7 1 2
transition to fire according to priorities and 8 2 2
weights. If timed transitions are enabled, choose 9 3 2
one transition according to firing rates. 10 4 2
11 5 2
l Fire the selected transition and compute the new 12 6 2
marking.
l Update statistics of estimates.
WS3 serve as transfer points between loops.
For both the analytic and the simulative analysis Workstation WSl has two parallel and identical
alternatives, the GreatSPN package developed at the machines, whereas the other workstations have only
Universita di Torino is used (Chiola et al. 1995). The one machine each. The AGVs travel distances
package provides a friendly framework to experi- between stations are shown in Table 4. The loaded and
ment with GSPN modeling techniques and state-of- unloaded speeds of the AGVs are assumed to be equal
the-art analysis algorithms for validation, to 125 feet per minute. The AGVs travel, in both
Markovian analysis, and simulation of models. The loops, in a clockwise direction. The system produces
output of the sofhvare tool is used to compute the four types of products, denoted A, B, C, and D. The
desired performance metrics. four product types are produced according to the data
given in Table 5. The AGVs are capable of loading
and unloading loads from and to P/D points at a rate
Experimentation and Results of 360 unit loads per hour. Also, the system operates
Consider a manufacturing system with two loops under a CONWIP operating policy; that is, the WIP
and six workstations as shown in Figure 6. Notice that level for each product is held constant. The experi-
the P/D points of workstation WS5 and workstation mental set of system parameters is shown in Table 6.

245
Journal ofManuj&turing Systems
Vol. 20iNo. 4
2001

In Table 7, the results for the single-load AGVs ther throughput improvement.
are presented while the results for the dual-load The behavior of the AGV utilization is depicted in
AGVs are presented in Table 8. Three important Figures 9 and 10 for each of the two AGV capacity
results are shown in these tables: the throughput scenarios, respectively. The AGV capacity has, as
per hour for each product type, the AGV utiliza- expected a positive effect on the throughput metric.
tion, and the workstation utilization. The results for In the same fashion, the AGV capacity has a positive
cases 1, 2, 7, and 8 were computed using the ana- effect on the utilization of the workstations. This can
lytical analysis alternative. For the rest of the cases, be visualized in Figures II and 12 for each of the
the simulative analysis alternative was used and a two AGV capacity scenarios, respectively. In these
95% confidence interval for each performance figures, the shape of the curves is nearly identical.
metric was also computed. The batch means Note that in the experimentation above, the AGV
approach was used when conducting the simula- utilization is rather low with respect to the worksta-
tions. The number of batches per simulation was tion utilization. Thus, increasing the WIP level from
variable; for each case, statistics were collected 1 to 2 units of each product type has a much larger
until the accuracy of the confidence interval for all impact on product throughput than converting the
performance metrics was within 10%. AGVs from single-load to dual-load vehicles. Now,
The throughput per hour for each product type let the loaded and unloaded speeds of the AGVs be
increases as the WIP level for the products increas- equal to 10 feet per minute and consider the results
es. This behavior is depicted in Figures 7 and 8 for shown in Table 9.
each of the two AGV capacity scenarios, respective- As expected, with reduced loaded and unloaded
ly. The curves for all the product types appear to speeds, the AGV utilization is rather high with
reach the point of critical WIP, that is, the point respect to the workstation utilization. In Table 9, it is
where further WIP increment will not produce fur- seen that converting the AGVs from single-load to

Table 7 Table 8
Experimental Set Results Experimental Set Results
Case Throughput AGV Utilization Workstation Utilization Case Throughput AGV Utilization Workstatmn Utilization
per Hour Loop I Loop II per Hour Loop I Loop II

I / A 0.7042 0.1790 1 ~~~ ~


O.lZS I mmPEl 05418 7 A -0.7257 0.1696 ~ 0.1235 I WSI 0.5486
5 0.4433 , WS2 0.4274 B - 0.4486 W’S2 0.4302
C-O.5682 ~ WS3 0.4641 C - 0.5676 WS3 0.4697
D -- 0.2853 WS4 0.4433 D - 0.2883 WS4 0.4486
I WS5 0.4274 W’S5 0.4302
WS6GO.3181 I W’S6 0.3233
2 A - 1.0000 0.2634 0.1738 WSI 0.6800 8 A - 1.0260 0.2343 0. I577 WSI 0.6903
B 0.5753 WS2 0.6177 B - 0.5830 WS2 -0.6217
c - 0.7335 WS3 0.6845 C 0.7520 ~ WS3 0.695 I
D 0.4340 WS4 - 0.7246 D-O.4389 ws4 0.7379
WSS 0.6176 WS’5 0.6230
i WS6 -0.4378
3 (A - I .l I52 + 0.0247 0.3029 i 0.003 0.1944 i 0.0020 WSI 0.7359 i 0.0107 9 ‘A 1.1700 zt 0.0305 0.2607 zt 0.0016 0.1709 + 0.0020 WSI 0.7410 i 0.0065
!B-0.6349iO.01611 WS2 - 0.7120 i- 0.0217 5 0.6337 i 0.0072 WS2 0.7241 f 0.0149
W’S3 0.8014 i 0.0080 C - 0.7896 i 0.0293 WS3 0.8067 zk 0.0227
‘D - 0.5200 + 0.0098
~ WS4 - 0.8570 + 0.0552 D 0.5277 i 0.0063, WS4 0.8694 l 0.0232

WS5 0.72 I8 t 0.0283 WS5 0.7254 + 0.0087


~~ ~~~~_~ WS6 0.4784 * 0.0081 ~ fS6 -@X75 +m&O087
4
4 A - 1.1800 i 0.0416 0.3246 + 0.0029/ 0.2053 + 0.0038tWSI - 0.7672 f 0.0102 IO A - 1.2088 + 0.0254 0.2760 z+ 0.0025 0.1762 i 0.0020 WSI - 0.7650 + 0.0042
-I
B - 0.6572 * 0.01 I8 WS2 0.7803 f 0.0325 B 0.6601 i 0.0066 _~-r WS2 0.7792 zt 0.0277
C - 0.8264 -t 0.0197 WS3 - 0.8649 zt 0.0581 C 0.8344 + 0.0210 1WS3 0.8728 i 0.0472
~D 0.575710.0088 WS4 0.9251 i 0.0581 II 0.5710 + 0.0032
I
’ ws4 0.9337 * 0.0437
WS5 0.7814 i 0.0284 WS5 0.7796 i 0.007 I
WS6 0.5017 f 0.0065 WS6 - 0.5 123 f 0.0058
5 A - I .I884 i 0.0243 IO.33 I5 * 0.001 0.2056 * 0.0030 WSI 0.7729 * 0.0066 II A - I.2316 zt 0.0139 0.2777 i 0.0020’ 0.1769 + 0.0030 WSI 0.7688 + 0.0102
B - 0.6730 + 0.0056 ~~~~~ (I IWS2 0.8225
~~~~ * 0.0412
~~~ E 0.6669 i 0.0112 WS2 0.8 I33 + 0.065 I
WS3 0.9126 i- 0.0340 C 0.8316 i 0.0235 WS3 0.9 I62 + 0.0409
,:;“,:z::::g WS4 - 0.9615 + 0.067 I D - 0.6170 + 0.0077 ~ WS4 0.9667 i- 0.0629
I
WS5 0.8140 + 0.0542 , WS5 0.8183 f 0.0680
WS6 0.5 149 i 0.0092
_~ ~~_~ _PPIm~ ~ W’S6 0.5045
~~~~ i _~~
0.0071
6 ‘km- I.1884 * 0.0513 IO.3370 + 0.0032. 0.2093 * 0.00331 WSI 0.7689 + 0.0098 I2 A - 1.1932 * 0.0326 0.2842 + 0.1761 l 0.0016 1WSI 0.7734 * 0.0033
B - 0.6719 l 0.0094 WS2 0.8455 * 0.0698 B 0.6761 i 0.0052 ws2 0.855 I i 0.0440
C- 0.8580
D - 0.6287
+ 0.0344
+ 0.0079 iE I ::;:;:: : ::::;: C -0.8704
D - 0.6278
i 0.0243
+ 0.0043
WS3
WS4
0.9452
0.9X21
f 0.0736
f 0.0855
KS5 0.8409 i 0.0544 WS5 0.8434 IO.0277
WS6 0.5057 -t 0.0135 WS6 0.5 165 i 0.0082

246
Journal of Manufacturing Systems
Vol. 2a/No. 4
2001

1.2 1.2

1 1

Cl ;
e 0.6
$
5
$ O6
E
+ 0.4
Product A --+-- Product A ---+-
0.2 Product B ---+--- 0.2 Producti? ---+ -x
‘,..“,’ ,,
Pmduct C ___J+__ .,>,‘,, Product C --o--
1,,, Product D ---x---
Product D ___)(___ ,’
0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
WIP level per product type WIP level per product type

Figure 7 Figure 8
Throughput Curves for Single-Load AGVs Throughput Curves for Dual-Load AGVs

: I
,/’
0.05 :I-
:‘,, AGVl --a--
; ,’
,;i,’ AGVZ---+--
06
0 1 2 3 4 5 6
WIP level per product type WIP level per product type

Figure 9 Figure 10
Single-Load AGV Utilization Dual-Load AGV Utilization

I-
f

0.6 - -+____ __ 0.6

0.6 -

i 8 _,A’
$,‘.* WSI
w.9
--u--
---+_-_
I wsl
m
m
--Q__
--_+_--
___D__
WSQ ___x_-_
m ___*_--
m __.*._-

"0 1 2 3 4 5 6 “0 1 2 3 4 5 6
WIP level per product type WIP level per product type
Figure 11 Figure 12
Workstation Utilization for Single-Load AGVs Workstation Utilization for Dual-Load AGVs

247
Journal of Manufacturing s?/stems
Vol. 20Mo. 4
2001

Table 9
Additional Experimental Set Results simple and their logic intuitive. Moreover, combin-
Loop I I ws2 Throughput
ing the modules is almost mechanical. Thus, the
approach generalizes quite well and in contrast to
discrete event simulation, in which computer pro-
gramming is required, model verification and vali-
dation is simplified.
0.5844 0.1470 The modeling assumption, in the case of the ana-
0.5800 0.1878 22.77% lytical analysis alternative, is the use of exponential-
0.6543 0.2062
0.7438 0.2897 1 17.31%
ly distributed interarrival and service times. In the
I case of the simulative analysis alternative, no such
restriction is required. Thus, general distributions
dual-load vehicles has a much larger impact on for both interarrival and service times can be speci-
product throughput when the material handling sys- fied. This is important because, with no congestion
tem is the bottleneck of the system. Thus, an inter- and blocking, AGV travel times in a tandem AGV
action exists between the AGV utilization and the system are unlikely to follow the exponential distri-
workstation utilization. This interaction needs to be bution.
taken into account when designing tandem AGV When exact, yet conservative, solutions are
systems. The study of this interaction in detail is desired, the flexibility provided by the Petri net mod-
beyond the scope of this paper. eling approach has a drawback; the numerical solu-
tion of a generalized stochastic Petri net is limited by
the state-space explosion problem. Extensive com-
Summary and Conclusions puting power and resources are the only way to over-
The advantages of AGV systems relative to other come this obstacle. The simulative alternative, on
types of material handling systems include: reliable the other hand, allows point estimates to be obtained
and automatic operation, improved positioning when memory resources are exceeded with the ana-
accuracy, rapid and flexible response to changes in lytical approach. Without robust modeling, however,
the material handling requirements, and easily any solution technique or numerical analysis is use-
expandable layout and system capacity, among oth- less. The modeling approach for tandem AGV sys-
ers. The design and further performance analysis of tems presented in this paper represents a step toward
AGV systems, however, can be quite complicated. In such a robust modeling framework.
general, discrete event simulation is used as the pri-
mary tool in analyzing AGV systems; analytical References
Ajmone Marsan, M.; Balbo, G.; Conte, G.; Donatelli, S.; and
tools are preferred when certain modeling assump-
Franceschinis, G. (1995). Modeling with Generalized Stochastic Petri
tions and constraints hold. Nets (New York: John Wiley & Sons).
In this paper, a Petri net approach for analyzing Alanche, D.; Benzakour, K.; Gilld, D.F.; Dolle, E; Gill& P; Rodriguel, P;
and Vallete, R. (1984). “PSI: a Petri net based simulator for flexible
tandem AGV systems has been presented. The pro- manufacturing systems.” Advances in Petri Nets, Lecture Notes m
posed approach provides, within a single modeling Computer Science 188 (New York: Springer-Verlag).
framework, analytical as well as simulative results Bartholdi, J.J. and Platzman, L.K. (1989). “Decentralized control of auto-
mated guided vehicles on a simple loop.” IZE Trans. (v2 I, n 1), ~~76-8 I.
with much of the modeling flexibility of discrete Bilge, U. and Tanchoco, J.M.A. (1997). “AGV systems with multi-load car-
event simulation. The modeling flexibility of the riers: basic issues and potential benefits.” Journal ofMom@cturing
proposed approach allows the consideration of criti- Systems (~16, n3), ~~159-174.
Bozer, Y. and Park, J.H. (I 992). “New partitioning schemes for tandem
cal design issues that have not been addressed previ- AGV systems.” Proc. of Int’l Material Handling Research Colloquium.
ously in the literature for the class of tandem AGV Bozer, Y. and Srinivasan, M.M. (1991). “Tandem configurations for auto-
mated guided vehicle systems and the analysis of single vehicle loops.”
systems. These critical design issues require the rep-
IIE Trans. (~23, nl). ~~72-82.
resentation of concurrency, synchronization, and Bozer, Y. and Srinivasan, M.M. (1992). “Tandem AGV systems: a parti-
resource sharing; properties that cannot be tioning algorithm and performance comparison with conventional AGV
systems.” European Journal qf’Operational Reseurch (~63, n2). ppl73-
expressed easily in queuing theoretical formalisms. 191.
In the proposed approach, tandem AGV Petri net Chiola, G.; Ajmone Marsan, M.; Balbo, G.; and Conte, G. (1993).
models are built by combining the three basic mod- “Generalized stochastic Petri nets: a definition at the net level and its
implications.” IEEE Trcms. on S#ware Engg. (~19. n2), ~~89-107.
ules, move, transfer, and process. The modules are Chiola, G.; Franceschinis, G.; Gaeta, R.; and Ribaudo. M. (1995).

248
Journal of Manufacturing Systems
Vol. 2a/No. 4
2001

“GreatSPN 1.7: graphical editor and analyzer for timed and stochastic Raju, R.K. and Chetty, O.VK. (1993). “Design and evaluation of automat-
Petri nets.” Performance Evaluation (v24), ~~47-68. ed guided vehicle system for flexible manufacturing systems: an
David R. and Alla, H. (1992). Petri Nets & Grafcet: Tools for Modeling extended timed Petri net-based approach.” Int ‘l Journal of Production
Discrete Event Systems (New York: Prentice-Hall). Research (v3 1, n5), pp1069-1096.
Davis, W.J.; Jackson, R.H.F.; and Jones, A.T. (1989). “Real time optimiza- Schroer, B.J. and Tseng, ET. (1988). “Modeling complex manufacturing
tion in the automated manufacturing research facility.” Progress in systems using discrete event simulation.” Computers and Industrial
Material Handling and Logistics (New York: Springer-Verlag). Engg. (~14, n4), ~~455-464.
Egbelu, PJ. (1987). “The use of non-simulation approaches in estimating Venkatesh, K.; Zhou, M.-C.; Kaighobadi, M.; and Caudill, R. (1996). “A
vehicle requirements in an automated guided vehicle based transport Petri net approach to investigating push and pull paradigms in flexible
system.” Material Flow (v4), ~~17-32. factory automated systems.” Int ? Journal of Production Research (~34,
Egbelu, PJ. and Tanchoco, J.M.A. (1984).“Characterization of automatic n3), ~~595-620.
guided vehicle dispatching rules.” Int’l Journal ofproduction Research Vosniakos, G.C. and Mamalis, A.G. (1990). “Automated guided vehicle
(~22, n3), pp359-374. system design for FMS applications.” Int ? Journal of Machine Tools &
Egbelu, PJ. and Tanchoco, J.M.A. (1986). “Potentials for bi-directional Manufacture (v30), pp85-97.
guidepath for automated guided vehicle based systems.” Int’l Journal of Wang, H.-PB. and Hafeez, S.A. (1994). “‘Per$ormance evaluation of tandem
Production Research (~24, n5), pp1075-1097. and conventionalAGVsystems using generalized stochastic Petri nets.”
Hsieh, L.-F. and Sha, D.Y. (1996). “A design process for tandem automat- Int’l Journal of Production Research (~32, n4), ~~917-932.
ed guided vehicle systems: the concurrent design of machine layout and Yim, D. and Linn, R.J. (1993). “Push and pull rules for dispatching auto-
guided vehicle routes in tandem automated guided vehicle systems.” mated guided vehicles in a flexible manufacturing system.” Int’l
Integmted Mfg. Systems (~7, n6), ~~30-38. Journal of Production Research (~31, nl), ~~43-57.
Hsieh, L.-F. and Sha, D.Y. (1997). “Heuristic algorithm for the design of
facilities layout and AGV routes in tandem AGV systems.” Int ‘1Journal
of Industrial Engg. (~4, nl), ~~52-61.
Lin, J.T. and Dgen, P-K. (1994). “An algorithm for routing control of a tan- Authors’ Biographies
dem automated guided vehicle system.” Int’l Journal af Production Ignacio Castillo is an assistant professor of management science at the
Research (~32, n12), ~~2735-2750. University of Alberta School of Business. He received a PhD in industrial
Mahadevan, B. and Narendran, T.T. (1990). “Design of an automated guid- engineering from Texas A&M University in 2001, an MSE in industrial
ed vehicle based material handling system.” Int ‘IJournal of Production engineering from Arizona State University in 1998, and a BS (magna cum
Research (~28, n9), ~~1611-1622. laude) in applied sciences from Universidad San Francisco de Quito,
Malmborg, C.J. (1990). “A model for the design of zone control automated Ecuador, in 1995. His research and teaching interests include manufactur-
vehicle systems.” Int’l Journal of Production Research (v28), pp1741- ing logistics, facility layout and material handling systems, and operations
1758. research applications in production systems. Dr. Castillo is a member of
Martinez, J. and Silva, M. (I 98 1). “A simple and fast algorithm to obtain all HE, INFORMS, Alpha Pi Mu, Phi Kappa Phi, and a former FUNDA-
invariants of a generalized Petri net.” Proc. of 2nd European Workshop CYT/LASPAU scholar.
on Application and Theory of Petri Nets.
Maxwell, W.L. and Muckstadt, J.A. (1982). “Design of automated guided Sergio A. Reyes received an MS in industrial engineering from Texas
vehicle systems.” HE Trans. (~14. n2), ~~114-124. A&M University in 2000 and a BS in mechanical and electrical engineer-
Murata, T. (1989). “Petri nets: properties, analysis and applications.” Proc. ing from the Institute Tecnologico y de Estudios Superiores de Monterrey,
of IEEE (~77, n4), ~~541-580. Mexico, in 1996. His research interests include the real-time scheduling,
Nayyar, P and Khator, S.K. (1993). “Operational control of multi-load analysis, and control of manufacturing and material handling systems. Mr.
vehicles in an automated guided vehicle system.” Computers and Reyes is a member of HE, INFORMS, Alpha Pi Mu, and a former CONA-
Industrial Engg. (~25, nl-4), ~~503-506. CYTiFulbright scholar.
Ozden, M. (1988). “A simulation study of multiple-load-carrying automat-
ed guided vehicles in a flexible manufacturing system.” Int’l Journal of Brett A. Peters is an associate professor in the Industrial Engineering
Production Research (~26, n8), ~~1353-1366. Dept. at Texas A&M University. He received a PhD and MS in industrial
Peters, B.A.; Smith, J.S.; and Venkatesh, S. (1996). “A control classification engineering from the Georgia Institute of Technology in 1992 and 1988,
of automated guided vehicle systems.” Int ‘IJournal af Industrial Engg. respectively, and a BSIE from the University of Arkansas in 1987. His
(~30, nl), ~~29-39. research and teaching interests include facilities design, material handling
Peterson, J.L. (1981). Petri Net Theory and the Modeling of Systems systems, and the design, analysis, operation, and control of manufacturing
(Englewood Cliffs, NJ: Prentice-Hall). systems. Dr. Peters is a member of SME, HE, and INFORMS.

249

You might also like