correlations between emergent synthesis classes: due date based control and planning of job shops

ADVANCED ENGINEERING

www.elsevier.com/locate/aei

Advanced Engineering Informatics 20 (2006) 289–300

INFORMATICS

Correlations between emergent synthesis classes: Due date basedcontrol and planning of job shops

Attila Lengyel a,b,*, Kanji Ueda b

a Production Engineering Research Laboratory (PERL), Hitachi Ltd., 292 Yoshida, Totsuka, Yokohama, Kanagawa 244-0817, Japanb Research into Artifacts, Center for Engineering (RACE), The University of Tokyo, 5-1-5 Kashiwanoha, Kashiwa, Chiba 277-8568, Japan

Received 15 December 2005; accepted 16 January 2006

Abstract

Emergent synthesis classifies problems of artifactual system behavior into three classes depending on the completeness in the descrip-tion of the system environment and specifications. This paper introduces correlations between the problem classes and their problemsolvers. To illustrate the discussed correlations, a job shop model with make-to-order manufacturing environment is presented. Theproblem frame of the control and planning in the model is shown to be a Class III type problem and approached by using the correlatedproblem solvers of the three classes. The purpose of the job shop is to evaluate the overlapped space between the specifications of thecustomers and the capabilities of the manufacturing system and to form the behavior of the system in order to fulfill orders with highaccuracy. The structure of the model and the developed solvers indicate that to solve a Class III type problem, various Class I and ClassII problem solvers are relevant.� 2006 Elsevier Ltd. All rights reserved.

Keywords: Emergent synthesis; Manufacturing control; Production planning

1. Introduction

Emergent synthesis offers a great methodology to handleand resolve complexity in artifactual systems. Harmonizingtop-down and bottom-up features in forming the behaviorof the system, the approach provides efficient, robust andadaptive solutions to the problem of synthesis [1]. In emer-gent synthesis related solutions the global behavior of thesystem is dynamically formed bottom-up through locallyinspired interactions between the artifacts attempting toachieve the purpose of the whole system. To verify theemerging global order, top-down features are introducedthat are able to modify the order by rendering the globalpurpose to the artifacts top-down. With taking into

1474-0346/$ - see front matter � 2006 Elsevier Ltd. All rights reserved.

doi:10.1016/j.aei.2006.01.003

* Corresponding author. Address: Production Engineering ResearchLaboratory (PERL), Hitachi Ltd., 292 Yoshida, Totsuka, Yokohama,Kanagawa 244-0817, Japan.

E-mail addresses: [email protected] (A. Lengyel), [email protected] (K. Ueda).

account the local and global goals, the artifacts build uptheir emerging behavior in order to accurately achieve thepurpose of the whole system.

Emergent synthesis introduces three types of problemclasses and their emergent related problem solvers depend-ing on whether completeness of information could beachieved in the description of the system environmentand the specifications of the system. In Class I type prob-lems full completeness can be achieved in both the descrip-tion of the environment and the specifications. Althoughall constraints to be taken into account are known, to finda solution satisfying all the constraints leads to combinato-rial explosion. Therefore, emergent related methods thatcan handle combinatorial explosion are implemented inthis class. The problem solvers are evolutionary computa-tion methods such as genetic algorithms and evolutionaryprogramming. In Class II type problems the descriptionof the specifications is complete, but the description ofthe environment is incomplete. The proposition of the sys-tem is to cope with the dynamic properties of the unknown

mailto:[email protected]

mailto:ueda@ race.u-tokyo.ac.jp

mailto:ueda@ race.u-tokyo.ac.jp

290 A. Lengyel, K. Ueda / Advanced Engineering Informatics 20 (2006) 289–300

environment. To deal with this problem, the environmentalconstraints have to be determined through being in interac-tion with the environment. Learning and adaptation basedapproaches such as reinforcement learning and adaptivebehavior based methods are feasible to this class of prob-lems. In Class III types withal the incomplete environmen-tal descriptions, the description of the specifications is alsoincomplete. Besides ascertaining the dynamic environmen-tal constraints, this class has to cope with the iterativedetermination of the system structure. Emergent proper-ties, such as interactivity, self-coordination, co-evolutionand self-reference are essential in this class.

In this paper correlations between emergent synthesisclasses are under examination. One correlation can be seenbetween Class II and Class III as the implementation of theClass II approaches to handle the unknown environmentalchanges in Class III type problems. In the aspect of the pre-sented research work it is necessary to establish further cor-relations between Class I, Class II and Class III classes tosolve complex problems in artifactual systems. The paperfirst draws a schematic functional description of emergentsynthesis classes and their correlations. Although it is anoversimplified model, it shows, in the same manner, theexamined correlations between the classes. Literaturereview follows the schematic model to support the necessityof establishing the correlations. After the review, a ClassIII type problem is described as the due date based controland planning of a job shop model with make-to-order man-ufacturing environment. The developed problem solversand system structure illustrate that to solve a Class III typeproblem, various Class I and Class II problem solvers andtheir synthesis are relevant.

2. Correlations between the problem classes

2.1. Schematic functional model

In Class I problems the description of the environmentand the specification is complete. Let E denote the set ofenvironmental constraints, S the constraints of the specifi-cation and P the constraints of the human purpose, thenthe symbolical model of a Class I problem can be seen inEq. (1)

f ðE; SÞ ! R

subject to Pð1Þ

In Eq. (1), f function denotes the search method that is ableto find optimal or quasi-optimal solutions and R the resultof the method as that found near optimal or exact solutionfor the problem. The main difference between Class I andClass II is the dynamic approach to the problems. In ClassI the emergences of the solution is not time related, but inClass II the emerging solution is valid for time periods inbetween the decision makings where the dynamic environ-mental constraints need to be considered. To solve a ClassII type problem the incomplete description needs to be

completed for the decision making and the dynamic prop-erty of the environment handled. In Eq. (2), the g functiondenotes the learning or adaptation based method that con-verts the environmental changes to determined constraintsat any time t. The task of the learning and adaptationmethod is to signify the non-linear property of the uncer-tain environment, thus the g function is non-linear by t.

gðDE; tnÞ ! En; where n 2 N ð2Þ

In Eq. (2), tn denotes the time the decision is requested tobe made in the Class II type problem. In case the environ-mental changes are accurately indicated by g, then withcomplete description a Class I problem can be defined atany time and a search method can be applied (see Eq.(3)). The Class II problem is approached as a dynamicClass I problem. Rn represents the emerging solution forthe decision making that is denoted by tn.

f fgðDE; tnÞ; Sg ! Rn

subject to Pð3Þ

In Class III problems, besides the incomplete descriptionof the environment the system has to be prepared forincomplete specifications as well. The system interactingwith its superiors and collaborators determines goals andspecifications to form its behavior and achieve collectiveand individual purposes.

One can suggest that if superiors and collaborators areconsidered as environmental factors then in this approacha Class III type problem would not differ from a Class IIproblem. An architect having created a new design stylethat inspires other architects to follow did not solve a ClassII problem with the design of his first building in the newstyle. Even if the customer who ordered the building is con-sidered as an environmental factor, he certainly left incom-plete specifications for the architect that allowed him todesign in the new style. The architect using his creativityrendered these incomplete specifications to individuallycomplete specifications that drove his conceptions. More-over, his actions have been learned and adopted by theenvironment. A successful system in a Class III problemis able to perform actions which can drive itself and itsenvironment to a higher level of development without com-plete specifications.

In the schematic model, specifications are handled asdynamics of the system without considering their external-ity or internality. Let h denote the iterative method (co-evolutionary, self-organization or interactive) that is ableto determine the set of constraints of the specifications atany time t (see Eq. (4)). Note that the function h is alsonon-linear by t due to the uncertainty of the specifications.

hðDS; tnÞ ! Sn ð4Þ

In case the specifications become determined, the Class IIIproblem can be approached by the synthesis of a Class IIand a Class I problem at time tn (see Eq. (5)).

A. Lengyel, K. Ueda / Advanced Engineering Informatics 20 (2006) 289–300 291

f fgðDE; tnÞ; hðDS; tnÞg ! Rn

subject to Pð5Þ

The schematic model of the correlations between emergentsynthesis classes can be seen in Fig. 1.

In other words, when a decision needs to be made by thesystem at time tn, the specifications and the environmentalchanges need to be closely approximated to complete thedescription of the problem and through an optimizationor a local optima search algorithm the solution can befound or evaluated. In case the emerged solution doesnot meet the human purpose, the specifications can be real-located, the dynamic environment reconsidered and thesearch algorithm can attempt to find a new solution. Thisprocedure can be repeated as long as the solution is not sat-isfying. It is often discussed that if a system is able to deter-mine specifications and environmental factors at any timethen the system itself never faces more complex problemthan a Class I. A computer could certainly not be calleda calculator for the arithmetic logic unit of its central pro-cessing unit. A part of a system cannot describe the whole.The synthesis of the parts will give the whole. A systemsolving a Class III type problem will face several Class Itype and Class II type problems during its actions. Run-ning iteratively a Class I problem solver inside a Class IIor Class III problem frame will not simplify the systemto a simple solver.

In real industrial problems, exact solutions are oftenvalid only in certain situations and time intervals. The abil-ity of flexibly allocating environmental changes, ambiguousspecifications and purposes can advance problem solvers toadapt to competitive and challenging environments. Cor-porations quickly entering into emerging market segmentsby purchasing premature, but satisfying products can oftenobtain large market shares easier than those corporationsthat are trying to deliver the best products in a certain envi-ronment with static specifications. Customers often prefer

Fig. 1. Correlations between emergent synthesis classes.

quick, near optimal solutions over moderate optimal solu-tions. Reaction to these challenges is the key factor in themodern market environment. However, keeping customersatisfaction and easily acquired market shares, corpora-tions have to evolve their products to meet with the increas-ing consumer’s requirements. Fig. 1 proposes correlationsbetween emergent synthesis classes to be considered indeveloping quick response artifactual systems. Emergentsynthesis classes associate problem frames with artificialintelligence (AI) tools and the presented correlations intro-duces interactions between these tools. It is upon thesystem designer’s ingenuity to classify problems into emer-gent synthesis classes, apply the competent AI tools anddefine the parameters of the correlated interactions. How-ever, the simple usage of AI tools as engineering tools, pro-posed in this paper, is yet to come into reality, emergentsynthesis at the moment offers a straightforward approachto complex problems of artifactual systems in a compre-hensive point of view.

2.2. Literature review

In the literature several references can be found to thecorrelations between emergent synthesis classes and theirproblem solvers. The extended model of Class II includinga Class I problem in its structure can be found in the liter-ature of classifier systems where the combination of agenetic algorithm and a reinforcement learning method isimplemented to improve the rule set of the classifier system.Holmes et al. [2] point out that the learning classifier sys-tems (LCS) are adaptive and scalable to optimization prob-lems. LCS has been successfully implemented in variousfields, such as autonomous robotics, knowledge discoveryand computational economics. Other implementations ofcombined Class I and Class II problem solvers can be seenin control theory related works. Aler et al. [3] introduced acontrol system based on the combination of reinforcementlearning and genetic programming (GP). The initial popu-lation of the GP is evaluated by the learned control knowl-edge instead of a random one. The accuracy of solvingcontrol problems by the combined system increased withgreat measures. Jedrzejowicz [4] and Barbucha [5] deve-loped the combination of evolutionary and learning meth-ods further and introduced the social or populationlearning algorithm. The algorithm models the learning abil-ities of individuals in the social systems. Through the edu-cational system the individuals having greater abilities tolearn are selected in higher positions in the social system.During the evolutionary selection of the algorithm theseindividuals are preferred in generating new populations.Applications of the algorithm in the generalized segregatedstorage problem show that the social learning algorithmcan effectively handle the combinatorial explosion of thelogistic problem.

The model of Class III with the introduced correlationsbetween Class I, Class II and Class III problem solversto handle the unpredictable environmental dynamics by

Fig. 2. Problem types and their barriers introduced by Brezocnik et al. [8].


taking into account changing specifications can be found instudies on engineering applications of artificial intelligence.Cass and DePietro [6] presented an example of combiningthese methods to process monitoring, fault anticipationand aversion, fault diagnosis and resolution, and processoptimization in metal casting. Tan and Li [7] utilized evo-lutionary algorithms to design process control systems thatare customized for various specifications. In the automizedsolution of the design problem, the evolutionary algorithmevaluates the overlapped space of the specification con-straints and the possible architectures of the controller.In the problem the environment is given, but the specifica-tion is changing. Brezocnik et al. [8] classify this type ofproblems with the divergent barrier (see Fig. 2). Emergentsynthesis is not considered by Brezocnik as a suitableapproach to this type of problems. Using the model inFig. 1, the approach to the problem inside the divergentbarrier can be deducted. Moreover, emergent synthesisoffers problem solvers to all problem types discussed in [8].

As the literature review shows, the correlations betweenAI tools have been already proposed by several researchworks. This paper suggests adopting these correlations inspecifying interaction between emergent synthesis classes.

3. Due date based planning and control of job shops

3.1. Solving manufacturing related Class III problems

Approaches to Class III type problems can be found inemergent synthesis related research works. In the industrythis problem arises from the shortening life cycle of prod-ucts corresponding with the dynamic changes of themarket. The technology is being rapidly developed. Aproduct of the latest purchase, for instance, in informationtechnology may become an ‘‘old-timer’’ in less than a year.To keep up with the demand for the new products holdingthe latest technology and to avoid the high expense of

building new production systems, manufacturers have tobuild production systems that are reconfigurable andadaptable to the dynamic demand. Ueda et al. [9] intro-duced biological manufacturing systems (BMS) to copewith this problem. As a case study, a model of a line-lessassembly system is presented that is able to self-organizethe layout of the shop floor dynamically for new products.In the simulation analysis it is shown that the modeledassembly system is as productive as a line-wise system,but the reconfiguration of the line-less system for a newtype of product holds advantages in time and cost reduc-tion. Realization of line-less manufacturing systems hasyet to come due to the technological difficulties, however,Car et al. [10] show that the idea can be implemented ina simulation model to be run before the physical reconfig-uration of the system to evaluate the optimal shop floorlayout for the new product. In these studies the Class IIItype problem to be solved by BMS is considered as the dif-ficulties in the physical organization of the manufacturingsystem.

Specifications in manufacturing systems can vary notonly in the product type, but also in the optimization crite-ria for producing one type of product during its life cycle.For instance, it is well known that the cost, quality andproductivity objectives in manufacturing systems are con-flicting. In case customers request different evaluation ofthese objectives for their orders, the system also faces theproblem of changing specifications. A job shop, forinstance, with make-to-order manufacturing environmentmay face different customer specifications. An assemblyline owner ordering a specialized, one of a kind part thatis needed to repair an idle line is most certainly ready toprioritize quick lead time over the cost of the manufactur-ing. An academic researcher ordering a newly developedmechanical part probably prefers low manufacturing costover short lead times. In such an environment each cus-tomer may require different evaluation on these parame-ters. The aim of this paper is to utilize emergent synthesisapproaches and the correlations between the problem clas-ses to develop a job shop planning and control model thatsupports shop floor managers to evaluate the coherencebetween lead-times and manufacturing costs taking intoaccount avoiding bottlenecks as early as in the productionplanning phase. The job shop model is imaginary and onlyserves as an environment to illustrate the performance ofthe problem solvers built up by following the model shownin Fig. 1. The implemented cost functions can be extendedby more realistic approaches, the layout of the system cus-tomized and the type of products replaced by real types.

3.2. The job shop type manufacturing system model

The model consists of several unique workstations thatare capable of executing specialized operations. The work-stations are identical and cannot be replaced by each other.The job shop operates in one of a kind make-to-order envi-ronment. Customers order unique products with unique


specifications. For instance a customer may require a prod-uct within the possible shortest delivery and accepts therelated increasing cost, while others may be ready to waitfor the delivery longer, in case the cost is decreased. Thus,the system receives unique specifications in product typeand optimization criteria for each order. In one of a kindmake-to-order environment the demand for products isunpredictable. Periods of high and low demand may followeach other indicating different work-in-process (WIP) levelsof the shop floor. WIP is one of the main environmentalconditions that influence lead-times of jobs and the accu-racy of the system in delivering the product at the timethe customer requested. For instance, in the case of highdemand to accept orders with relatively short delivery datesmay cause overdue jobs in the system. It is significant toindicate WIP and loads of machines at order acceptationto set feasible due dates for individual orders.

The discussed job shop problem is a Class III type,where the description of the specifications and the environ-ment is incomplete. The formed decision makings dependon the dynamic properties of the system, where the purposeis to organize the operation of the shop floor avoiding dis-turbances, delivering products just in time and acceptingorders by taking into account the environmental conditionsto ensure the accuracy of deliveries. A matter of course, thesystem also desires maximizing profit by allocating theusage of the resources cost effective.

3.3. The structure of the model

The structure of the model corresponds with the ClassIII model of Fig. 1 (see Fig. 3). The proposed model con-sists of a management, job shop control and productionplanning. Their interactions and scope of duty are builtup by following the model of Class III with correlations.The interactions start with the order for a product by acustomer.

Fig. 3. The proposed job shop model.

After an order having been placed, the production plan-ning interacts with the shop floor control to evaluate thecapabilities of the system by the adaptive lead time estima-tion method (see Section 3.3) and attempts to create a fea-sible production plan for the product that can be realizedby the requested due date using a genetic algorithm searchmethod (see Section 3.4). The management receives theproduction plan, adjusts the manufacturing costs relatedto the plan and offers a price to the customer. The customermay reject the price. In this case further interactions areneeded, where the system offers longer lead times and theirindicated lower manufacturing costs. The deal of the orderends in case the customer accepts an offered combinationof the lead time and cost. The next task of the system isto deliver the ordered product in the accepted time andcost. To execute this task the system has to solve severaldynamic problems described in the following paragraphs.

3.4. Job shop control

The job shop control schedules jobs on the shop floorusing feasibility function based real-time scheduling [11].In this real-time scheduling environment it also evaluatesthe appropriate due date tightness using completion tan-gent based adaptive lead-time estimation [12]. The follow-ing notation is used in the equations of this paper:

pi,j
operational time of task j of job i
pi
processing time of job i
wi
the wth tasks of job i waiting to be processed by theworkstations
si
the number of tasks of job i
di
due date of job i ri arrival time of job i
n
number of jobs in the system t simulation time.
The feasibility function (FF) based real-time schedulingcan be found in Eqs. (6) and (7).

FFiðt;wiÞ ¼1; if t ¼ djPsi

j¼wipi;j

ðdi�tÞ2 ; otherwise

8<: ð6Þ

max

Psij¼wi

pi;j

pwi ;iFFiðt;wiÞ

� �; i 2 W ; if jLj ¼ 0

minpwi ;iPsij¼wi

pi;jFFiðt;wiÞ

� �; i 2 L; otherwise

8>>><>>>:

ð7Þ

Eq. (7) represents a real-time scheduling decision makingprocess at a workstation, where W denotes the set of jobswaiting in the input buffer and L the set of overdue jobsseparated from the waiting jobs into a designated bufferat the workstation. If the designated buffer is empty thenthe method will choose the job which has the largest valueof its feasibility function multiplied with an allowanceshowing the significance of the waiting tasks of jobs. In


case there are overdue jobs then the workstation willchoose the job with the lowest value. The reciprocate ofthe allowance is taken to indicate a shortest processing timerule modified by dividing the operational time of the wait-ing task by the square lateness. The advantages of using thefeasibility function based real time scheduling can be foundin [11].

The completion tangent (CT) based adaptive lead-timeestimation can be seen in Eqs. (8)–(10).

CTiðtÞ ¼t � riPwi�1j¼1 pi;j

ð8Þ

DTnþ1ðtÞ ¼1; if

Pni¼1

wi ¼ 0Pn

i¼1wiCTiðtÞPn

i¼1wi

; otherwise

8>><>>: ð9Þ

dnþ1 ¼ rnþ1 þ pnþ1DTnþ1ðtÞ ð10Þ

In Eq. (8), CTi(t) denotes the completion tangent of jobsand in Eq. (9), DTn+1(t) the assigned due date tightnessof the incoming job at time t. Completion tangent basedlead-time estimation method has been shown to be capableof assigning due date tightness to jobs adaptively to theWIP level of the system [12]. However, the later introducedsimulation results of this paper will also confirm the adap-tive behavior of the method. At the beginning of the simu-lation when there are no tasks finished in the system, thedue date tightness of the arrival jobs will be equal to 1,since the jobs will not need to wait for processing and theirlead time will equal to their processing time. When theshop load increases and there are several finished tasksthe method will assign due date tightness to arriving jobsby the weighted average of the completion tangents ofthe jobs currently in the system at time t (see Eqs. (9)and (10)). Thus, the method uses only real-time informa-tion that is available at any time t in the shop floor.

3.5. Production planning

The production planning creates tasks and routings ofjobs. It is also responsible for setting machining parametersfor operations. The operational time of a task assumingeffective routing and setups depends on the scaling of theseparameters. Toth et al. [13] introduced intensity type vari-ables to integrate control and planning of manufacturingprocesses. Instead of rigid plans for optimal machiningparameters, such as cutting speed, depth of cut, etc., plan-ning is based on technological intensity that allows real-time scaling of these parameters to adapt to the dynamicmanufacturing environment. The approach, for instance,can be utilized in case bottlenecks appear in the manufac-turing system running a risk of overdue deliveries of jobs.Increasing technological intensity on a machine appearingas a bottleneck will cause higher machining costs conflict-ing local optima in the viewpoint of the machine itself,but the global optima of the whole system as minimizingtardiness penalty costs will be enhanced. In this research

work it is assumed that the production planning can eval-uate a range of production process times and their relatedmachining costs. When an order arrives in the systemthe production planning creates two draft plans: the tech-nologically possible shortest plan (pmin) and economicallyreasonable longest plan (pmax). Using Eq. (10), the maxi-mum (dmax) and minimum (dmin) due date can be calculatedfor the orders. The customer is able to choose a due date(dexp) in between dmax and dmin (see Eq. (11))

if dminnþ1 6 dexp

nþ1 6 dmaxnþ1 then pexp

nþ1 ¼dexp

nþ1 � rnþ1

DTnþ1ðtÞð11Þ

The task of the production planning is to find a plan thatindicates a sum of operational times equal to the expected(pexp). The implemented search method is a simple geneticalgorithm [14]. A genome (X) in the search space of themethod is represented by Eq. (12), where l denotes the res-olution of the problem and dec( ) function transforms bin-ary numbers to decimal.

pXnþ1 ¼

Xsnþ1

j¼1

pminnþ1;j þ

pmaxnþ1;j � pmin

nþ1;j

2l xj

!ð12Þ

where X 2 ½0; 1�snþ1xl and xj = dec([Xj,1, . . . ,Xj,l]).The object function of the search method gives scores to

individuals by two criteria: minimization of the machiningcosts and minimization of the operational times on heavilyloaded machines. In the machining cost estimation twotypes of costs are taken into account: machine tool costs(see Eq. (13)) and tooling costs (see Eq. (14)). The modelis simplified based on sophisticated cost estimation models[15]. As it can be seen in the equations, shortening opera-tional times will cause higher machining costs by themodel. The sum of machining costs indicated by a genomecan be seen in Eq. (15). In the equations, Mi,j denotesthe machine the jth task of job i is assigned to, cmt themachine tool cost allowance and ct the tooling cost allow-ance of machine Mi,j.

Cmtnþ1 pX

nþ1

� �¼Xsnþ1

j¼1

ctMnþ1;j

pxj

nþ1;j

60

!ð13Þ

Ctnþ1 pX

nþ1

� �¼Xsnþ1

j¼1

ctMnþ1;j

pmaxnþ1;j

pxjnþ1;j

� �pmax

nþ1;j � pxj

nþ1;j

60

!0B@

1CA ð14Þ

Cmnþ1 pX

nþ1

� �¼ Cmt

nþ1 pXnþ1

� �þ Ct

nþ1 pXnþ1

� �ð15Þ

The above described machining cost estimation in thissimple form cannot be implemented in a real industrialenvironment. In case a job shop owner is interested in eval-uating the effectiveness of the developed model for hissystem, customized data and cost estimation can beimplemented in the simulation model with minor modifi-cations.

Heavily loaded machines are estimated real-time, usingthe same approach introduced in the completion tangentbased lead-time estimation model of the job shop control.


For each machine the mean average of ratios between thetime the active jobs in the system have waited for being exe-cuted by the machine and the operational times of jobs onthe machine is calculated. Let JM denote the set of jobs onthe shop floor that have already visited the machine M atleast once, wti,M the time the job i waited at the machine,and pi,M the operational time of the job on the machine.The load on a machine is estimated by

LM ¼1

jJ M jXjJM j

i¼1

wti;M þ pi;M

pi;M

ð16Þ

LM shows an average multiplier that estimates how muchtime will be needed to process an operational time on themachine M. In case the value of this multiplier is too highcompared to DT(t), than the machine is being a bottleneckof the system. If the system aims at avoiding bottlenecks orheavily loaded machines then the production planning canset machining parameters of the plan so that on the ma-chines having high LM, operational times will be shorterand to recover the lost cost, processing times will be longeron the machines having low LM. The score of a genome X

is given by Eq. (17)

if pXnþ1 ffi pexp

nþ1

then score ¼ cCnþ1 pmin

nþ1

� �Cnþ1 pX

nþ1

� �þ l

Psnþ1

j¼1 pmaxnþ1;jLMnþ1;j

� Psnþ1

j¼1 pXnþ1;jLMnþ1;j

� ð17Þ

In Eq. (17), c denotes the weight of machining cost minimi-zation and l the weight of minimizing the expected lead-time of the job. The production plan and the minimummachining cost of the expected order are evaluated by theindividual to be found having the highest score. The firstpart of the score evaluation gives higher score to those indi-viduals that have lower machining costs and the second partfavors individuals that minimize their operation times onhighly loaded machines. The management is responsiblefor interacting with customers to form the deal of the duedate and cost for their order. When an order is passed tothe production planning, the found minimum machiningcost is assigned to the required due date and the manage-ment evaluates the price of the ordered product. If the cus-tomer is not satisfied with the price, the due date can beenlarged to reduce the price. This procedure can be repeateduntil the appropriate due date and cost are not found.

The management is also responsible for setting weightsfor the introduced score evaluation (see Eq. (17)). Selectingthe weights in different ways will result in different perfor-mances of the system. Giving more weight to minimizinglead-times of jobs will cause higher machining costs. Onthe other hand, in a cost driven market giving more weightto cost minimization is a better management strategy. Themanagement strategies are the long term operation policiesof the system. Forming the weights according to thedynamic specifications of the customers or market predic-tions will increase average customer satisfaction. The

future work of the research will focus on implementing alearning method to evaluate appropriate weight allocationsin different market environments to develop self-coordina-tion ability of the system.

In the following chapter three sets of simulation resultswill be shown. In the first simulation the weight of machin-ing cost minimization is set to one and the weight of lead-time minimization to zero [c = 1, l = 0]. In the second,both the criteria are set to one [c = 1, l = 1] and in thethird only lead-time minimization is considered [c = 0,l = 1].

4. Simulation results

4.1. The simulation model

The manufacturing system model described in the paperis imaginary, illustrating a complex job shop problem. Thetype of the manufacturing system is classic job shop withrecirculation. The number of machines in the job shop is35. Machines can perform unique tasks and they cannotbe replaced by each other. The layout of the simulationmodel can be seen in Fig. 4. The model was developed withTecnomatix eM-Plant object oriented, discrete event drivensimulation software.

The work pieces of jobs are transported individually byAGVs assigned to each job. The number of AGVs is equalto the number of jobs in the system. The transportationtime is not considered in the simulation runs. Althoughthe routing of AGVs between machines is optimized, thepurpose of the model is to generate results that can be ver-ified by any simulation model independently from the lay-out. Due to the abilities of the simulation software thetransportation times can easily be set and the layout recon-figured, thus the model is fully customizable. The fluctuat-ing demand for products is modeled by setting the timeintervals between arrivals of jobs shorter for high (1 h20 min) and longer (1 h 35 min) for low demand. A simula-tion run takes one year time period and at every 100th daythe demand is changed in the following order: high, low,

high and low. The number of tasks in the task list of ajob is chosen between 15 and 55 using uniform distribution.Each job has different routing through the shop floor and ajob may visit a machine more than once. The operationtime of a task at a machine is assigned randomly between3 and 180 min by uniform distribution as well. The produc-tion planning is allowed to change the randomly set oper-ational times by ±16% to indicate the technologicallypossible shortest and economically reasonable longest pro-cess times. To model free choice of possible due dates, thedue date of a job is assigned randomly between the evalu-ated dmin and dmax.

4.2. Simulation results of the operation policies

The following due date based performance measureshave been considered:

Fig. 4. The layout of the simulation model created in eM-Plant.

Table 1Simulation results of the operation policies


PUP
summarized unit penalty
[c = 1, l = 0] [c = 1, l = 1] [c = 0, l = 1]
FP number of finished jobs in one year PU 157 155 106
T
summarized tardiness [days/h/min] PF 5151 5807 5902
Ltmax
maximum lateness [days/h/min] PT 85/22/46 50/15/45 24/18/34 Ermax maximum earliness [days/h/min]
Ltmax 00/17/48 00/12/22 1/04/22
MAL mean absolute lateness [days/h/min] Ermax 43/07/31 8/11/46 8/15/04 MSL mean square lateness [s2] MAL 3/06/59 00/18/42 00/14/20
MSL 2.144 * 1011 8.569 * 109 5.698 * 109

Rmax 7.911 4.019 2.993

Rmin 0.669 0.744 0.710MR 1.155 1.068 1.080MDT 19.08 8.21 5.53

MLM 20.19 5.77 4.62PCm 122191860.09 131203355.97 146177910.72

MCm 23717.95 25466.54 28369.63

To measure on-time completion performance of jobsbesides the above classical measures the completion ratiois introduced (see Eq. (18)).

Ri ¼di � ri

ci � rið18Þ

Ri denotes the completion ratio and ci the completion timeof job i. In case the completion ratio is equal to one the jobis on-time; if it is less than one the job is late. Related tocompletion ratio three more measures will be presented:

Rmax
maximum completion ratio Rmin minimum completion ratio MR mean completion ratio
To measure the introduced machining cost and machineloads the following data have been collected.

MDT
mean due date tightness assigned to jobs MLMP mean machine loads
Cm
summarized machining costsfor the one year run [$]
MCm
mean machining cost [$]
The simulation results with the three operation policiescan be seen in Table 1.

One of the main criteria in make-to-order manufactur-ing environment as delivering the product before the duedate has been achieved by all the three policies with highaccuracy. Although the summarized unit penalty of simula-tion runs is 3%, 2.7% and 1.8% of the finished jobs, respec-tively, the maximum lateness and summarized tardiness arerelatively low. The classical on-time performances showthat in case of the first run job, earliness is obviously nota-ble. Maximum job earliness is 43 days; mean absolute late-ness is more than three days and the mean square latenessis two magnitudes larger than in case of the other runs.Completion ratio performance measure confirms this fact.When lead-time minimization is considered in the objectfunction of the genetic algorithm search method in the pro-duction planning, on-time performance of the systemincreases significantly. Moreover, the mean due date tight-ness assigned to jobs is 2.3 times smaller in the second and3.5 times smaller in the third simulation run compared tothe first. Since the summarized unit penalty and lateness


measures are very similar in all runs, lead-time minimiza-tion can be confirmed. The mean machining cost of jobsand the summarized machining costs represent that settingthe weights in the object function signify the performanceof the system accordingly. Although in Table 1 all theassumptions of the formal chapters can be seen, moredetailed analysis of the simulation results are necessary tounderstand the behavior of the system with different oper-ation policies. In Figs. 5–7 data samples can be seen takenat each job arrivals reflecting the available information ateach decision making the job shop control, productionplanning and management are facing to.

Figs. 5 and 6 represent the adaptation ability of thecompletion tangent based lead-time estimation method.The due date tightness is assigned according to the WIPlevel of the system. In Fig. 7 the on-time performance ofjobs can be seen. At the beginning of the simulation runs,due date performance is not satisfactory due to initially set-ting the due date tightness to one. When there are moredata samples in the system the performance of the lead-

Fig. 5. The assigned due d

Fig. 6. Work-in-process

time estimation is convincing. The lateness measures shownin Table 1 were collected in the first days of the simulationruns, later on there are no late jobs in the system in case ofall operation policies. Too early jobs appear, but theirnumber is only significant in the first run. In Fig. 8 lowermachining cost is indicated in the first run and it is increas-ing in the other runs, respectively.

Figs. 5 and 6 present that in the first run the WIP leveland the due date tightness of jobs are monotone increasing.The due date tightness is around 36 at the end of the sim-ulation. It means that when a job arrives at the system, theestimated lead-time is 36 times larger then the evaluatedprocessing time of the incoming job. To recognize the causeof this high due date tightness, the following data samplesare shown (see Figs. 9–11).

As it can be seen in Fig. 10 when lead-time minimizationis not considered, bottlenecks appear in the system. Onlytwo machines out of the 35 are withholding the perfor-mance of the whole system. In case machine load is consid-ered in the objective function, then bottlenecks are handled

ate tightness of jobs.

level at job arrivals.

Fig. 8. Machining costs of jobs.

Fig. 7. Completion ratio of jobs.

Fig. 9. Data samples of estimated machine loads (LM) with [c = 1, l = 0].


by the production planning by increasing technologicalintensity on the heavily loaded machines. The merit ofthe model is that besides easily recognizing the heavily

loaded machines, it gives the answers for how much theintensity should be increased and on which machine theaction should be taken to avoid bottlenecks.




4.3. Discussion of the simulation result

Three simulation runs with different operation policieshave been analyzed. Due to the same random seeds usedin the simulations, each run received the same input data,therefore the differences in the performances reflected onlythe different effects of the policies on the system. The man-agement giving two straightforward weights to the produc-tion planning is able to influence the whole behavior of thesystem. Simulation results confirm that in the case higherweight is given to machining cost minimization the systemwill form its behavior to produce products with lowercosts. The machining cost is reduced by setting machiningparameters cost effectively. The accuracy of delivering theproduct before the due date is not influenced by differentoperation policies. In the first simulation run, bottlenecksappeared weakening the performance of the system. Settingthe object function so that only machining cost minimiza-tion is considered does not necessarily causes bottlenecksin the system. Boundaries of random parameters of incom-

ing jobs were set by the authors so that the bottleneckavoiding ability of the model could be reflected.

5. Conclusion

In this paper, correlations between the problem classesof emergent synthesis have been introduced. Based on thisapproach to the Class III type problems, the authors devel-oped a job shop type system model in one of a kind make-to-order manufacturing environment. The aim of themodel is to adapt to the dynamic environment and chang-ing specifications, representing a Class III problem.

The production planning of the model solves a Class Itype problem using a genetic algorithm based optimumsearch method. Before running the search method the envi-ronmental constraints and the constraints of the specifica-tion are determined by the job shop control and themanagement. The job shop control using an adaptivemethod to estimate due date tightness of incoming jobs isthe Class II problem solver of the model. The management


interacting with customers to form the deal of the due dateand cost can evaluate the appropriate weights for themachining cost and lead-time minimization so that the sys-tem can adapt to different market environments. The syn-thesis of the behavior of the production planning, jobshop control and management deals with the Class III typeproblem of the system.

Simulation results with a random job shop show thatwhilst the environment is dynamic and the specificationsare changing during the simulation runs, the system perfor-mance for on time delivery of products, one of the main cri-teria in make-to-order manufacturing environment isaccurate. The management is able to control the range oftime and cost constraints of the system by evaluating theoperation policies. Emergent synthesis of top-down andbottom-up features are harmonized in the model to achievethe global purposes of the whole system.

Future work will focus on implementing a reinforcementlearning algorithm (Class II type solver) in the manage-ment to create long term operation policies of the systemin different market environments.

References

[1] K. Ueda, A. Markus, L. Monostori, H.J.J. Kals, T. Arai, Emergentsynthesis methodologies for manufacturing, Annals of the CIRP 50(2) (2001) 335–365.

[2] J.H. Holmes, P.L. Lanzi, W. Stolzmann, S.W. Wilson, Learningclassifier systems: new models, successful applications, InformationProcessing Letters 82 (1) (2002) 23–30.

[3] R. Aler, D. Borrajo, P. Isasi, Using genetic programming to learn andimprove control knowledge, Artificial Intelligence 141 (1–2) (2001)29–56.

[4] P. Jedrzejowicz, Social learning algorithm as a tool for solving somedifficult scheduling problems, Foundation of Computing and Deci-sion Sciences 24 (2) (1999) 51–66.

[5] D. Barbucha, Three approximation algorithms for solving thegeneralized segregated storage problem, European Journal of Oper-ational Research 156 (1) (2004) 54–72.

[6] R. Cass, J. DePietro, Computational intelligence methods for processdiscovery, Engineering Applications of Artificial Intelligence 11 (5)(1998) 675–681.

[7] K.C. Tan, Y. Li, Performance-based control system design automa-tion via evolutionary computing, Engineering Applications of Arti-ficial Intelligence 14 (4) (2001) 473–486.

[8] M. Brezocnik, J. Balic, Z. Brezocnik, Emergence of intelligence innext-generation manufacturing systems, Robotics and Computer-Integrated Manufacturing 19 (1–2) (2003) 55–63.

[9] K. Ueda, I. Hatono, N. Fujii, J. Vaario, Line-less production systemusing self-organization: a case study for BMS, Annals of the CIRP 50(1) (2001) 319–322.

[10] Z. Car, I. Hatono, K. Ueda, Layered based architecture for adaptivereconfiguration of manufacturing systems, in: Proceedings of the 4thInternational Workshop on Emergent Synthesis (IWES’02) 2002,Kobe, Japan, pp. 115–120.

[11] A. Lengyel, I. Hatono, K. Ueda, Scheduling for on-time completionin job shops using feasibility function, Computers & IndustrialEngineering 45 (1) (2003) 215–229.

[12] A. Lengyel, I. Hatono, K. Ueda, Dynamic due date assignment in jobshops based on emergent synthesis, in: Proceedings of the 36th CIRPInternational Seminar on Manufacturing Systems 2003, Saarbrucken,Germany, pp. 463–470.

[13] T. Toth, F. Erdelyi, F. Reyagani, Intensity type state variables in theintegration of planning and controlling manufacturing processes,Computers & Industrial Engineering 37 (1-2) (1999) 89–92.

[14] W. Matthew, GAlib: a C++ library of genetic algorithm components,Mechanical Engineering Department, Massachusetts Institute ofTechnology, MIT, 1996. Available from: <http://lancet.mit.edu/ga/>.

[15] E. Cuesta, J.C. Rico, S. Mateos, C.M. Suarez, Times and cost analysisfor sheet metal cutting processes in an integrated CAD/CAM system,International Journal of Production Research 36 (6) (1998) 1733–1747.

http://lancet.mit.edu/ga/

correlations between emergent synthesis classes: due date based control and planning of job shops

Documents