Upload
terry-williams
View
213
Download
0
Embed Size (px)
Citation preview
Towards realism in network simulation
Terry Williams *
Department of Management Science, Strathclyde University, Graham Hills Building, 40 George Street, Glasgow G1 1QE, UK
Received 1 April 1998; accepted 1 October 1998
Abstract
The use of networks handling uncertainty to provide a temporal risk analysis of projects is now widespread.However, such analyses frequently give rise to very wide probability distributions, and thus in practice are describedas not credible. This is largely because the simulations do not re¯ect the actions that management would take to
bring late-running projects under control. These are di�cult to include in models, not because the actionsthemselves are complex, but rather because the e�ects of those actions are not well-understood. These e�ects areoften much less e�ective than expected and some are counter-intuitive. However, much work has been done in
modelling projects using system dynamics, and this work can give some useful insights into the e�ects ofmanagement actions in projects, both their behaviour and indications of their cumulative impact. This paper hasattempted to describe these indications and then to apply such lessons to network simulations, to gain the bene®t ofthe insights without losing the operational advantages of the networks. Some small illustrative models of the e�ects
are given. It is hoped that the use of such modelling can help to bring additional realism to probabilistic networkmodelling. # 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Networks; Simulation; Project risk; System dynamics
1. Introduction
Temporal analyses of projects are nearly always
based on the network (PERT or CPM) concept
(e.g. [34]). In recent years, much work has been done
in trying to assess the temporal risk in projects by
including uncertainty in such networks. These analyses
have become more sophisticated, but are still very fre-
quently giving results that are not believed, frequently
because they have very wide probability distributions
when the project management knows the probable
range is very much smaller.
After introducing the ideas of network simulation,
and the bene®ts it brings, this paper traces recent
developments in network simulation attempting to
enhance the realism in the modelling. The reason for
the wide probability distributions is then identi®ed as
the need to model management action and the conse-
quences thereof. Attempts to include such action are
reviewed, but these have not been successful. This is
largely because the e�ects of such actions are not well
understood. To give some insights into the e�ects of
management actions on projects, the ideas of system
dynamics (SD) are considered. SD has been used a
number of times to model project behaviour, but
nearly always at a strategic level, unrelated to oper-
ational considerations such as the project network
(hence its lack of usage by practising project man-
agers). This paper looks at the SD work and tries to
draw out some of the lessons from the modelling that
has taken place; these lessons are then used to amend
the operational network models. This can help to give
a simple model of what would actually happen in pro-
jects, leading to more believable results. This, it is
hoped, will provide some recovery of the credibility
that network simulation has lost by providing unbelie-
vable results.
Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314
0305-0483/99/$ - see front matter # 1999 Elsevier Science Ltd. All rights reserved.
PII: S0305-0483(98 )00062-0
PERGAMON
* Tel.: +44-141-552-4400; fax: +44-141-552-6686; e-mail:
The paper is written from the point of view of theauthor who was a project risk manager for a decade
before returning to academia ®ve years ago. It aims toexplore what is useful to the project management pro-fession, rather than merely what is academically inter-
esting. A recent conference similarly concluded thatwhile there could clearly be extremes of, on the onehand, academic modelers who have no impact on pro-
jects and, on the other hand, project managers whoreject all mathematical techniques to the detriment oftheir planning. Steering a middle course, development
of technique must always be driven by the needs ofproject managers and the desire to improveperformance [51].
2. Analytical methods
Network analysis is a well-established analytical tool
when the project is deterministic. When uncertaintiesare introduced into the analysis, various analyticalapproaches have been tried. The ®rst line of attack onstochastic networks is of course straight analytic sol-
ution, and there are excellent reviews of this work byAdlakha and Kulkarni [3] and Ritchie [39]. However,since there is no simple analytic solution of the dur-
ation of the stochastic network, other than for speci®ccases, e.g. exponentially-distributed networks [30],other authors (for example Dodin [17], Devroye [16]
and Kamburowski [25, 26]) have attempted to deriveexact bounds. Finally, others analytically derive ap-proximate solutions, such as Anklesaria and
Drezner [5], using stochastically dominating paths, andMehrotra et al. [33], using the commonality of activi-ties on paths and employing a Normal approximation.However, while these methods can be shown to work
on some small or specially-constructed networks, thereis no general indication of their accuracy, nor explora-tion of their use for real-life networks.
However, the main problem with the analyticalapproaches is the restrictive assumptions that they allrequire, making them unusable in any practical situ-
ations; this means that currently they are not really ofinterest to practitioners. First of all, there are therestrictions of commission: most require very speci®c(and sometimes unrealistic) duration-distributions, and
are not applicable unless all the activities are of thisdistribution. Secondly, there are the very signi®cantlimitations of omission: these techniques are not appli-
cable if there are resource-constraints on the network(indeed, it shall be seen below that activity-criticalityindices are not even de®ned in this circumstance), and,
perhaps more importantly, there is the plethora ofcomplexities that the practising project modelerrequires to include in his analysis for the results to be
both relevant and credible, such as the following (seee.g. Ref. [47]):
. e�ects that operate across a range of activities and/or resources, such as third-party e�ects or common-
cause e�ects: the major risks within a project moreusually than not involve more than one activity (theassumption that activity-durations are iidrv's isenough to lose a technique credibility);
. unusual activity±duration-distributions; one obviousinstance that frequently occurs in practice is theneed to combine 0±1 and continuous uncertainties
when several risks a�ect the same activity, (particu-larly where some of those risks are epistemic (i.e.due to a lack of knowledge) and others are aleatoric
(i.e. of an underlying probabilistic nature) [49]; (asecond instance is given in Section 3);
. resource availabilities or requirements that are
uncertain (and possibly varying over time): these cansometimes be the critical uncertainties, for example,the precise timing of the availability window for avital resource can be the deciding factor in whether
a project meets its deadline;. uncertainties in the project network structure: whileall classical network methods assume that the net-
work itself is ®xed, in practice there can be probabil-istic or conditional branching;
. domain-speci®c uncertainties which can have very
particular uncertainty pro®les, such as the randomfailure of test-rigs, or weather-windows in deep-seaoil work etc;
and so on. Furthermore, of course, these complexitiesoccur not individually but severally: for example, whendesigning a military aircraft, there is a possibility that
one or more of the test-aircraft crashes: such an occur-rence would a�ect many streams of activity simul-taneously (avionics, engine and air-frame), it wouldcause the resource availability to be uncertain (so later
test-activities will take longer), it would cause individ-ual durations to have nonsimple distributions, andthese e�ects might occur at random.
It is perhaps worth noting ®nally that some analyti-cal methods (e.g. Mehrotra et al.'s [33]) provide onlycertain moments of the project duration. However, the
main purpose of this analysis in practise is to give aidin bidding, when the project manager is interested inquestions such as:
. what is the probability of meeting the project due-date?
. if there are liquidated damages, what is the expected
penalty (i.e.
�t
p�t� � LD�t�dt
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314306
where p(t) is the PDF of project duration and LD(t)is the damages payable for a duration of t)?.
. what is the (say) 90% con®dent project duration?
Such questions require an appreciation of the whole
duration distribution, and such distributions cannot beassumed to be any of the standard models: even for asimple stochastic network, the total duration is theresult of a combination of convolutions and taking
maxima; of course, where there are the more complexuncertainties noted above, the duration can take quitecomplicated shapes.
3. Network simulation
A number of analysts are therefore working onusing simulation to model the type of projects foundin practice. There has not been universal agreement,
obviously. Chapman [10], for example, says thatalways using Monte Carlo simulation is a simple gen-eral solution, but it is often ine�cient `using a sledge-
hammer to crack a nut'. More important, to continuethe analogy, it encourages an approach which makesthe walnuts very unattractive eating, and di�cult to
®nd. Similarly, Adlington and Christensen [4] feel thatthe use of sophisticated mathematics, simulation andspecialist risk managers is misplaced; they say that thefundamental aim of a risk analysis is to establish a
`picture' of the future which enables management tomake informed decisions, to grasp the opportunitiesand to control or make provision for the risks; how-
ever, even they point to the problem of dependencieswithin projects, implying a need for simulation.Despite these concerns, many practitioners have
arrived at similar solutions: network-type models,using classical network logic, within a general simu-lation framework to enable the uncertainties to be
modelled. Such authors include the following:
. Ragsdale [38] gives a good summary of the literatureof network simulation to that date.
. Williams [46, 47] describes an early such system, andapplications to real projects within the defencedomain.
. Kidd [28, 29] describes the VERT system, which has
some similar characteristics (also looking at simulat-ing cost and performance).
. Similar work is given by Berny and Townsend [7] in
their VISIER software.. A user-friendly fully-engineered package similarlyable to give a high degree of generality to uncertain
networks is provided by @Risk for Project [36],which applies the @Risk concept (essentially aspreadsheet, but allowing cells to be allocated ran-
dom variables) to a standard network package
(MicroSoft Project).
. Similarly, the MonteCarlo package [37] allows most
features of a network analysis to be given variability
by the user, allowing custom distribution models,
duration correlations, conditional branching, prob-
abilistic resource requirements and availabilities,
repeat-chain logic and so on.
. There are also important simulation models of
speci®c projects Ð for example, Nicolo [35] tries to
capture the whole project environment (including
nonphysical parameters), and simulates a project
network with all of these e�ects, using a scenario
generation technique, building a project GERT net-
work covering the whole environment of the project,
then using the SLAM simulation package.
Although straightforward simulation is more fre-
quently used in practice, there have been theoretical
advances in improving the information derived from
simulations, generally using either variance reduction
(e.g. [6], particularly using antithetic variables as
in [44]) or conditional simulation, choosing which arcs
to simulate (e.g. [2, 18, 31]). The aim of this is to
reduce the number of iterations required. However, the
requirement for this has decreased markedly as compu-
ter power has increased, and furthermore since these
studies are generally carried out on networks popu-
lated by subjective probability-distributions whose pro-
venance is fairly weak, the search for precise accuracy
in practical applications can be rather spurious.
These simulations are aimed not just at estimating
the project duration distribution of the project dur-
ation, but also the relative importance of the activities.
Until the late 1980's, that meant estimating the activity
criticality indices [19]. However, it is now well-known
that these can be misleading and of course they cannot
be calculated when there are the more complex model-
ling features described above (although work such
as [8] is seeking to rede®ne criticality in a resource-con-
strained setting). These problems motivated the devel-
opment of the cruciality index, de®ned for an activity
as the correlation between the (activity-duration) and
the (total project-duration), to be used in collaboration
with (not simply replacing) traditional criticality
indices [48]. This index does not depend on the critical
path. This means that the output of a network simu-
lation (even if resource-constrained) can provide cruci-
ality ®gures and can also be used to provide cruciality
indices for aspects other than activity-durations (e.g.
resource availability-levels, or the choice at a stochastic
branch-point [48]).
Finally, one further purpose for which such network
simulations have shown themselves useful is in the
rational allocation of contingency to activities within a
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314 307
network, a perennial problem with large, dispersed,projects [55].
4. Problems with realism
Such simulations, while popular amongst risk ana-
lysts, have been slow to gain popularity with practisingproject managers. One issue that has arisen resultsfrom the nature of estimates populating such models.
There is clearly a relationship between setting PERTestimates and setting targets (e.g. Little®eld andRandolph [32] relate PERT with management byobjectives (MBO)). A natural consequence of this is
the observation that the actual duration is rarely lessthan the target. To some extent, this is due to the ideathat activity-durations tend to be skewed Ð some esti-
mate skew to be twice as much to the right of themode as to the left [20]. However, a major cause isobviously `Parkinson's e�ect': indeed, Gutierrez and
Kouvelis [24] attempt to de®ne some models in whichthe duration of an activity is dependent upon the allo-cated duration, including mathematical statements of`expect all activities to be late' and `it is the busiest
man who has time to spare'. This led Williams [50] todevelop a particular duration-distribution based onthese e�ects that has been found in some circumstances
to correspond well to managers' subjective judgementson duration-distributions.Besides this, it is still a feature of many network
simulation outputs, that the distributions are oftenvery wide. A report that gives a 95% con®dence inter-val to the project duration falling between (say) 2 and
8 years will simply be ignored. A key reason for thisproblem is that the simulations described above simplycarry through each simulation run in a dumb fashion.In reality, of course, management would take action to
control a project going out of control, but untilrecently simulations have not been able to take cogni-sance of such actions. This has, in the opinion of this
author, gravely weakened the credibility of the time-risk analyses used in common practice.A few authors have attempted to include such man-
agement actions within analytical models or complexmathematical models using simulation to reach a sol-ution. Key amongst these is Golenko-Ginzburg: [21]describes a line of work through GERT leading to
CAAN (controlled alternative activity networks),which attempts to make decisions to optimise the net-works, although the decisions do not depend on state
of project at that time. The author understands thatsuch methods are not widely used in practice sincethey have a high level of complexity while still not
incorporating su�cient generality with su�cient trans-parency for practitioner-acceptance (although see adefence of this area and subsequent discussion in [51]).
(This work is continued in, for example, Golenko-Ginsberg and Gonik [22]).
Therefore, work is ongoing to attempt to includemanagement actions within project simulations.Tollersrun [45] makes a ®rst step towards this with
Statoil's TOPPS system by allowing a degree of correc-tive action during the project, based on the currentoutcome; for him, a project is de®ned as a dynamic de-
cision process. Also from Norway is TerreMar'sDynRisk package described by Skogen andHuseby [43], which again attempts to include decisions
within the simulations.However, there is a key problem with such work. It
has indeed recognised the need to incorporate manage-ment control into the models and has gone a little way
in this direction. However, these are simple controlactions with simple results, whereas in practice theresulting e�ects of such actions are often not obvious:
there are multiple in¯uences that combine to producethe resulting e�ects (often counter-intuitive). Thesemultiple in¯uences are not taken into account, and
thus not modelled, in current network simulationmodels.How can we learn about, and thus try to model, the
e�ects that occur when management take actions tocontrol projects? One area of work, which has studiedsuch actions and their resulting e�ects, has used theidea of system dynamics.
5. System dynamics
This alternative approach has ignored discrete-eventsimulation and project networks altogether and con-
centrated on the dynamics within the management de-cision-making process, in particular using systemdynamics (SD). Such work was begun by authors par-
ticularly such as Brooks [9] and Cooper [12], and thereis now a considerable body of work using SD tomodel the temporal ¯ow of projects (see a review byRodrigues and Bowers [40]). The SD modelling tech-
nique has four important bene®ts.
. SD shows the compounding e�ect of combinationsof in¯uences (the `2 + 2 = 5 rule').
. It is able to explicate the e�ects of feed-back on theproject.
. It is helpful in modelling the soft e�ects involved in
project management (schedule pressure, motivation,etc.).
. SD's original design was to model the e�ects of the
information ¯ows; this helps to allow decision-rulesto be built in to try to mimic the actions that man-agement would take.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314308
These advantages of SD have enabled it to be used
to give interesting, and sometimes somewhat counter-
intuitive, insights into the behaviour of large projects.
This is particularly so when it has been used to expli-
cate past project behaviour in post-mortem
analysis [52]. In particular, it has been shown that
management actions, taken to try to achieve di�cult
schedules, or particularly to try to regain schedules
that have slipped, will often have much less e�ect than
expected, due to a number of interrelated e�ects, for
example:
. more workers have to be taken on than originally
planned, and/or extra shift-patterns and more over-
time have to be worked in order to keep to the sche-
dule; this exacerbates the feedback loops to cause
increasing ine�ciency; this well-known phenomenon
is commonly referred to as the $2000 hour, a term
originally coined by Cooper [14];
. the lack of system freeze, combined with a tight
time-constraint, forces management to work on pro-
ject-elements for which the surrounding system is
not yet frozen (which has to be reworked if there are
changes in the as-yet unfrozen surrounding system);
. such action has secondary e�ects, such as disincenti-
vising the design sta� as they work with unclear par-
ameters and knowing that their work may turn out
to be nugatory;
. when a concurrent manufacturing phase is con-
sidered, there are additional e�ects, both because de-
sign activities ®nish later and thus increase
concurrency, but also because items begin manufac-
ture and are then changed, which leads to retro®t,
degradation of manufacture learning and so on.
Cooper [15] gives a good overview of the modelling
of such e�ects in complex projects.
However, this SD modelling takes an overall, sys-
temic view of the whole project. It puts aside ideas
such as activities in a network and work breakdown
structures, and thus loses much of the operational in-
formation about the project. This means that, while it
has shown interesting insights in post mortem analyses,
and is able to give some useful guidelines before the
start of a project, it has been virtually ignored at the
operational project management level which monitors
and controls ongoing projects.
One answer to this is to try to combine the classical
simulation methods with SD and gain a management
structure in which each informs each other; this has
been tried on at least one real large project by
Rodrigues and Williams [41] and Rodrigues is further
developing formalised methodologies for its implemen-
tation. This is certainly a possible way forward,
although needing more work to operationalise the
method, and a lot of work to sell it to practising risk
managers. An alternative is to seek to gain lessonsfrom the SD work and apply it to the current network
simulation models, which can then be used in a moreintelligent way than at present.
6. A modi®ed network simulation
We have established that network simulations lackrealism and credibility because they lack modelling oftwo features:
. management actions in response to a late-runningproject;
. the e�ects resulting from such actions; clearly, sim-
plistically modelling the fact that management act tobring a late-running project into line would implythat projects all end up on time Ð however, the sys-
tem dynamics work provides some useful lessonsabout the e�ects of such actions, which could besimpli®ed and included in our network simulations.
It is not the aim of this paper to give an exhaustiveaccount of all possible management actions, nor amodel of the e�ects of these actions generally appli-
cable to all projects. The examples below aim to indi-cate the type of network simulation models that can bedeveloped drawing upon the lessons from systemdynamics.
There are two generic types of action which appearto be most important:
1. Bringing an activity forward. If activity i has a stan-dard ®nish±start dependency (or dependencies) asshown in Fig. 1a, this action would start activity iearlier than would normally be the case, resulting
(theoretically) in the dependency shown in Fig. 1b.2. Reducing an activity's duration: thus if activity i
has duration ti, this action would change its dur-
ation to lti (where l is a crashing proportion, suchthat 0 < l< 1). Of course, this idea of crashing ac-tivity durations is well-known and there is a fair
amount of literature on ®nding the optimum projectduration when each of a subset of activities inde-pendently has crashing potential. However, this is ahypothetical situation and it is not clear to what
extent the ideas are used in practice;Kamburowski [27] described an optimum algorithmrecently, but in the subsequent discussion
Chapman [11] comments that work on the proper-ties of exact project scheduling network models`is . . . of little practical interest'. One reason for this
is because the down-stream e�ects of crashing oneactivity are not included in the model; we can lookto SD to gain insight into what these e�ects are.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314 309
What are the e�ects of such management actions?
Typical papers that discuss these e�ects include the fol-lowing.
1. Rodrigues and Williams [42] discuss (inter al) the
e�ects of reducing an activity's length (based onwork in the software development domain), show-
ing that it:
Ðleads to more downstream problems due toquality problems (errors in code)
Ðleads to more problems in parallel activities due
to increased parallelism between activities.
2. Cooper [13] similarly uses quality as a criticalmetric, being de®ned as the fraction of work being
executed that will not require subsequent rework.
3. Williams et al. [54] describe a series of interlinkedfeedback loops (based in a manufacturing domain)
where bringing activities earlier (thus doing workbefore engineering judgement said it should be
done) increased rework, which increased delays,which led to more parallelism and thus increased
cross-relations between activities, hindering a systemfreeze and hence more work done before it should
be done.
4. Graves [23] discusses the e�ect of compressingR&D project times, giving a time±cost curve
suggesting that if duration is reduced to lt (where0 < l < 1) then cost becomes divided by l k where
typical ®gures of k ranging from 0.9 to 2 are given.
5. Cooper [14] (referring to Brooks [9]) discusses thefeedback loops involved when extra personnel are
brought in to reduce activity-times, leading again toincreased durations and increased costs.
6. Abdel-Hamid and Madnick [1] give full details ofsome SD models designed to investigate such
e�ects.
How should such indications be modelled in our net-
work simulations? Looking at the two managementactions above, simple models could include the follow-ing, where activity i has nominal duration ti and costci:
1. Bringing activity i forward by t (where t is a timesuch that 0 < t < ti). This will have the e�ect that
Ðeven being optimistic, the remaining activity-dur-ation after the dependency, theoretically tiÿt,will in fact be a ®gure greater than that, say
tiÿat, where a is a factor such that 0 < a < 1;Ðthe new cost, rather than ci will be ci (1ÿ b(t))
where b is a function of t such that 0 < b(t) < 1;
Ðboth parallel and down-stream activities willhave their nominal durations increased from tjto tj(1 + d(j, t)), where d(j, t) is a nonnegativefunction of t, that decreases exponentially as ac-
tivity j moves towards the end of the project.
2. Reducing an activity's duration from ti to lti.3. the work described above gives a crashing time±cost
function more realistic than a simple linear func-tion, where the cost is divided by l k;
4. parallel activities j will be a�ected: that is, their dur-
ations will become tj(1 + g(l)) (where g(l) is a non-negative function of l);
5. down-stream activities will have their nominal dur-
ations increased from tj to tj(1 + e(j,l)), whereagain e(j,t) is a nonnegative function of l thatdecreases exponentially as activity j moves towards
the end of the project.
Furthermore, these secondarily lengthened activities
will in general themselves be crashed, to stop thembecoming critical, so their costs will increase, and thee�ects described above will be repeated in a secondary
wave of knock-on e�ects.
7. Example
Modelling of these e�ects has to depend upon theproject being modelled. Given here is the very simplest
possible example, to indicate the type of modelimplied, and the type of result. Suppose that a projecthas an important mid-point mile-stone; we will com-
bine together all of the activities before this point intoa single activity A. Suppose further that there is a sub-sequent activity B that management can control. We
Fig. 1. (a) Standard ®nish±start dependency. (b) Finish±start
dependency where management takes the decision to start ac-
tivity i early.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314310
will combine together all of the other activities into
two: activity C, which represents activities parallel to B
and activity D, representing down-stream activities.
The resulting simple network is shown in Fig. 2.
Suppose the activity-durations have distributions as
shown in Table 1 (where the unit is months).
Assuming the activities are independent, this is
obviously simple to simulate. Suppose that the ®nal
date for completion of the project (beyond which liqui-
dated damages are payable) is 35 months; simulation
shows a 90.2% probability of achieving this.
But suppose that the milestone by which A is sup-
posed to be ®nished is at month 23, and management
knows that if this slips, the ®nal target date of 35
months is unlikely to be met. If activity A slipped then
management would take two actions:
. if activity A completed after month 23, part of ac-
tivity B would be brought forward, so that two
months work would be done in parallel to A. As
always, Fig. 2 represents the planned activities, but
management usually has the option of bringing for-
ward part of an activity if the project is running
late, for example setting engineers to work on de-
signs even though the concept has not been ®nalised,
or taking o� detailed designs that are not fully com-
pleted (so not really ready for release) and beginning
construction of the basis of them. Practising project
personnel will recognise these occurrences from their
own experience. However, due to the e�ects dis-
cussed above:
Ðthe remaining duration of activity B would be
reduced only by 1.5 months (i.e. a= 0.75);
Ðthe cost of activity B would be increased by 6.5/6;
Ðactivity C (parallel to B) would be increased by0.5 months;
Ðactivity D(downstream from B) would beincreased by 0.5 months.
. if activity A completed after month 24, activity B(that element which was not brought forward tobefore the milestone) would be crashed, so that itslength would be reduced by 20%. However, again:
Ðthe cost would be increased by 25% (i.e. k= 1)Ðthe duration of activity C would be divided by 0.9Ðthe duration of activity D would be divided by
0.95.
This network has been simulated 1000 times using
the @Risk package [36], both the initial version (withno management control) and the latter version. Figure3 shows the probability of the total project duration
exceeding certain times for both the crude simulationand this latter controlled version; in particular, thecontrolled version now has a 95.2% chance of achiev-
ing the deadline. Furthermore, the cruciality ofActivity A (i.e. the correlation between its durationand the duration of the whole project) is reduced from
95 to 93% in going from the crude to the controlledversion, showing that management can take someaction to reduce the e�ect of activity overrun, but can-not simply catch up completely (implying a determinis-
tic project duration and thus a cruciality index ofzero).Thus we have shown:
. the e�ects of management control, restricting thelonger project durations;
. the resulting project duration which is less optimistic
than might be ®rst thought of by management, sincethe time saved by bringing an activity forward orcrashing it is lost to a certain degree by the e�ects of
the project dynamics;. additional costs: since the ®rst of the above actionsoccurred in 14.9% of iterations, and the second in
8.7% of iterations, the cost of activity B wasincreased by 3.2%.
8. Conclusion
The use of networks handling uncertainty to providea temporal risk analysis of projects is now widespread.However, the results of such analyses are frequently
not regarded as credible, partly because of the resultingwide probability distributions; this is because themodels do not re¯ect the actions that management
Fig. 2. Example network.
Table 1
Activity Distribution Minimum Mean Maximum
A beta 15 20 30
B beta 5 6 10
C beta 2 36
D beta 4 5 8
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314 311
would take to bring late-running projects under con-
trol. These are di�cult to include in models, notbecause the actions themselves are complex, but ratherbecause the e�ects of those actions are not well-under-
stood. However, the work in system dynamics model-ling of projects can give some indications of whatthose e�ects would be. This paper has attempted to
describe these indications and give some small illustra-tive models of the e�ects. It is hoped that the use ofsuch modelling can help to bring additional realism to
probabilistic network modelling.
References
[1] Abdel-Hamid T, Madnick SE. Software project
dynamics: an integrated approach. NJ: Prentice-Hall,
1991.
[2] Adlakha VG. An improved conditional Monte Carlo
technique for the stochastic shortest path problem.
Management Science 1986;32:1360±7.
[3] Adlakha VG, Kulkarni VG. A classi®ed bibliography of
research on stochastic PERT networks: 1966±1987.
INFOR 1989;27:272±96.
[4] Adlington PS, Christensen PJ. Risk analysis: realism of
just mathematics? In: Proceedings of the Internet '94
12th World Congress, Oslo, vol. 2. Zurich: International
Project Management Association, 1994:104±11.
[5] Anklesaria KP, Drezner Z. A multivariate approach to
estimating the completion time for PERT networks.
Journal of the Operational Research Society
1986;37:811±5.
[6] Avramidis AN, Bauer KW, Wilson JR. Simulation of
stochastic activity networks using path control variates.
Naval Research Logistics 1991;38:183±201.
[7] Berny J, Townsend PRF. Macrosimulation of project
risks: a practical way forward. International Journal of
Project Management 1993;11:201±8.
[8] Bowers JA. Criticality in resource constrained networks.
Journal of the Operational Research Society 1995;46:80±
91.
[9] Brooks F.J. The mythical man-month. Chapel Hill, NC:
University of North Carolina, 1975.
[10] Chapman CB. Risk analysis, a view from 1990. I.M.A.
Journal of Mathematics. Applications in Business and
Industry 1990;2:79±95.
[11] Chapman CB. On the minimum cost project schedule:
some comments. OMEGA 1995;23:467±8.
[12] Cooper KG. Naval ship production: a claim settled and
a framework built. Interfaces 1980;10:20±36.
Fig. 3. Distribution of project duration.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314312
[13] Cooper KG. The rework cycle: benchmarks for the pro-
ject manager. Project Management Journal 1993;24(1):0.
[14] Cooper KG. The US$2000 hour: how managers in¯uence
project performance through the rework cycle. Project
Management Journal 1994;25:11±24.
[15] Cooper KG. System dynamics methods in complex pro-
ject management. In: Williams TM, editor. Managing
and modelling complex projects. NATO ASI Series.
Dordrecht, Netherlands: Kluwer Academic Publishers,
1997.
[16] Devroye LP. Inequalities for the completion times of sto-
chastic PERT networks. Mathematics of Operational
Research 1979;4:441±7.
[17] Dodin B. Bounding the project completion time distri-
bution in PERT networks. Operations Research
1985;33:862±81.
[18] Dodin B. Minimum number of arcs in conditional
Monte Carlo sampling of stochastic networks. INFOR
1986;24:33±44.
[19] Dodin BM, Elmaghraby SE. Approximating the critical-
ity indices of the activities in PERT networks.
Management Science 1985;31:207±23.
[20] Golenko-Ginzburg D. On the distribution of activity
time in PERT. Journal of the Operational Research
Society 1988;39:767±71.
[21] Golenko-Ginzburg D. Controlled alternative activity net-
works for project management. European Journal of
Operational Research 1988;37:336±46.
[22] Golenko-Ginzburg D, Gonik A. On-line control model
for cost-simulation network models. Journal of the
Operational Research Society 1996;47:266±83.
[23] Graves SB. Why costs increase when projects accelerate.
Research and Technology Management 1989;0:16±18.
[24] Gutierrez GJ, Kouvelis P. Parkinson's law and its impli-
cations for project management. Management Science
1991;37:990±1001 (Erratum in Management Science,
1991;37:1507).
[25] Kamburowski J. An upper bound on the expected delay
of completion time of PERT networks. European
Journal of Operational Research 1985;21:206±12.
[26] Kamburowski J. Normally distributed activity durations
in PERT networks. Journal of the Operational Research
Society 1986;36:1051±7.
[27] Kamburowski J. On the minimum cost project schedule.
OMEGA 1995;23:463±5.
[28] Kidd J. A comparison between the VERT program and
other methods of project duration estimation. OMEGA
1987;15:129±34.
[29] Kidd J. Do today's projects need powerful network plan-
ning tools?. International Journal of Production Research
1991;29:1969±78.
[30] Kulkarni VG, Adlakha VG. Markov and Markov-regen-
erative PERT networks. Operations Research
1986;34:769±81.
[31] Kulkarni VG, Provan JS. An improved implementation
of conditional Monte Carlo estimation of path lengths in
stochastic nets. Operations Research 1985;33:1389±93.
[32] Little®eld TK, Randolph PH. PERT duration times:
mathematics or MBO. Interfaces 1991;21:92±5.
[33] Mehrotra K, Chai J, Pillutla S. A study of approximat-
ing the moments of the job completion time PERT net-
works. Journal of Operations Management 1996;14:277±
89.
[34] Moder J. Network techniques in project management.
In: Cleland DI, King WR, editors. Project management
handbook, 2nd ed. New York: Van Nostrand Reinhold,
1988:324±73.
[35] Nicolo E. Metaproject analysis: a systemic methodologi-
cal aid for strategic planning. In: Proceedings of the 11th
Internet World Congress, Project management without
boundaries Florence (Italy) June, vol 2. Zurich:
International Project Management Association,
1992:227±36.
[36] Palisade Corp. @Risk is produced by Palisade
Corporation, New®eld, NY, US, 1997.
[37] Primavera Systems Inc. MonteCarlo1 is a registered tra-
demark of Primavera Systems Inc, Bala Cynwyd, PA,
US, 1995.
[38] Ragsdale C. The current state of network simulation in
project management theory and practice. OMEGA
1989;17:21±5.
[39] Ritchie E. Network base planning techniques: a critical
review of published development. In: Rand GK, Eglise
RW, editors. Further developments in operational
research. Oxford, UK: Pergamon Press, 1985.
[40] Rodrigues A, Bowers J. The role of system dynamics in
project management. International Journal of Project
Management 1996;14:213±20.
[41] Rodrigues A, Williams TM. Systems dynamics in soft-
ware project management: towards the development of a
formal integrated framework. European Journal of
Information Systems 1997;6:51±66.
[42] Rodrigues A, Williams TM. Systems dynamics in project
management: assessing the impacts of client behaviour in
project performance. Journal of the Operational
Research Society 1998;49:2±15.
[43] Skogen S, Huseby AB. Dynamic risk analysis: the
DynRisk concept. In: Proceedings of the 11th Internet
World Congress, Project management without bound-
aries Florence (Italy), June, vol 2. Zurich: International
Project Management Association, 1992:511±20.
[44] Sullivan RS, Hayya JC, Schaul R. E�ciency of the anti-
thetic variable technique for simulating stochastic net-
works. Management Science 1982;28:563±72.
[45] Tollersrun PK. TOPP: Total project planning. In:
Proceedings of the Internet '94 12th World Congress,
Oslo, vol. 2. Zurich: International Project Management
Association, 1994:396±403.
[46] Williams TM. Project risk management. In: MILCOMP
89 Conference Proceedings. UK: Microwave Publications
Ltd., 1989:107±13.
[47] Williams TM. Risk analysis using an embedded CPA
package. International Journal of Project Management
1990;8:84±8.
[48] Williams TM. Criticality in probabilistic network analy-
sis. Journal of the Operational Research Society
1992;43:353±7.
[49] Williams TM. Managing risk in development and initial
production. International Journal of Production
Research 1994;32:1591±7.
[50] Williams TM. What are PERT estimates?. Journal of the
Operational Research Society 1995;46:1498±504.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314 313
[51] Williams TM, editor. Managing and modelling complex
projects. NATO ASI Series. Dordrecht, Netherlands:
Kluwer Academic Publishers, 1997. ISBN 0-7923-4844-3.
[52] Williams TM. The risk of safety regulation changes in
transport development projects. In: Kahkonen K, Artto
KA, editors. Managing risks in projects. London: E&FN
Spon, 1997:284±293. ISBN 0-419-22990-6.
[53] Williams TM. Allocation of contingency in activity dur-
ation networks. Construction Management and
Economics 1998, submitted for publication.
[54] Williams TM, Eden CL, Ackermann FR, Tait A. The
e�ects of design changes and delays on project costs.
Journal of the Operational Research Society
1995;46:809±18.
T. Williams / Omega, Int. J. Mgmt. Sci. 27 (1999) 305±314314