Intelligent Agent Based Framework for Manufacturing Systems Control

bootlessbwakInternet and Web Development

Nov 12, 2013 (3 years and 8 months ago)

159 views

560 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
Intelligent Agent Based Framework for
Manufacturing Systems Control
Sunderesh S.Heragu,Robert J.Graves,Byung-In Kim,and Art St.Onge
Abstract Existing modeling frameworks for manufacturing
system control can be classified into hierarchical,heterarchical,
and hybrid control frameworks.The main drawbacks of existing
frameworks are discussed in this paper.A new hybrid modeling
framework is also described.It is a hybrid of the two:hierarchical
and heterarchical frameworks.In this proposed framework,
entities (e.g.,parts) and resources (e.g.,material handling devices,
machines,cells,departments) are modeled as holonic structures
that use intelligent agents to function in a cooperative manner so as
to accomplish individual,as well as cell-wide and system-wide ob-
jectives.To overcome the structural rigidity and lack of flexibility,
negotiation mechanisms for real-time task allocation are used.
Lower-level holons may autonomously make their negotiations
within the boundary conditions that the higher-level holons set.
Horizontal,as well as vertical decisions,are made between various
levels of controllers,and these are explicitly captured in the model.
Index Terms Heterarchical,hierarchical,holon,hybrid,intelli-
gent agent,manufacturing systems control.
I.I
NTRODUCTION
T
HROUGHa workshop and a delphi survey,the committee
on Visionary Manufacturing Challenges for 2020 has iden-
tified intelligent agents as one of the key enabling technolo-
gies that will help companies overcome one of the six grand
challenges or fundamental objectives to remain productive and
profitable in the year 2020 [1].This grand challenge is to in-
stantaneously transform information from a vast array of di-
verse sources into useful knowledge and effective decisions
[1].Intelligent agents are mentioned as an enabling technology
that will help in modeling and decision support systems.In this
paper,we discuss how a new modeling framework,involving
holons and intelligent agents negotiating between themselves,
can be used to develop a manufacturing control system.
The theory of holonic systems has evolved fromseveral early
roots,with one in the biological systems arena,where obser-
vations of stability and coherence were made about biological
subsystems.Koestler [2] observed that subsystems could be co-
herent and stable while interacting in a larger hierarchical sub-
system.He coined the term holon to describe nodes in the
Manuscript received September 23,2000;revised July 14,2001.This work
was supported by National Science Foundation Grant DMI 9900039.This paper
was recommended by Associate Editor S.Narahari.
S.S.Heragu and R.J.Graves are with the Department of Decision Sciences
and Engineering Systems,Rensselaer Polytechnic Institute,Troy,NY 12180,
USA (e-mail:herags@rpi.edu;graver@rpi.edu).
B.-I.Kimis with the University of Memphis,Memphis,TN38152,Industrial
and Systems Engineering,(email:bkim@memphis.edu).
A.St.Onge is with St.Onge Company,York,PA 17402,USA (e-mail:
srtstonge@stonge.com).
Digital Object Identifier 10.1109/TSMCA.2002.804788
hierarchical tree that exhibit whole and part characteristics.De-
pending upon how one looks at them,holons behave partly as
wholes and wholly as parts.According to Koestler [2],holons
have dual tendencies to preserve and assert their individuality
while,at the same time,function as an integrated part of a larger
unit.
Holonic manufacturing systems include individual units
(part,material handling device,or machine controller) that
make autonomous decisions based on guidelines,or system
wide constraints provided by higher-level controllers,such
as,a cell controller or system controller.The holonic concept
clearly recognizes the two levels of interactions that take place
among unit and higher level controllers.
In the past,two major frameworks hierarchical and heter-
archical frameworks,have been proposed to model the inter-
actions among controllers in a manufacturing system.In this
paper,a new modeling framework that has both features of hi-
erarchical and heterarchical frameworks is presented.It recog-
nizes that horizontal,as well as vertical decisions,are made
between various levels of controllers,and that these have to
be captured explicitly in the model.The focus of the holonic
modeling framework is on the study of design and planning in
the manufacturing domain,while still recognizing the fact that
the material handling system (MHS) is an integral part of the
broader manufacturing system.For example,processing time
variability impacts the transportation schedule of the material
handling device that serves the machine and must be considered
during the negotiation process between machine and material
handling agents.Entities (e.g,parts) and resources (e.g.,mate-
rial handling devices as well as machines) are modeled as holons
using intelligent agents to function in a cooperative manner so as
to accomplish individual as well as cell-wide and system-wide
objectives.Scheduling,and other decisions,are taken by these
holons in a dynamic and real-time fashion,based on conditions
that exist at the time these decisions are made.Such time-based
decisions and material flow control are a driving force in man-
ufacturing.The framework can be used to study the impact of
design decisions,operational decisions as well as the impact of
design decisions on operational aspects.For example,it can be
used to determine the effect on traffic congestion of introducing
a specific number of automated guidedvehicles (AGVs) in a net-
work and on the input and output queues of a key work-center
served by the AGVs.Or it can be used to study the impact of
reducing a key work-centers processing time variability on the
transportation schedule of a subset of material handling devices.
Research addressed in this paper has received significant at-
tention and interest from industry,especially in the recent past.
There is a relatively large body of literature demonstrating that
1083-4427/02$17.00 © 2002 IEEE
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 561
holonic ideas have been applied successfully to resource al-
location problems in distributed computing environments.Re-
searchers have used the intelligent agent framework for mod-
eling operational problems in general manufacturing systems
(see [3],for example).While there are limited efforts at mod-
eling the overall systems design problem,many of them do
not explicitly capture the interactions of the material handling
system with the manufacturing system.It is important to con-
sider these interactions because material handling systemdesign
problems have an integrative influence on material flow over a
manufacturing shop-floor.
This paper is organized as follows.After we classify the ex-
isting control architectures as hierarchical,heterarchical,and
hybrid models in Section II,we summarize the drawbacks of
hierarchical and heterarchical models and discuss the features
of hybrid models in Section III.Our new modeling framework
is presented in Section IV and a conceptual framework is pre-
sented.Application of this framework to an industrial problem
can be found in Kim et al.[4],[5].Distinctive features of our
framework and conclusion are presented in Section V.
II.L
ITERATURE
R
EVIEW
There are several classifications of control frameworks.
Duffie and Piper [6] present a spectrum of architectures of a
centralized controller,a hierarchical controller with dynamic
scheduling,and a fully distributed heterarchical controller
with intelligent parts.Lin and Solberg [7] present four con-
trol paradigms:centralized information-centralized decision
making,distributed information-centralized decision making,
centralized information-distributed decision making,and
distributed information-distributed decision making.Dilts et al.
[8] identify four basic control architecture forms for automated
manufacturing systems:centralized,proper hierarchical,modi-
fied hierarchical,and heterarchical.They also summarize their
characteristics,advantages,and disadvantages.
In this paper,we classify control frameworks into hier-
archical,heterarchical,and hybrid control frameworks.The
hierarchical frameworks map to the centralized and proper
hierarchical classifications in [8].The hybrid frameworks
include the modified hierarchical architecture in [8] and their
recent evolutions.The hierarchical framework assumes there is
a hierarchy and a master/slave relationship between higher and
lower levels of control.The hierarchy is introduced to handle
the complexity of a manufacturing system.Sensory data flows
in an upward direction from units at the lowest level to higher
level supervisory controllers [9].Based on this,command data
is generated and sent in a downward direction from supervi-
sory controllers to units.Optimization and knowledge-based
methods have been extensively applied in these traditional
approaches at the subsystem optimization level rather success-
fully.However,because they ignore important aspects,such
as,uncertainty and complexity of the real-world system,they
have not necessarily been adequate in accurately modeling,
or providing timely prescriptions in todays distributed and
complex manufacturing environment.The slow response times
for these sensory and command information flows adversely
affect the quality and timeliness of decisions in manufacturing
since the environment at the time a control decision is executed
is different fromthe one under which the decision was made.In
addition,they ignore,almost completely,important interactions
between unit controllers.Decisions are therfore made entirely
by the master controller [8].
The heterarchical framework focuses on interactions be-
tween unit controllers to allowsystemflexibility,while ignoring
those between higher and lower-level controllers.The lack of
predictability and global perspective are major drawbacks of
this framework.The hybrid framework has features of both
hierarchical and heterarchical frameworks.It allows direct
interactions amongst the lower-level controllers as well,as
those between those higher and lower.
Although intelligent agents and holons have different mean-
ings and roots (see [10]),they are used interchangeably in the
literature.While discussing existing papers in this section,we
keep the terminology used by the respective authors.
A.Hierarchical Control
The ubiquitous presence of hierarchy motivated researchers
and practitioners to build hierarchical manufacturing control
systems.In this subsection,past research attempts at building
hierarchical systems are briefly reviewed.Jones and McLean
[11] and Jackson and Jones [12] present a hierarchical control
model developed for automated manufacturing systems.To
limit the size,complexity,and functionality of individual con-
trol modules,they adapted the hierarchical structure.There are
five layers in their model:facility,shop,cell,workstation,and
equipment layers.While longer-term planning is theoretically
possible in the higher layers,only the lower three layers are
implemented in their prototype model.Each module decom-
poses the input command from its supervisor into simpler
subtasks,assigns them to appropriate subordinates,monitors
their executions,and provides status feedback to the supervisor.
Each module has one supervisor and several subordinates,and
there is no direct communication between modules of the same
level.While they propose a hierarchical modeling methodology
for real-time scheduling,its feasibility and optimality are not
proven.
Chryssolouris et al.[13],[14] present a work center level
controller,called manufacturing decision making (MADEMA),
which assumes four levels of hierarchy:factory,job shop,work
center,and resource.The factory level represents the entire fac-
tory and controls the plant-wide capacity requirement.The job
shop level consists of work centers and assigns work to them.
The work center consists of groups of production resources.The
resource level refers to a single production resource.MADEMA
gets the work requirement (type,quantity,release times,and
due dates of jobs) from the job shop-level,determines feasible
alternatives of task-resource pairs,relevant criteria,the conse-
quences of the alternatives with the multiple criteria,applies
decision-making rules,and then selects the best alternative.
Boulet et al.[15] compare three hierarchical control architec-
turesthe ones developed by Jackson and Jones [12],European
Strategic Program for Research and Development in Informa-
tion Technology (ESPRIT Project 932) and Chryssolouris et al.
[14].OGrady and Lee [16] propose a multiblackboard archi-
tecture for the intelligent cell controller,which is a middle-
562 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
level hierarchy of factory,shop,cell and equipment.There are
scheduling,operation dispatching,monitoring and error han-
dling blackboards in their architecture.Cho and Wysk [17],[18]
develop a three-level hierarchical shop floor control system.The
levels in the system correspond to the shop,workstation and
equipment levels in the hierarchy.Their design focuses on the
middle-level workstation controller.
B.Heterarchical Control
Intelligent agent frameworks have been used extensively to
determine resource allocation in distributed computing environ-
ments.The White House has used one to handle thousands of
requests for information [19].In an address to the 21st Confer-
ence on Decision and Control,Vamos [20] discussed the inade-
quacy of traditional control approaches and defined the cooper-
ative systemas:1) a free coalition of systems,2) a systemwhere
complete knowledge on the systemis not available,3) a system
that operates by exchange of information.It has no central con-
troller.The control is distributed and based on the agreements of
the systems components.Thus,it can handle unplanned events
such as machine breakdown.Hatvany [21] suggests cooperative
heterarchies as an alternative to the hierarchical control system.
He points out the need for designing behavior rules,local ob-
jectives,and global objectives (which the autonomous entities
follow),to prevent anarchy.
Duffie and Piper [6],[9] demonstrate the advantages of the
heterarchical control architecture by comparing three control
systems:centralized,hierarchical,and heterarchical.The advan-
tages of the latter include reduced complexity,reduced software
development costs,high modularity,high flexibility,and im-
proved fault tolerance.But they also list a few of its problems
including the contradiction problembetween local objective and
the overall system performance as well as deadlock detection
and resolution problems.
Duffie and Prabhu [22] present a look-ahead cooperative
scheduling algorithmto enhance the global systemperformance
of the heterarchical manufacturing system,which was pointed
out as a problemby Duffie and Piper [6].Because they observe
unstable dynamics in their experiments,Duffie and Prabhu [23]
develop a local feedback algorithm for controlling the arrival
of parts at a machine.They also mention that a primary reason
why the architectures have not been used in real-world manu-
facturing is due to the perception that there is a high degree of
variability in the performance of a heterarchical system.They
list several design principles for constructing a heterarchical
system,and attempt to remove hierarchies in their model.
Shaw[24] describes a distributed scheme for dynamic sched-
uling in a cellular manufacturing system.There are several one-
level cells in their model.Each cell controller maintains its local
information and there is no global controller.The job assign-
ment task is done dynamically by negotiation between cell con-
trollers and the scheduling task within each cell is completed by
a local scheduler.The augmented Petri Net is used to model the
bidding scheme and the performance of the proposed scheme
is compared with a centralized,shortest processing time dis-
patching scheme via a simulation model.
Lewis et al.[25] present the data flow model for a manufac-
turing control system.Under the model,a message consisting of
a sequence of tasks and a current task pointer,is broadcast to all
the machines in the system.If a machine is eligible to perform
the current task,it saves the message in its internal queue.When
the machine becomes idle,it selects a task in its internal queue
using the first in,first out principle.When the machine finishes
the task,it updates the messages task pointer and broadcasts a
corresponding message.There is no supervisory controller.
Lin and Solberg [3] propose and apply a heterarchical intel-
ligent agent framework at the unit process level in a manufac-
turing system.Their shop-floor control model treats each part
and resource unit as an agent.Each agent negotiates with others
in real-time via a market-like bidding mechanismto achieve in-
dividual objectives.For example,when a part enters a system
with some processing requirement,its agent negotiates with re-
source agents (machines) to optimize a weighted objective that
is a function of due date,price,quality,and other user defined
factors.The agent evaluates bids from several resource agents
and selects the one that optimizes its objective.Each agent has
the power to accept or deny an offer submitted by another agent.
Though no specific details comparing the performance of the
proposedframework are provided in the paper,theymention that
their framework achieves better individual decisions and overall
systemperformance compared with traditional models.
Baker [26],[27] described a market driven contract net
control architecture for advance factory scheduling (not factory
dispatching) in a heterarchical architecture.Each agent controls
one or more manufacturing resources and is connected with
others in a network.The agents negotiate autonomously with
others using estimated costs and market-based negotiation
mechanisms.Application of the architecture to a small com-
pany is also presented.
Parunak et al.[28] and Baker et al.[29] described an agent
architecture for shop floor control and scheduling.Manufac-
turing resources,managers,part types,and unit processes (the
knowledge of how to combine resources and parts to make
other parts) are modeled as intelligent agents in their archi-
tecture.They attempt to provide a mechanism for direct dialog
between customers and distributed manufacturing system for
mass customization using intelligent agent technology.If a
customer gives a pre-order to the system,the system can give
possible delivery times and the corresponding cost.Then the
customer selects the best delivery time and cost and gives the
systema final order.The systemmakes a schedule according to
this order.To solve the scheduling problem,they sequentially
take one job at a time and optimize it based on the previously
given schedule.Because they feel that the top-down flow of
control impedes the good characteristics of the agent-based
system,they do not incorporate a hierarchical structure in their
architecture.However,they design the manager agent,which
represents a human manager,for overall systemperformance.
Peng et al.[30] present the consortium for intelligent inte-
grated manufacturing planning execution (CIIMPLEX) agent
systemarchitecture,which is designed for integrating planning
and execution in a manufacturing system.They focus on ex-
ceptions handling and their resolution.They build several spe-
cialized agents such as data-mining/parameter agent,process
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 563
rate agent,monitoring agent,and scenario coordination agent,to
provide the functionality for the exceptions handling.Two ser-
vice agents,agent name server and facilitator (broker) agents,
are built for facilitating collaboration between agents.Though
each agent usually communicates with others through a broker
agent,it can also communicate directly.The gateway agent is
used for communication with existing systems such as enter-
prise resource planning (ERP),manufacturing execution system
(MES),and Capacity Analysis system.There is no hierarchical
structure in their architecture.
Defining agent as an active object with initiative Parunak
[31] views it as the next extended step to object-oriented
programming in software evolution.He mentions that agent
technology is best suited to the problems which have charac-
teristics of modular,decentralized,changeable,ill-structured
and complex [31].He also pointed out two open research
areas:methods and tools for system design and development,
and methods for understanding the dynamics of agent based
systems.Jennings and Wooldridge [32] define the character-
istics of intelligent agents.They are autonomous,responsive,
proactive,and social.They discuss the application domain for
intelligent agents.They also point out that the problems of
the agent paradigm are no overall system controller and no
global perspective.
Baker [33] surveyed dispatching,scheduling and pull algo-
rithms of factory control from the view point of implementing
them in a multi-agent system.
Agent Cooperation and Coordination:In the literature,
there are several heterarchical models that use cooperation and
coordination methods for agent interactions.Some of these are
briefly reviewed in this subsection.Smith and Davis [34][36]
present the Contract Net protocol and Distributed Sensing
System as an example system for distributed problem solving.
They discuss suitable application areas and limitations of the
contract net as well as open research problems in this domain.
The application areas include.
1) decomposable problems where the subtasks are large and
there is a need for intensive computation;
2) problems where the main objectives are reliability,bottle-
neck prevention,and distributing control;
3) domains where the node-task allocation is not known in
advance.
The limitation and open problems include
1) determining the information that should be given with
task-announcement and bid-submission,
2) result-sharing protocol,and
3) method of enhancing global systemperformance.The ad-
vantages are simplicity,reliability,and scalability.
Tilley [37] discusses a number of protocol areas that are
poorly defined in manufacturing contract nets.They are task
announcement contents,method of constructing bids,selection
of task announcement receivers,negotiation interval,etc.The
inability to predict the performance of a system and lack
of global optimality are pointed out as the main drawbacks
of manufacturing contract net protocol.Tilley [37] presents
the study of a bidding-based heterarchical systems behavior
from the computational and communication points of view,
and shows that the main constraint of the system operation is
the length of time taken by contractors to interpret the task
announced by task manager.
Nwana et al.[38] provide an overview of four coordination
techniques:organizational structuring,contracting,planning,
and negotiation.Faratin et al.[39] present a reasoning model
for service-oriented negotiation between autonomous agents.
Using several tactics and strategies to generate bids,the model
evaluates proposed bids and makes counter proposals.Oliveira
[40] presents a multi-agent systemarchitecture for an assembly
robotics cell.Their base architecture is heterarchical.Since
conflicts are possible in a multi-agent systemdue to differences
in agent objectives and global system objectives,several
cooperation policies are proposed to minimize the occurrence
of conflicts.Finin et al.[41] propose knowledge query and
manipulation language (KQML) as a communication language
between software agents.It has predefined performatives to
clarify message intentions.
Ramamritham and Stankovic [42] point out that if all re-
sources have similar capabilities and workloads,the bidding
mechanism would not be effective in real-time distributed sys-
tems.This argument,however,is based on a specific situation
in which the deadlines of tasks are firmand if one node cannot
guarantee completion of a task by its deadline,it tries to hand
over the task to another.Since their application involves all
nodes having similar capabilities and workloads,the bidding
mechanism does not work effectively.
C.Hybrid Control
Although hierarchical and heterarchical models have limita-
tions,they also have several desirable characteristics.Some re-
searchers have attempted to capture the positive aspects of both.
One of the earliest agent-based manufacturing systems,called
yet another manufacturing systems (YAMS) was developed by
Parunak [43].Astatic hierarchical structure of the global sched-
uler,workstations,and workcells,is used in the architecture.
The workstation is a production unit and workcell is a collec-
tion of workstations or other workcells.While a coarse schedule
of broad time-windows is created centrally for the entire fac-
tory based on a global viewby the global scheduler,the detailed
schedules are determined by real-time negotiation between suc-
cessive layers of the control hierarchy.Even though a hierar-
chical structure is used in the model,an agent is allowed to ne-
gotiate with its sibling agents,as well as,its parent and children
agents.
Butler and Ohtsubo [44] present a distributed scheduling ar-
chitecture,called architecture for distributed dynamic manufac-
turing scheduling (ADDYMS).There are several levels of work
cells in their model.The workcell is a physical division of re-
sources,has a site agent,and may have subworkcells.The site
agent allocates its task to the subwork sites by negotiation.To do
this,it has the database of its capabilities,resource list,knowl-
edge of the assigned operations,and their states and addresses
of other agents for communication.The local resources at a site
select an operation dynamically based on a heuristic.They can
be shared among different sites.
Tawegoum et al.[45] introduce a hybrid control architec-
ture for a flexible manufacturing system.If a sub-level con-
564 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
troller in their model is unable to meet a schedule because of
unexpected events,it requests the upper level controller to solve
them.Brennan et al.[46] discuss a hybrid control architecture
suitable for constantly changing manufacturing environments.
Concepts and reference architectures of partial dynamic hierar-
chies are also presented.
Ou-Yang and Lin [47] propose a limited version of a hy-
brid shop floor model.They use a bidding method for job dis-
patching.When the shop controller announces the new order
arrival,each cell controller prepares and submits its bids based
on processing time,inventory cost and shortage cost.Then the
shop floor controller selects the proper cell based on the sub-
mitted price and utilization level.Each cell has fixed part routing
information and there is no part agent concept in their model.
Overmars and Tonich [48] suggest a hybrid flexible numerical
control (FNC) architecture in which each servo axis has an in-
telligent controller.These controllers cooperate with each other
in a heterarchical manner and also respond to commands from
a hierarchical host scheduler.
Ottaway and Burns [49],[50] propose an adaptive produc-
tion control system (APCS),in which the transition between
heterarchical control and hierarchical control occurs dynami-
cally based on system work load.There are job,resource and
supervisory agents in their model and each agent has coordi-
nation knowledge,production knowledge,interface knowledge,
and an inference engine.When a resource agent determines that
the resource it represents is not being properly utilized,it re-
quests a supervisory agent,which has jurisdiction over the re-
sources in question.Thus,one level of hierarchy is introduced
dynamically.They develop a prototype,present comparison re-
sults between a nonhierarchical control system and APCS,and
argue that the APCS performs better.The bidding price calcula-
tion mechanismof the simulated nonhierarchical control system
in their model,however,does not appear to be well designed,
since the machines are not utilized evenly.While the utilization
level of one machine is 98.96%,the other is only 17.71%.
Maturana [51] and Maturana et al.[52] propose MetaMorph,
a multi-agent architecture,for distributed manufacturing
system.They use two types of agents:resource agents for
representing physical resources,and mediator agents for
coordination.The mediator agents use brokering and recruiting
mechanisms for coordination.Intelligent agents find other
agents through mediator agents.The mediator agents play
the role of system coordinator by encouraging cooperation
among intelligent agents.The individual resource agents need
to register themselves with mediator agents.Virtual clusters,
or organizations of intelligent agents,can be created and
disbanded dynamically.To minimize communication overload,
they propose selective communication and agent-cloning
mechanisms.They implement a prototype and test it on two
shop floors and three products.To facilitate coordination,
learning mechanisms are proposed and demonstrated on
capacity planning process.They use learning from future by
simulation,as well as,learning from history.KQML protocol
is used as communication standard in MetaMorph.Shen et al.
[53] extend the MetaMorph architecture to integrate enterprise
activities with those of its suppliers,partners and customers in
their MetaMorph II project.They use hierarchical mediators
and bidding mechanism for the cooperative negotiation among
resource agents.
Brussel et al.[54],Valckenaers et al.[55],and Wyns [56]
present the reference architecture,product-resource-order-staff
architecture (PROSA),a holonic manufacturing system
(HMS).Their objective is to attain stability with disturbances,
adaptability,and flexibility with change and efficient use of
resources.There are three types of basic holons:order,product,
resource and staff holons.Each of the basic holons is respon-
sible for logistics,technological planning (including process
planning),and the determination of resource capabilities,
respectively.The staff holon can provide centralized algorithm
such as,scheduling algorithm and can assist the basic holons.
By including staff holons,the system can have hierarchical
control behavior so that the global system performance can be
enhanced.The system hierarchy can be formed by aggregated
holons (holarchies).The aggregated holons can be created
dynamically by the self-organization of interacting holons,or
initial system design.Individual resource holons may belong
to several holarchies at several different levels.Wyns [56]
also presents a deadlock handling mechanism and a PROSA
application framework.Valckenaers et al.[55] present an
overall holonic manufacturing system design approach and
design principles for development of the control software.
Bongaerts et al.[57] propose a reactive scheduler for holonic
manufacturing system,which tries to capture the effect of local
decisions on a global performance measure by using partial
derivatives of the global performance to the local decision
parameter.Brussel et al.[58] present methods of identifying
holons and holarchy.As example cases,they discuss resource
allocation holarchy and process planning holarchy.
Bussmann [10] compares the concept of holonic manufac-
turing with agent-oriented manufacturing.Both approaches
recognize that the manufacturing system consists of au-
tonomous,as well as,cooperative manufacturing units.He
concludes that the holonic manufacturing concerns the overall
structure of the manufacturing process,and agent-based sys-
tems concentrate on the design of information processing in
a control system.He suggests the use of agent-technology to
design and implement the information processing of a holon.
Bussmann and McFarlane [59] discuss control requirements of
future manufacturing systems as decentralized/resource-based
control architecture,generalized and flexible control inter-
action,reactive,proactive,and self-organizing control.They
argue that a holonic manufacturing control scheme can support
the requirements since it has properties of autonomy,coop-
eration,self-organization,and reconfigurability.McFarlane
and Bussmann [60] review the use of holonic manufacturing
concepts in production planning and control.
Gou et al.[61] and Luh and Hoitomt [62] propose a holonic
scheduling algorithm based on the Lagrangian relaxation
method.They define a holonic model for the algorithm.The
model has two hierarchical levels which are,factory level and
cell level.The factory level has product,part,cell,factory
coordinator,and factory holons.The cell level has part,ma-
chine-type,cell coordinator,and cell holons.The algorithm
is an iterative,decomposition algorithm.The part precedence
constraints of the factory level and the machine capacity con-
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 565
straints of the cell level are relaxed with Lagrangian multipliers.
The scheduling problems are decomposed to cell and part-level
subproblems.The latter are solved via dynamic programming.
The cell-level and factory-level problems are solved by the
conjugate subgradient method based on the subproblem objec-
tive values.The factory coordinator and the cell coordinator
holons generate appropriate coordination information to guide
the schedule quality.The part precedence and machine capacity
constraints are relaxed,however,and the generated solution
may not be a feasible one.
Giebels [63] proposes a control concept,EtoPlan (Engi-
neer-to-order Planning),to integrate the design,as well as,
process and production planning tasks in manufacture-to-order
environments.Three generic information structures for prod-
ucts,resources,and orders are presented for integration of the
three planning tasks.They argue that a hierarchical control
structure,although structurally rigid,is necessary to predict the
manufacturing environments and enhance the global objectives
of the company.They propose an evolution-based control
model.The temporary planning hierarchies of Applicability
Groups (AGs),which correspond to the orders,can be dynam-
ically created and deleted in their model.An AG is defined
as a virtual group of all resources which are considered
applicable for the (partial) execution of a given order. The
AGs can be drawn up,split up manually or automatically,
and are controlled by its autonomous AG controller.The AG
controller has four functions which are,planning,dispatching,
monitoring,and diagnostics.Each AG controller can interact
with its parent,children,and siblings within the same order
directly,while interacting with other agents via the resources.
For the shop floor level scheduling function,a scheduling group
(SG) is introduced,but does not make detailed operational
plans.The lower level AGs make the detailed operation plan
autonomously within the boundary constraints set by the SG.
An aggregate order planning method is also presented for
higher level integration of macro process planning and resource
loading.
Shen and Norrie [64] survey agent-based systems for a man-
ufacturing area,which are,enterprise integration and supply
chain management,planning and scheduling,holonic manufac-
turing systems,and they also present key issues with an anno-
tated bibliography.
D.Evaluation
Leung and Suri [65] provide an overview of performance
evaluation techniques for discrete manufacturing systems,
such as,physical experimentation,static allocation model,
probability model,queuing network,simulation,perturbation
analysis,Petri Nets,algebraic,and qualitative and hybrid
models.The evaluation of a modeling framework,or system
architecture for effectiveness in achieving proper control,is
quite important.Although other methods have been used,sim-
ulation is the most popular method used in the literature.From
the point of view that heterarchial shop floor control systems
are message-based and communication intensive,Veeramani
and Wang [66] propose a performance analyzing methodology
of auction-based shop floor control systems.Their analysis
focuses on measures such as,auction throughput and auction
time,rather than manufacturing system related measures such
as,production throughput and part-sojourn time.They use
closed queuing network models and simulation approaches to
estimate the performance of auction-based shop floor control
schemes.The proposed closed queuing network model is used
for rough-cut evaluation and more accurate evaluation is done
by simulation.
Nandula and Dutta [67] present a modeling methodology
of auction-based heterarchical system by using Colored Petri
Net for system performance evaluation to estimate system
throughput,work-in-process,deadlock existence,and resource
utilization.They apply the methodology to AGV path eval-
uation in a manufacturing system,however,because simple
random selection of machines is used for the modeling of
auction mechanism in their model,it represents an extreme
case of auction-based manufacturing.
Prabhu and Duffie [68] and Prabhu [69] present a differ-
ence equation-based model to analyze the dynamics of their
heterarchical feedback control algorithm proposed in Duffie
and Prabhu [22].While the prediction of part arrival time
trajectories is possible by the model,it is limited to a single
machine and a specific control mechanism.
III.A
NALYSIS OF
E
XISTING
F
RAMEWORKS
The main drawbacks of hierarchical control systems men-
tioned in the literature can be summarized as follows.
 Structural rigidity:It is difficult to add,modify,or delete
resources.To modify structure,the system is required to
be shutdown and all data structures of higher levels need
to be updated [54].
 Difficulty of control system design:It is necessary for a
hierarchical systemdesigner to consider the large number
of interrelationships related to failures and to explicitly
program the relationships in order to get a fault tolerant
system [8].
 Lack of flexibility:Production planner and scheduler of
higher level controllers assume deterministic behavior of
their lower level components.Unforeseen disturbances
such as machine breakdown invalidate the plan and
schedule [54].
The main problems of heterarchical control systems can be
summarized as follows.
 Lack of global information:Since each intelligent
agent only attempts to achieve its objective without
considering the global objective,there might be a contra-
diction-problem between local objective and the overall
system performance [8].
 Difficulty in predicting systemperformance:Since the in-
teraction of intelligent agents may lead to unstable dy-
namics,it is difficult to predict systemperformance or the
behavior of individual parts [23].
 Sensitivity of market rules:the global performance is very
sensitive to market rules and to the fine tuning of the rules
followed by the agents and holons [54].
Developers of the hybrid framework have attempted to over-
come some of the problems mentioned above by combining
features of both hierarchical and heterarchical frameworks.
566 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
They have developed hierarchical structures to enhance global
performance with coordination between holons and agents.
While [51][56] and [63] adopt dynamic hierarchical struc-
tures,other researchers use static hierarchical structures in their
modeling framework.Some researchers for example,[45],
[49] and [50],choose dynamic introduction of a supervisor
or high-level controller when a low-level controller is unable
to meet forecasted schedule.Others use static introduction of
supervisory controllers.To overcome the structural rigidity and
lack of flexibility,most researchers in the hybrid framework
area adopt a negotiation mechanism for real-time task alloca-
tion.Although MetaMorph uses a learning module and PROSA
uses a centralized scheduling holon,optimization and learning
techniques are rarely used in the existing models.Drawbacks
of some of the existing hybrid frameworks are discussed in the
next section.
IV.N
EW
M
ODELING
F
RAMEWORK
Our modeling framework incorporates elements of the hi-
erarchical and heterarchical frameworks and is thus,a hybrid.
Since a manufacturing control hierarchy does not change in
a short time period (daily or weekly),and the dynamic hier-
archy makes the control system complex,we use a static hier-
archy in our model.However,manufacturing components can
be added,deleted,or modified while the system operates.To
overcome the structural rigidity and lack of flexibility,a nego-
tiation mechanism for real-time task allocation is used.Lower
level agents (holons) may autonomously make their negotia-
tions within the boundary conditions that the higher level agents
set.Thus there are horizontal and vertical negotiations in the
model.In this description,we use the terms of agents and holons
interchangeably.
A.Holon/Agent Types
Parts (orders),machines,and material handling devices are
equipped with intelligent agents.There are also cell and system
level agents.The intelligent agents may reside in their corre-
sponding physical entities or in a separate computer system.De-
tails of each are described below.
Part Agent:When an order enters the system,related part
agents are created.For example,if parts A and B are included
in an order,part agents A and B are created with appropriate
parameters and objectives.Not all agents of a given type have
the same objective.While achieving high-quality could be the
objective of one part agent,fast delivery may be that of another.
When a part agent negotiates with machine or material handling
device agents,it selects the best one that satisfies its objective.
Each part agent obtains its process plan fromthe database after
it is created.Multiple process routes can be used for flexible
routing.The part agent also has its target bulletin boards through
which it negotiates with other agents using a bidding protocol.
Fig.1 shows an example of part agent architecture.
Machine Agent:Machine agents represent machines.Each
machine may have a different set of objectives.A machine
agents objective could be to take on high priority jobs,or jobs
requiring least set-up,or jobs with similar processing require-
ments.Amachine agent has information concerning capability,
Fig.1.Part agent.
objective,quality level,speed level,current status,part setup,
and processing time table of the associated machines.It also
has a bidding protocol,task evaluation,and selection logic as
well as bid calculation logic.
Material Handling Device (MHD) Agent:While there are
limited efforts at modeling the overall systems design problem,
many of them do not explicitly capture the interactions of the
material handling system with the manufacturing system.It is
important to consider these interactions because material han-
dling system design problems have an integrative influence on
material flow over a manufacturing shop-floor.Material han-
dling,and therefore the material handling devices,are a vital
link between the manufacturing resources that process the parts.
Even if machines are processing at peak-levels of performance,
a slow MHD may become a bottleneck and drastically degrade
performance of the system.
Material handling device (MHD) agents represent MHDs
such as automated guided vehicle (AGV),fork-lift truck,gantry
robot,overhead crane,etc.Similar to the machine agent,each
MHD agent has information concerning capability,objective,
speed level,current status,transportation time equation,bid-
ding protocol,task evaluation,and selection logic,as well as,
bid calculation logic of the associated MHD.
Cell,System Agent:A cell agent represents a department or
a cell to which the part,MHD or machine agents are assigned.
Agents in a cell may be grouped physically or logically.Phys-
ical grouping occurs when the agents are co-located.Logical
grouping occurs when the agents are distributed.Cell agents
(controllers) are those that control the MHDand machine agents
assigned to that cell.Thus,we may have a MHD controller and
a machine controller for each cell.In some systems it may be
desirable to assign machines and MHDs of a given type to a
particular cell,e.g.,surface finishing equipment to one cell,au-
tomated guided vehicles (AGVs) to a fully automated section
of the shop-floor,stacker cranes to a warehouse,and so on.In
such systems the cell controller is the controller that controls
the work-stations or MHDs of the given type.Thus,we could
have a cell controller that in fact,controls all stacker cranes or
all AGVs or all surface finishing equipment.The manufacturing
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 567
Fig.2.Three levels of decisions and interactions.
plant where the production takes place comprises the system.
System agents control all cell controllers.A high-level system
controller controls cell controllers in a warehouse,general man-
ufacturing area,automated manufacturing area,painting area,
finishing area,assembly area and so on.
Cell and systemagents have similar structures but their roles
differ in capability,objective,global optimization algorithms,
and reactive scheduling mechanism.The relationship among
global optimizer,reactive scheduler,and individual part agents
is described in the scheduling architecture section.Examples
of cell or system agents objectives are:transferring the most
number of parts,processing the most variety of parts,having
the least number of MHD or machine breakdowns,achieving
high processing and transfer throughput rates,transferring parts
with special processing,handling requirements,and so on.
B.Interaction Among Agents
The framework presented in this paper expands each limited
approachhierarchical and heterarchical,to one encompassing
both through the use of a time-to-decision concept.A holonic
or any other intelligent agent-based framework that is entirely
based on intra-level negotiation or communication,is likely to
be inadequate.In practice,in addition to the intra-level inter-
action among the agents,there are vital interactions that take
place between different level agents.For example,a globally
optimized schedule can be generated in a higher level agent and
be released to lower level agents.
The proposed framework assumes that decisions are made at
three levelsindividual (part,machine,and material handling
device agents),cell,and systemlevels (Fig.2).Individual agents
negotiate with one another to achieve their individual objectives.
As pointed out earlier,not all agents of a given type need to
have the same objective in the presented framework.Decisions
made at upper levels may take into consideration their own indi-
vidual interests as well as those of the lower level entities.This
could be achieved by appropriately combining and weighting
the objectives of the agents at the two levels.For example,a cell
agents decision should not only try to maximize order picking
throughput rates in that cell but also respect preferences and lim-
itations of the individual agents belonging to that cell.Similarly,
a systemagent must take into consideration cell agent interests
and limitations in addition to system performance objectives.
Fig.3.Modules used in decision making by agents at various levels.
A negotiation based planning algorithmis a good candidate for
this purpose.It could be used not only for real-time decision
making,but also for higher level agents planning.
Just as upper level agents must take into consideration in-
terests of the lower level agents,lower level agents must also
contribute toward achievement of upper level agents objectives.
Thus,while attempting to satisfy their personal objectives,in-
dividual agents must also attempt to satisfy their cells objec-
tives.These objectives can be translated into several specific
constraints for the agents.In general,a cells objectives need not
always be in conflict with that of an agents objectives.In fact,
the rules for action,agents local objectives as well as their in-
teraction mechanisms need to be carefully designed by a system
designer to enhance cell or system objectives.Just as an indi-
vidual agent must contribute toward meeting a cells objectives,
a cell must also contribute toward achievement of the system
objectives.A proposed scheduling and control architecture for
this cooperation is presented in the next subsection.
Another responsibility of the cell agent,and more so of the
systemagent,is to ensure stability of the systemso that the ne-
gotiations at various levels lead to a convergent solution.When a
pure heterarchical negotiation is used in a control system,there
is a possibility that the system behaves in a chaotic,unstable
manner (see,for example,[22],[70],and [71]).Higher level
agents have a broad view and are thus able to make decisions
to prevent such unstable behavior.For example,if no individual
agent is willing to process or transport a part,a cell agent or
systemagent has the power to force an agent to accept specific
processing or transportation assignment(s).The systemagent is
also concerned about the overload or underload of cells for sta-
bility reasons and may force cell agents to move toward stable
solutions by declining permission to cell agents on specific pro-
cessing/transfer requests and/or providing incentive to bid on
others.Just as the system agent is concerned with stability and
convergence at the system level,cell agents are similarly con-
cerned at the cell level and use discretion in approving permis-
sion requests from individual agents or forcing assignments on
them.
Several modules can be used for decision making (Fig.3).
The learning module may include neural network based learning
algorithms.For example,when a part agent type always selects a
568 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
specific machine agent for a task via negotiation,it could learn
the pattern using the learning module and for future bids,di-
rectly contact the machine agent to check its availability before
announcing its task to the task board.The database module may
include process plan information.The knowledge base module
may include behavior rules and a cost function for negotia-
tions.The algorithm module may include mathematical pro-
gramming models,genetic algorithm,heuristics,and queuing
network models for obtaining key system performance mea-
sures.Early queuing network approaches (e.g.,Solberg [72]) are
considered by some practitioners to be unrealistic,but a newer
alternative approach with great potential has emerged [73].Ac-
cording to Whitt [74],this approach is an approximate analysis
of a more exact model.It builds a more realistic model by re-
laxing many restrictive assumptions (e.g.,exponential inter-ar-
rival and service times) of the earlier approach.In fact,it is em-
bedded in publicly available software such as QNA [74] and
RAQS [73].The analysis technique is based on two-moment ap-
proximations.A major advantage of this latter approach,is that
it requires only the mean and variance of the service and inter-
arrival times and is especially suitable for embedding in our
framework,because the agents rely on the algorithm module,
not for an exact solution,but to get approximate values of key
performance measures quickly.
Each agent uses a subset of these modules.For example,
individual agents primarily use the data and knowledge base
modules,and to some extent,the learning module in their
real-time negotiation with one another and with cell/system
agents.They use the algorithm module less frequently than
higher level agents.The decision making process is triggered
whenever an event occurs.An event is said to occur when
1) part comes into the system;
2) part is almost ready for the next processing operation;
3) part is almost ready for the next transportation;
4) machine or MHD breaks down;
5) request for expediting a job comes in;
6) unanticipated events (that require a decision to be made)
occur;
7) a machine or MHD becomes available,and so on.
Decision making at the lowest level does not usually require
the algorithmmodule.Much of the decision at this level is made
to satisfy individual interests using simple rules.Of course,our
framework can be expanded to allow individual agents to uti-
lize the algorithmmodule in decision making,when necessary.
The algorithm module is used primarily by the cell and system
agents in making cell-wide or system-wide decisions.Negotia-
tions are conducted by the cell and systemagents with the help
of the learning and algorithm modules.While the primary ob-
jective is to optimize systemperformance,a critical objective is
also used to ensure that the systemis stable and the negotiation
process leads to a convergent solution.
Not all decisions require interactions between all the agents
and all levels of agents.For example,the decision to assign a
part to a cell is usually negotiated by the part agent,cell agents,
and systemagent.The cell agent already has information on its
transportation and operation capabilities,and hence,there is no
need for involving MHD and machine agents for this decision.
Besides,the higher level agents are also concerned about cell
Fig.4.Negotiation protocol.
loading for stability reasons and it is reasonable to leave the
MHD and machine agents out of this decision.Similarly,the
decision to transport a part requires interaction between the part
agent,MHDand machine agents and the cell agents.There is no
need for the systemagent to get routinely involved at this lower
level of decision making.
C.Decision Making Process
Below is a brief description of how the intelligent
agent/holonic framework can be applied to the decision
making process.Fig.4 shows our negotiation protocol,which
is a modified version of the Contract Net [34].Assume a part
is about to complete its current operation and is ready to be
transferred to the next operation.It broadcasts a message to
all the other agents via a bulletin board that it is almost ready
for the next operation as well as the transportation.MHD and
machine agents that are capable and willing to perform the
transportation and operation,respectively,submit bids to the
part agent.
When a machine or MHD agent constructs a bid,it uses the
part information contained in the data base,its current and fu-
ture workload and any special information included in the bid
solicitation,as well as its objectives in setting a price for the bid.
For example,if a part has loading restrictions and requires a rel-
atively long time to load and the MHDs primary objective is
to maximize transfer rate,then the price set can be expected to
be relatively high because the resource agent may not be able to
maximize its objective if it undertakes transfer of this part.On
the other hand,another MHD agent whose primary objective
may be to transfer parts with special needs may set a relatively
lower price if indeed it is available in the time windowspecified
by the part agent.The price setting is dynamic in the sense that
the same resource may set different prices for the same opera-
tion for the same part at different times.This is because the price
in our model depends upon current utilization of the resource
agent,jobs to which this agent is already committed,current
cell and system wide objectives,and other factors.
After a part agent selects a best bidder,it needs to get per-
mission from its higher level agent for committing to the con-
tract.Sometimes assigning a MHD or machine for a particular
part may be in the best interests of the part agent,but not in the
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 569
Fig.5.Scheduling and control architecture.
best interests of the cell.Clearly,the part agent cannot make
this determination in an unbiased manner and that is why it is
required to receive permission from a cell agent.Higher level
agents can have a greater degree of look-ahead capabilities
than lower level agents.Look ahead capabilities can be incorpo-
rated in a number of ways.For example,a higher level agent may
reduce the time available on key resources (individual agents)
due to an anticipated increase in processing or material han-
dling operations required of the key resources.Although an in-
dividual part agent may prefer a certain machine or MHD,a
higher level agent may prohibit such a matching because it may
foresee use of the resource agents in other high priority opera-
tions.In approving or denying the permission request from an
individual part agent,a cell agent may consult the data base,
learning module,knowledge base and algorithm module to de-
termine whether approving the request has any adverse effect on
the cell performance and see whether it is in the overall interests
of the system.
D.Scheduling and Control Architecture
Fig.5 shows our proposed scheduling and control architec-
ture for robustness and global optimization.There are three
levels of agents:low-level,middle-level guide,and higher
level global optimizer agents.Low-level agents correspond to
part,machine,and MHD agents in the conceptual framework.
Middle-level guide and high level global optimizer agents may
be considered as components in cell and systemagents.
A higher level agent,which has a global perspective,gener-
ates a globally optimized schedule.This schedule includes re-
source assignment for each part and their sequence.While the
middle-level guide agent and lower level agents handle current
actual manufacturing,the higher level agent prepares the next
set of manufacturing tasks.The middle-level guide agent takes
the schedule and guides the lower level agents.The part agents
know the preliminary resources assigned to them by the higher
level agent when they enter the system.However,the part agent
can reselect resources based on the real-time shop floor con-
ditions by negotiating with the machine and material handling
device agents.When necessary,it announces its tasks on the
task board,gets bids from the available resources and selects
the best bid.If the resource is different from the one assigned
by the higher level agent,the part agent asks the middle-level
guide agent for permission to change its assigned resource.The
part agent must provide information on the current participants,
especially the originally assigned resource,the best bidder and
their bid data to the middle-level agent so that the middle-level
agent can decide whether or not to grant permission.The logic
behind this is since the middle-level guide agent has a more
global view,the final decision has to be made by the middle-
level agent rather than the lower level agent.
The middle-level guide agent can be classified as a reactive
scheduler.It maintains updated aggregate planning information
such as the number of parts assigned to each resource.When
it gets a request to change the original assignment,it examines
the projected workload of resources in its control area and grants
the request only if the assignment change is necessary or desir-
able.For example,a particular machine originally assigned to
process a part may have broken down and therefore a new ma-
chine must be assigned to the part.Notice that the hybrid archi-
tecture requires interactions between higher level,middle-level
and lower level agents.
While the hybrid architecture attempts to keep the decision of
the higher level global optimizer because it can provide higher
performance under normal operating conditions,it also provides
flexibility in reacting to disturbances.Moreover,this architec-
ture can be easily converted to a pure hierarchical or heterar-
chical architecture if necessary.If the higher and middle-level
agent is bypassed,we have a heterarchical structure.If nego-
tiation at the lowest level is prohibited,we have a hierarchical
architecture.
While the proposed architecture is similar to the architecture
proposed by Bongaerts [75],there are also differences.Our ar-
chitecture gives ultimate authority to the higher or middle-level
agents,whereas Bongaerts architecture gives it to the lower
level agents.The higher level agents have an advisory role in
Bongaerts model whose focus is on the design of an online
manufacturing control (OMC) agent,and the interaction mech-
anism between an offline reactive scheduling agent and the
OMC agent.Our focus,however,is on the interaction between
middle/higher level agents and lower level agents.
There are several factors that make this new framework
attractive.First,traditional models cannot adequately represent
any manufacturing system,let alone distributed and next
generation manufacturing systems.Thus,the applicability of
traditional models to MHSs which interact with processing
equipment is questionable.For example,a queuing network
model,which can quickly provide steady-state estimates of
performance measures,cannot adequately capture the nego-
tiations that take place among agents or between an agent
and cell/system.They cannot represent different individual
objectives and their interactions.Second,there have been
several successes in applying natural phenomena to design
and operational optimization problems.For example,neural
networks,genetic algorithm,and simulated annealing have all
tried to mimic natural phenomena in solving problems.In fact,
they have been applied extensively to scheduling and other
operational problems in material handling systems.They have
been relatively successful as general purpose problem-solvers.
The constrained negotiation model in the presented approach
utilizes natural phenomena that have worked very well in all
aspects of society.
570 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
E.Comparison With Other Hybrid Frameworks
It is difficult to conduct a fair comparison of different frame-
works and architectures because of their conceptual nature and
wide range of production environments.A qualitative compar-
ison of the proposed framework and other existing hybrid frame-
works however,could be made.In this section,structural char-
acteristics of the proposed framework are compared with others.
Since general descriptions of existing hybrid frameworks are in
the literature review section,only the weak points of existing
frameworks compared with our proposed one are discussed in
this section.
In YAMS [43],while a coarse schedule of fairly broad time-
windows for classes of machines is created centrally based on
the global view by a global scheduler,a detailed schedule for
each individual machine is determined by real-time negotiation
between successive layers of the control hierarchy.Since the
detailed schedule is determined only by negotiation between
agents,it may still have a critical weak point of heterarchical
framework,which is generation of globally inferior schedule.
In MetaMorph [51],[52],the hierarchical structure of manu-
facturing control is dynamically changed by agents clustering
but the clustering mechanism is not clearly presented.While
agents need to register with mediator agents,and can negotiate
with other agents only through mediator agents in their frame-
work,agents can directly negotiate with each other through bul-
letin boards in our framework.Since planning and scheduling
is generated only by negotiation between agents in their frame-
work,a systemgenerated by their framework may result in glob-
ally inferior operation.In order to minimize communication,
they use the agent-cloning method.To get real-time status how-
ever,the cloned agents need to be updated,at near real time,thus
showing that there are no existing benefits.Furthermore,cloned
agents can develop only promissory plans for tasks and only re-
source agents assess these promissory plans at discrete periods
and can commit to a task.If the promissory plans are not good
when resource agents make decisions,all processing should be
repeated and there is possibility of deadlock.The other weak
point of their framework is that the Enterprise Mediator can be
a bottleneck.Since the Enterprise Mediator has a central coor-
dination role,its breakdown will cause the entire systemto stop.
In APCS [49],[50],when a resource agent determines that
the resource it represents is not properly utilized,it requests a
supervisory agent.When the services of the supervisory agent
are no longer required,the role of the supervisory agent will
be removed fromthe resource agents behavior logic.Thus,the
transition between heterarchical control and hierarchical con-
trol occurs dynamically based on the systems workload.It is
not clear however,howa resource agent knows that its physical
resource is not being used properly when it only has local infor-
mation.Moreover,since a part agent makes a schedule only by
negotiating with resource agents,the schedule may result in an
inferior schedule from a global system perspective.
PROSA [54][58] is a sophisticated reference architecture
developed by active participants in the Holonic Manufacturing
Systems Consortium (see http://hms.ifw.uni-hannover.de).In
this architecture,aggregated holons such as workstation,shop,
and factory could emerge out of self-organizing interaction
TABLE I
C
OMPARISON
B
ETWEEN
PROSA
AND THE
P
ROPOSED
F
RAMEWORK
or be explicitly designed.As far as we know,the method for
self-organizing interaction has not yet been developed and
aggregated holons (holarchies) are defined by the system de-
signer,not the self-organizing interaction in their example case
[58].It can be argued that the manufacturing system structure
must be designed in a more systematic way and not in an emer-
gent way.Manufacturing layout such as job shop,cellular,and
flowline,should be designed with the consideration of product
characteristics,material handling perspective,and company
strategy.Staff holons,such as scheduling,online control,and
CAD holons are independent holons in their architecture and
provide only advice to basic holons such as,product,resource,
and part holons.The ultimate decision power belongs to the
basic holons.While they follow the advice of staff holons
in normal conditions,basic holons are allowed to ignore the
advice and autonomous actions if the advice is deemed to
be of poor quality.The mechanism that permits basic holons
to decide whether staff holons are providing sound advice
when they only have a local perspective is questionable.In our
model,the function of staff holons are embedded in higher
level agents.Since a higher level agent has a broad view in
terms of upcoming workload and system status,a lower level
agent must get permission from a higher level agent if it wants
to change fromthe schedule provided by the higher level agent.
By doing this,a globally optimized schedule will be preserved.
If the higher level agent is out of service,however,the lower
level agent can act autonomously.While a product holon makes
and provides process plan in their architecture,part agent can
retrieve the process plan information from database module in
ours.While the scope of PROSA is an entire manufacturing
system that includes product development,process planning
generation,production planning,scheduling and control,ours
is mainly for production planning,scheduling and control.
Table I summarizes the differences between the two.
In summary,our framework is simpler (fewer levels and
smaller number of agents and minimal interaction),natural
(has elements of traditional hierarchy),easy to implement and
provides a mechanism to get a globally near-optimal solution
that is also robust against disturbances.
V.C
ONCLUSION
The holonic concept clearly recognizes the horizontal and
vertical interactions that take place among lower and higher
level agents.Traditional control frameworks do not adequately
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 571
model the holonic concept.The hierarchical framework almost
completely ignores the horizontal interactionsespecially
those at the lowest level.The heterarchical framework focuses
almost entirely on horizontal interactions and ignores the
vertical ones.To model the complexities and uncertainties in
modern manufacturing and distribution systems there is a need
for a framework that integrates features of the hierarchical
and heterarchical frameworks.This paper presents such a
framework for future generation manufacturing systems that
balances between the two frameworks and explicitly captures
horizontal,as well as,vertical decisions made between various
levels of agents.
It is distinctly different fromthe other holonic and intelligent
agent-based frameworks.
 It assumes decisions are made at three or more levels
and considers interactions among the entities making de-
cisions at the various levels.
 It explicitly captures interactions of the material handling
system with the manufacturing system.
 It uses a combination of optimization,learning and knowl-
edge based approaches in arriving at a solution.
 It has built-in features to ensure optimal system perfor-
mance,stability and convergence.
 It is simple (fewer agent types and interaction),natural
(use traditional hierarchy),easy to implement,providing
mechanismto get globally near-optimal solutions that are
robust against disturbances.
The negotiation and market-like models in our approach uti-
lize natural phenomena that have worked very well in aspects
of society at large.Augmenting these with optimization,knowl-
edge and learning-based models allows each agent to negotiate
wisely.We have applied the proposed framework to distribution
environments in order to validate the framework.Examples can
be found in Kim et al.[4],[5].In an application,we apply the
hybrid framework to solve an order picking and replenishment
problems in short cycle time environments where near real-time
decision making is necessary [76].
A
CKNOWLEDGMENT
The authors would like to thank the anonymous referees for
their valuable comments on this paper.
R
EFERENCES
[1] Visionary Manufacturing Challenges for 2020.Washington,DC:Na-
tional Academy Press,1998.
[2] A.Koestler,The Ghost in the Machine.New York:MacMillan,1967.
[3] G.Y.Lin and J.J.Solberg,Integrated shop-floor using autonomous
agents, IIE Trans.,vol.24,no.3,pp.5771,1992.
[4] B.Kim,R.J.Graves,S.S.Heragu,and A.St.Onge,Intelligent agent
based modeling of an industrial warehousing problem, IIE Trans.,vol.
34,no.7,pp.601612,2002.
[5] B.Kim,S.S.Heragu,R.J.Graves,and A.St.Onge,A hybrid sched-
uling and control systems architecture for warehouse management,
Tech.Rep.3801484,DSES Rensselaer Polytechnic Instit.,Troy,
NY,2001.
[6] N.A.Duffie and R.S.Piper,Non-hierarchical control of a flexible
manufacturing cell, Robot.Comput.Integr.Manuf.,vol.3,no.2,pp.
175179,1987.
[7] G.Y.Lin and J.J.Solberg,Effectiveness of flexible routing control,
Int.J.Flex.Manuf.Syst.,vol.3,pp.189211,1991.
[8] D.M.Dilts,N.P.Boyd,and H.H.Whorms,The evolution of control
architectures for automated manufacturing systems, J.Manufac.Syst.,
vol.10,no.1,pp.7993,1991.
[9] N.A.Duffie and R.S.Piper,Nonhierarchical control of manufacturing
systems, J.Manufc.Syst.,vol.5,no.2,pp.137139,1986.
[10] S.Bussmann,An agent-oriented architecture for holonic manufac-
turing control, in Proc.1st Int.Work.on IMS,Lausanne,Switzerland,
1998,pp.112.
[11] A.T.Jones and C.R.McLean,A proposed hierarchical control model
for automated manufacturing systems, J.Manufac.Syst.,vol.5,no.1,
pp.1525,1986.
[12] R.H.F.Jackson and A.W.T.Jones,An architecture for decision
making in the factory of the future, Interfaces,vol.17,no.6,pp.
1528,1987.
[13] G.Chryssolouris,K.Wright,J.Pierce,and W.Cobb,Decision making
on the factory floor:An integrated approach to process planning and
scheduling, Robot.Comput.Integr.Manufac.,vol.1,no.3/4,pp.
315319,1984.
[14]
,Manufacturing systems operation:Dispatch rules versus
intelligent control, Robot.Comput.Integr.Manuf.,vol.4,no.3/4,pp.
531544,1988.
[15] B.Boulet,B.Chhabra,G.Harhalakis,I.Minis,and J.M.Proth,Cell
controllers:Analysis and comparison of three major projects, Comput.
Ind.,vol.16,pp.239254,1991.
[16] P.OGrady and K.H.Lee,An intelligent cell control system for au-
tomated manufacturing, Int.J.Prod.Res.,vol.26,no.5,pp.845861,
1988.
[17] H.Cho and R.A.Wysk,A robust adaptive scheduler for an intelligent
workstation controller, Int.J.Prod.Res.,vol.31,no.4,pp.771789,
1993.
[18]
,Intelligent workstation controller for computer-integrated man-
ufacturing:Problems and models, J.Manufac.Syst.,vol.14,no.4,pp.
252263,1995.
[19] C.Guilfoyle and E.Warner,Intelligent Agents:The New Revolution in
Software:Cutter Information Co.,1996.
[20] T.Vamos,Cooperative systemsAn evolutionary perspective, IEEE
Control Syst.Mag.,pp.914,Aug.1983.
[21] J.Hatvany,Intelligence and cooperation in heterarchic manufacturing
systems, Robot.Comput.Integr.Manuf.,vol.2,no.2,pp.101104,
1985.
[22] N.A.Duffie and V.V.Prabhu,Real-time distributed scheduling of het-
erarchical manufacturing systems, J.Manufac.Syst.,vol.13,no.2,pp.
94107,1994.
[23]
,Heterarchical control of highly distributed manufacturing sys-
tems, Int.J.Comput.Integr.Mtg.,vol.9,no.4,pp.270281,1996.
[24] M.J.Shaw,A distributed scheduling method for computer integrated
manufacturing:The use of local area networks in cellular systems, Int.
J.Prod.Res.,vol.25,no.9,pp.12851303,1987.
[25] W.Lewis,M.M.Barash,and J.J.Solberg,Computer integrated manu-
facturing systemcontrol:Adata flowapproach, J.Manufac.Syst.,vol.
6,no.3,pp.177191,1987.
[26] A.D.Baker,Manufacturing control with a market-driven contract net,
Ph.D.dissertation,Rensselaer Polytechnic Inst.,Troy,NY,1991.
[27]
,Metaphor or reality:A case study where agents bid with actual
costs to schedule a factory, in Market-Based Control:A Paradigm for
Distributed Resource Allocations,S.H.Clearwater,Ed.River Edge,
NJ:World Scientific,1996,ch.8,pp.184223.
[28] H.V.D.Parunak,A.D.Baker,and S.J.Clark.The AARIA agent
architecture:From manufacturing requirements to agent-based system
design.presented at The workshop on Agent-Based Manufacturing,
ICAA 98.[Online].Available:http://www.erim.org/~vparunak/
icaa98.pdf
[29] A.D.Baker,H.V.D.Parunak,and K.Krol,Manufacturing over
the internet and into your living room:Perspectives from the AARIA
Project, ECECS Dept.,Univ.Cincinnati,Cincinnati,OH,Tech.Rep.
TR208-08-27,1997.
[30] Y.Peng,T.Finin,Y.Labrou,R.S.Cost,B.Chu,J.Long,W.J.Tolone,
and A.Boughannam,Agent-based approach for manufacturing integra-
tion:The CIIMPLEXexperience, Appl.Artif.Intell.,vol.13,no.1,pp.
3963,1999.
[31] H.V.D.Parunak.Agents in overalls:Experiences and issues in the de-
velopment and deployment of industrial agent-based systems.presented
at An Expanded Version of an Invited Talk at PAAM 99.[Online].
Available:http://www.erim.org/~vparunak/PAAM99Keynote.pdf
[32] N.R.Jennings and M.Wooldridge,Applications of intelligent agents,
in Agent Technology,N.R.Jennings and M.J.Wooldridge,Eds:
Springer,1998,pp.328.
572 IEEE TRANSACTIONS ON SYSTEMS,MAN,AND CYBERNETICSPART A:SYSTEMS AND HUMANS,VOL.32,NO.5,SEPTEMBER 2002
[33] A.D.Baker,A survey of factory control algorithms that can be imple-
mented in a multi-agent heterarchy:Dispatching,scheduling and pull,
J.Manufac.Syst.,vol.17,no.4,pp.297320,1998.
[34] R.G.Smith,The contract net protocol:High-level communication
and control in a distributed problemsolver, IEEE Trans.Comput.,vol.
C29,pp.11041113,1980.
[35] R.G.Smith and R.Davis,Frameworks for cooperation in distributed
problem solving, IEEE Trans.Syst.,Man,Cybern.,vol.SMC-11,pp.
6170,1981.
[36] R.Davis and R.G.Smith,Negotiation as a metaphor for distributed
problem solving, Artif.Intell.,vol.20,pp.63109,1983.
[37] K.J.Tilley,Machining task allocation in discrete manufacturing sys-
tems, in Market-Based Control:A Paradigm for Distributed Resource
Allocations,S.H.Clearwater,Ed.River Edge,NJ:World Scientific,
1996,ch.9,pp.225252.
[38] H.S.Nwana,L.Lee,and N.R.Jennings,Co-ordination in multi-agent
systems, in Software Agents and Soft Computing,H.S.Nwana and N.
Azarmi,Eds.New York:Springer,1997,pp.4258.
[39] P.Faratin,C.Sierra,and N.R.Jennings,Negotiation decision functions
for autonomous agents, Int.J.Robot.Autom.Syst.,vol.24,no.34,pp.
159182,1998.
[40] E.Oliveira,Cooperative multi-agent system for an assembly robotics
cell, Robot.Comput.Integr.Manufac.,vol.11,no.4,pp.311317,
1994.
[41] T.Finin,Y.Labrou,and J.Mayfield,KQMLas an agent communication
language, in Software Agent,J.M.Bradshaw,Ed.Cambridge,MA:
MIT Press,1997.
[42] K.Ramamrithamand J.A.Stankovic,Dynamic task scheduling in hard
real-time distributed systems, IEEE Software,pp.6575,July 1984.
[43] H.V.D.Parunak,Distributed artificial intelligence systems, in
Artificial Intelligence Systems:Implications for CIM,A.Kusiak,
Ed.Bedford,U.K.:IFS,1988,pp.225251.
[44] J.Butler and H.Ohtsubo,ADDYMS:Architecture for distributed dy-
namic manufacturing scheduling, in Artificial Intelligence Applications
in Manufacturing,A.Famili,D.S.Nau,and S.H.Kim,Eds.Menlo
Park,CA:AAAI Press/The MIT Press,1992,ch.9,pp.199213.
[45] R.Tawegoum,E.Castelain,and J.C.Gentina,Hierarchical and dy-
namic production control in flexible manufacturing systems, Robot.
Comput.Integr.Manuf.,vol.11,no.4,pp.327334,1994.
[46] R.W.Brennan,S.Balasurbramanian,and D.H.Norrie,A dynamic
control architecture for metamorphic control of advanced manufacturing
systems, in Proc.Int.Sym.Intell.Syst.& Adv.Manufac.,1997,pp.
213223.
[47] C.Ou-Yang and J.S.Lin,The development of a hybrid hierar-
chical/heterarchical shop floor control applying bidding method in
job dispaching, Robot.Comput.Integr.Manuf.,vol.14,pp.199217,
1998.
[48] A.H.Overmars and D.J.Tonich,Hybrid FMS control architectures
based on holonic principles, Int.J.Flex.Manufac.Syst.,vol.8,pp.
263278,1996.
[49] T.A.Ottaway and J.R.Burns,Adaptive,agile approaches to organiza-
tional architecture utilizing agent technology, Dec.Sci.,vol.28,no.3,
pp.483511,1997.
[50]
,An adaptive production control system utilizing agent tech-
nology, Int.J.Prod.Res.,vol.38,no.4,pp.721737,2000.
[51] F.Maturana,MetaMorph:An adaptive multi-agent architecture for ad-
vanced manufacturing systems, Ph.D.dissertation,University of Cal-
gary,Calgary,AB,Canada,1997.
[52] F.Maturana,W.Shen,and D.H.Norrie,Metamorph:An adaptive
agent-based architecture for intelligent manufacturing, Int.J.Prod.
Res.,vol.37,no.10,pp.21592173,1999.
[53] W.Shen and D.H.Norrie,Agent-based systems for intelligent manu-
facturing:A state-of-the-art survey, Know.&Infor.Sys.,an Int.J.,vol.
1,no.2,pp.129156,1999.
[54] H.V.Brussel,J.Wyns,P.Valckernaers,and L.Bongaerts,Reference
architecture for holonic manufacturing systems:PROSA, Comput.Ind.,
vol.37,pp.255274,1998.
[55] P.Valckenaers,H.V.Brussel,J.Wyns,L.Bongaerts,and P.Peeters,De-
signingholonic manufacturing systems, Robot.Comput.Integr.Manuf.,
vol.14,pp.455464,1998.
[56] J.Wyns,Reference Architecture for Holonic Manufacturing Systems:
The key to support evolution and reconfiguration, Ph.D.,Katholieke
Univer.Leuven,Belgium,1999.
[57] L.Bongaerts,H.V.Brussel,and P.Valckenaers,Scheduling execu-
tion using perturbation analysis, in Proc.1998 IEEE Int.Conf.Robot.
Autom.,Leuven,Belgium,May 1998,pp.27472752.
[58] H.V.Brussel,L.Bongaerts,J.Wyns,P.Valckenaers,and T.V.Ginder-
achter,A conceptual framework for holonic manufacturing:Identifi-
cation of manufacturing holons, J.Manufac.Syst.,vol.18,no.1,pp.
3552,1999.
[59] S.Bussmann and D.C.McFarlane,Rationales for holonic manufac-
turing control, in Proc.2nd Int.Work.IMS,Leuven,Belgium,1999,
pp.177184.
[60] D.G.McFarlane and S.Bussmann,Developments in holonic planning
and control, Prod.Plan.Control,vol.11,no.6,pp.522536,2000.
[61] L.Gou,P.B.Luh,and Y.Kyoya,Holonic manufacturing scheduling:
Architecture,cooperation mechanism and implementation, Comput.
Ind.,vol.37,pp.213231,1998.
[62] P.Luh and D.J.Hoitomt,Scheduling of manufacturing systems using
the Lagrangian relaxation technique, IEEE Trans.Automat.Cont.,vol.
38,no.7,pp.10661079,July 1993.
[63] M.M.T.Giebels,Etoplan:A concept for concurrent manufacturing
planning and controlBuilding holarchies for manufacture-to-order en-
vironments, Ph.D.dissertation,Univ.Twente,The Netherlands,2000.
[64] W.Shen,D.Xue,and D.H.Norrie.An agent-based manufacturing
enterprise infrastructure for distributed integrated intelligent manu-
facturing systems.presented at Proc.PAAM98.[Online].Available:
http://www.acs.ucalgary.ca/~wshen/papers/paam98.pdf
[65] Y.Leung and R.Suri,Performance evaluation of discrete manufac-
turing systems, IEEE Control Syst.Mag.,vol.10,no.4,pp.7786,
1990.
[66] D.Veeramani and K.Wang,Performance analysis of auction-Based
distributed shop-floor control schemes fromthe perspective of the com-
munication system, Int.J.Flex.Manufac.Syst.,vol.9,pp.121143,
1997.
[67] M.Nandula and S.P.Dutta,Performance evaluation of an auction-
based manufacturing systemusing colored Petri nets, Int.J.Prod.Res.,
vol.38,no.10,pp.21552171,2000.
[68] V.V.Prabhu and N.A.Duffie,Nonlinear dynamics in distributed ar-
rival time control of heterarchical manufacturing systems, IEEE Trans.
Control Syst.Technol.,vol.7,pp.724730,Nov.1999.
[69] V.V.Prabhu,Performance of real-time distributed arrival time control
in heterarchical manufacturing systems, IIE Trans.,vol.32,no.4,pp.
323331,2000.
[70] T.Hogg and B.A.Huberman,Controlling chaos in distributed sys-
tems, IEEE Trans.Syst.,Man,Cybern.,vol.21,no.6,pp.13251332,
1991.
[71] T.Ushio and N.Motonaka,Controlling chaos in a Hogg-Huberman
model of a manufacturing system, IEICE Trans.Fund.,vol.E81-A,no.
7,pp.15071511,1998.
[72] J.J.Solberg,The optimal planning of computerized manufacturing sys-
temsReport 9 (Revised) CAN-QUsers Guide, Sch.Industrial Engi-
neering,Purdue Univ.,West Lafayette,IN,1980.
[73] M.Kamath,S.Sivaramakrishnan,and G.M.Shirhatti,RAQS:A soft-
ware package to support instruction and research in quueing systems,
in Proc.4th IE Res.Conf.,Nashville,TN,May 1995,pp.944953.
[74] W.Whitt,The queuing network analyzer, Bell Tech.Syst.J.,vol.62,
no.9,pp.27792815,1983.
[75] L.Bongaerts,Integration of Scheduling and Control in Holonic Man-
ufacturing Systems, Ph.D.dissertation,Katholieke Univ.Leuven,Bel-
gium,1998.
[76] B.Kim,R.J.Graves,S.S.Heragu,and A.St.Onge,Realization of a
short cycle time in warehouse replenishment and order picking, Int.J.
Prod.Res.,2001,to be published.
Sunderesh S.Heragu received the B.Eng.degree in
mechanical engineering degree fromMalnad College
of Engineering,Hassan,India,affiliated to Univer-
sity of Mysore,the MBA degree at the University of
Saskatchewan,Saskatoon,SK,Canada and the Ph.D.
degree in industrial engineering from University of
Manitoba,Winnipeg,MB,Canada.
He is currently Professor in the department of De-
cision Sciences and Engineering Systems at Rensse-
laer Polytechnic Institute,Troy,NY.He previously
taught at State University of New York,Plattsburgh
and has held visiting appointments at State University of New York,Buffalo,
Technical University of Eindhoven and University of Twente,the Netherlands.
HERAGU et al.:INTELLIGENT AGENT BASED FRAMEWORK FOR MANUFACTURING SYSTEMS CONTROL 573
His current research interests are in the design of next generation factory lay-
outs,intelligent agent modeling of automated warehouse systems,integration
of design and planning activities in advanced manufacturing systems.He serves
as Principal Investigator on four National Science Foundation funded projects
in the above areas.His previous research has focused on the application of opti-
mization and/or knowledge-based techniques to facility location,layout,mate-
rial flow network,order picking in automated warehouses,scheduling,cellular
manufacturing,and group technology problems.He is the author of Facilities
Design published by PWS Company and author,or co-author,of over one hun-
dred technical articles,many of which have appeared in journals,magazines,
books,and conference proceedings.He serves on the editorial board of several
journals.
Dr.Heragu is an ABET examiner for Systems Engineering programs,a
senior member of Institute of Industrial Engineers,past Director of its Facilities
Planning and Design Division,a senior member of American Society for
Engineering Education,Institute for Operations Research and Management
Sciences,Production and Operations Management Society,Society for
Manufacturing Engineers and past member of College-Industry Council on
Material Handling Education.
Robert J.Graves received the B.S.degree in
industrial engineering from Syracuse University and
the M.S.and Ph.D.degrees in industrial engineering
from the State University of New York at Buffalo,
New York.
He is a Professor in the Decision Sciences and
Engineering Systems Department at Rensselaer
Polytechnic Institute and the Director of the Elec-
tronics Agile Manufacturing Research Institute
(EAMRI).He currently leads research to support
dramatically shorter product realization cycles in
electronics design and manufacturing.He also continues research work in
real-time scheduling and dispatching,material flow system design,and design
for manufacturing models.He is also the former U.S.Editor of Production
Planning and Control and is currently the Associate Editor of the Journal of
Manufacturing Systems.
Dr.Graves is a Fellowin the Institute of Industrial Engineers and a Fellowof
the Society of Manufacturing Engineers.
Byung-In Kim received the B.S.and M.S.degrees
in industrial engineering from Pohang University of
Science and Technology,Korea,in 1991 and 1994,
and the Ph.D.degree in decision sciences and engi-
neering systems at Rensselaer Polytechnic Institute,
Troy,NY,in 2002.
He is an Assistant Professor in the Department
of Mechanical Engineering (Industrial and Systems
Engineering Program) at the University of Memphis,
Memphis,TN.Prior to his Ph.D.program,he was an
Associate Research Engineer in the LG Electronics
Inc.,Korea.His research interest includes vehicle routing problems with time
windows,lead-time estimation,intelligent agent based shop floor scheduling
and control,logistics,informations systems,supply chain management,
optimization,systems integration,transportation systems,and manufacturing
systems.
Dr.Kimis a member of IIE,INFORMS,and Korean-American Scientists and
Engineers Association.
Art St.Onge is President of St.Onge Company,
York,PA,which is a specialized engineering firm
focused on advanced material handling information
technology applied to manufacturing and distribu-
tion.He and his company have extensive industrial
experience in creative thinking,distribution,in-
formation/control systems,logistics strategy and
modeling,manufacturing,simulation and graphic
services,having undertaken and successfully
implemented numerous projects for well known
companies throughout the world.
Mr.St.Onge is currently the President of The Material Handling Industry
Education Foundation,Inc.