See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/279531734

An Extended Critical Path Method for Complex Assembly Lines

Conference Paper · January 2012

CITATIONS READS
4 888

3 authors, including:

Evangelos Angelidis Oliver Rose

Universität der Bundeswehr München Universität der Bundeswehr München
10 PUBLICATIONS   42 CITATIONS    195 PUBLICATIONS   3,230 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

On the Dynamism in Dynamic Vehicle Routing Problems View project

Simulaton based Optimization Platform View project

All content following this page was uploaded by Oliver Rose on 26 July 2015.

The user has requested enhancement of the downloaded file.

Proceedings of the 2012 Industrial and Systems Engineering Research Conference
G. Lim and J.W. Herrmann, eds.

An Extended Critical Path Method for Complex Assembly

Lines
Evangelos Angelidis
Oliver Rose
Department of Computer Science
University of the Federal Armed Forces Munich, 85577 Neubiberg, Germany
André Naumann
Department of Computer Science
Dresden University of Technology, 01062 Dresden, Germany

Abstract

Our paper delves into the scheduling of complex assembly lines with workforce constraints. Formally it is defined
as a Multi-Mode Resource-constrained Multi-Project Scheduling Problem with splitting activities. One of the main
goals is to determine the maximum lateness for all projects leading to reduced production costs, reduced stock costs,
on-time delivery, etc. In the literature almost all algorithms for such problems are based on the calculation of the
critical path. This method calculates the path of all jobs’ activities contributing to the cycle time of each project.
Unfortunately a critical path cannot be computed in our case because every project has its own isolated project network
(with many possible graphs) including complex time, resource and precedence constraints. Furthermore the amount of
the resources and their availability are limited. As a consequence the projects inhibit one another because they share
the same system resources. In this paper we present an extended critical path method using a simulation approach
to calculate probable critical paths of the projects. The method takes into account the complex relationships of the
projects and it helps to minimize the choice of extended critical path candidates.

Keywords
complex assembly lines, MMRCMPSP with splitting activities, simulation-based optimization, critical path method

1. Introduction
In this paper we focus on solution strategies of scheduling problems of real world scenarios for complex assembly
lines. The companies need to decrease production costs more than ever and deliver on time for surviving on the global
market, especially in assembly lines with workforce planning. The typical products, we base our research upon, are
small series or even unique items, for example, turbines, planes and industrial machines. A scenario consists of various
products with different production plans (also referred by some authors as projects and project networks), precedence
constraints and thousands of activities. The activities and products have many time-bound requirements. The activities
do not have a fixed processing time and are dependent on the amount of the resources assigned to it. Additionally the
manufacture has limited resources with individual scheduling and finite buffers. The scheduling of such assembly line
productions is a very complex task, where even experienced foremen cannot always deliver suitable solutions.
The scheduling of such assembly lines is a large combinatorial optimization problem and is often mentioned in the
literature as a Multi-Mode Resource-Constrained Multi-Project Scheduling Problem (MMRCMPSP) with activity
splitting. Unfortunately this problem is NP-hard and cannot be solved with classic approaches. The research in
this area is mainly based on small or medium size problems, which are often solved by exact approaches or genetic
algorithms. These strategies are not suitable for our large problems and especially with respect to short runtimes. A
detailed explanation of this scheduling domain and its possible solution strategies are given in Section three of [3].
An alternative way to find a good solution in short runtime is the simulation-based optimization approach as it is
described in [3]. The basic idea of this approach is to simulate a scenario, analyze it and create a new one according
to the simulated scenarios. The goal of these iterations is to generate a feasible solution within only a few steps.
 Angelidis, Naumann and Rose
Almost all algorithms for this domain are based on the calculation of the critical path (CP). This method computes the
path of all jobs’ activities contributing to the cycle time of each product. As a result of goal-driven modification of the
critical path of the production plans, the algorithm controls production costs, reduced stock costs, on-time delivery,
etc. Unfortunately a critical path, as it is usually defined, cannot be computed in our case because every product
has its own isolated production plan (with many possible ways of production) including complex time, resource and
precedence constraints. Furthermore the amount of the resources and their availability are limited. As a consequence,
the products inhibit one another because they share the same system resources. In this paper we present a method to
compute a critical path using the simulation results to support these algorithms.
In Section 2 we will further discuss simulation-based scheduling. Section 3 is devoted to the presentation of an
extended critical path method. Section 4 provides a short overview of the experiments and the results. In Section 5 we
draw some conclusions and discuss future research.

2. Simulation-Based Scheduling Approach

The scheduling of our domain of interest needs solution strategies, which can handle huge search spaces and find a
suitable scenario in short runtimes. A framework to schedule complex assembly lines is described in [3], using the
simulation-based scheduling approach. In our research we use its data meta model. The basic idea is illustrated in
Figure 1. In the first step a real system is transformed to a MMRCMPSP model. The MMRCMPSP model describes
the manufacturing facility and contains all the possible scenarios of this assembly line production. The optimizer,
according to the solution strategy and the already simulated scenarios (SIMLOGs), creates new ones to increase the
current solution rating. The scenarios are classified as a Single-Mode Resource-Constrained Multi-Project Scheduling
Problem (SMRCMPSP).
Essentially the simulation model contains all the production plans with adjusted earliest start dates and latest end dates.
These are defined as activity-on-arc (AOA) networks with alternative production paths. The activities have adjusted
earliest start dates and latest end dates and a chosen mode with a current number of assigned resources. The resources
in the model own a subset of their scheduling plan. The model has a list of dispatching rules and other settings to run
the model according to the needs of the optimization. A more detailed explanation of the simulation model can be
found in [1].

MMRCMPSP

Real system Optimization Results

SMRCMPSP

SIM LOG

Simulation

Figure 1: Principle of the Simulation-Based Scheduling

2.1 Optimization Strategies and the Role of the Critical Path

The large space of the numerous parameters of the combinatorial problem have a big effect on the optimization
decision. Even few changes of the parameters can effect a big variation on the results, while the intention is to produce
many items at the same time, using limited resources with unexpected events, complex dispatching rules, etc. One
useful tool to weight the behavior of the activities over the costs and the time of the production, is the calculation of
the critical path.
 Angelidis, Naumann and Rose
Since the first presentation of the critical path in 1959 [5], it is a standard method in production planning and it
is used in many methods and applications. The classic Critical Path Method (CPM) we use, is defined in [4]. It
calculates the earliest possible completion time of a production plan and finds all activities, where a possible delay of
their start increases the makespan. For basic scheduling problems with parallel machines and unlimited resources the
computation is possible and feasible in short runtime. On the other hand for realistic scheduling problems, this is very
difficult. An interesting algorithm to minimize the time between the product’s completion time and the delivery time
is in [2], which it is based on the simple critical path method.

3. An Extended Critical Path Method

For our model description this theoretical critical path method, based only on a production plan, is not enough. In
the literature we found, various approaches of CPMs for similar problem domains. These CPMs and the theoretical
CPM are not applicable for the optimization of complex assembly lines. They consider production plans as indepen-
dent entities, without taking into account their interference (e.g., products competing for resources). A critical path
determined by these methods, might not represent the true criticality of the activities on it. We need a CPM for each
product, which involves the products’ interactions, the scheduling of the resources, the executed activities of the ac-
tivity splitting plan, the possible time gaps between the activities (through resource limitation and chosen dispatching
rules), etc.
In this Section we present a solution of an extended critical path method, which is based on the production plan as
well as on the simulation result of the scenario. For a better understanding we first define some terms as follows:
• An end activity is an activity, which has no successor activities.
• A possible critical end activity is the latest end activity or an end activity within a certain neighborhood around
the latest end activity. The size of this neighborhood can be adjusted in the algorithm’s settings.
• A not critical end activity is an end activity, which finished too early to have a significant influence on the
makespan.
• A path is set of linked activities within a production plan.
• A path statistic object (PSO) represents a path in our CPM, containing a set of activities and further statistics,
like the size of the activity set, the sum of all buffer times and operation times of all activities on the path. We
need this information to compare various paths.
• A buffer time of an activity is the timespan between its completion date and the start date of its earliest starting
successor activity.
The following pseudocode (Algorithm 1) is providing a rough overview of the algorithm’s structure. Within the next
paragraphs we will give more details on our CPM, its customization possibilities, and their effects.
The critical path method introduced in [4] consists of two procedures. A forward procedure determines the makespan
of the product and a backward procedure determines the critical path. In our simulation-based scheduling the makespan
is already known through the simulation, so the algorithm starts with the backward procedure upon a reduced produc-
tion plan. All not executed alternative activities are excluded from analysis (Figure 2). This restricts the search area
and accelerates the algorithm.
As a first step the algorithm divides all end activities into possible critical and not critical end activities. End activities,
which are completed long time before the completion time of the product, probably have little to no influence on
the product’s completion time. The algorithm considers these as not critical end activities. In addition, all activities
which are on a path which exclusively leads to a not critical end activity can be suspected as not critical activities.
Therefore the algorithm marks all paths exclusively leading to a not critical end activity and excludes them from
further investigation (Figure 3).
The algorithm starts the search at the remaining possible critical end activities. It creates a PSO for each of these
activities and begins to analyze these paths toward the start activities. Every time it reaches an activity, it adds the
activity to the current PSO and updates its statistics. In addition, at every reached node, the method executes several
checks.
• If the node links to only one not marked predecessor and to only one not marked successor, then it simply
continues with the predecessor activity.
 Angelidis, Naumann and Rose

Algorithm 1: extended critical path method

input : A simulated production plan Pl
output: A set of probable critical paths of the Pl
1 X ← set of possible critical end activities;
2 Y ← set of not critical end activities;
3 PossCPs ← 0/ //A set of PSOs;
4 for x ∈ X do
5 p ← a new PSO with one element x;
6 PossCPs.append(p);
7 for y ∈ Y do
8 mark all predecessor activities which are exclusively leading to y;
9 while PossCPs 6= 0/ do
10 for p ∈ PossCPs do
11 n ← predecessor node of the latest added activity of p;
12 if n has more than one not marked outgoing link then
13 nList ← list of outgoing PSOs of n;
14 nList.append(p);
15 if all outgoing PSOs ∈ nList then
16 remove all PSOs in nList from PossCPs;
17 compare PSOs using pathComparison(nList);
18 append returned PSOs to PossCPs;
19 mark all successors of n;
20 continue;
21 continue;
22 if n has no incoming link then
23 PossCPs.remove(p);
24 finalPossCPs.append(p);
25 else
26 if n has exact one not marked incoming link then
27 a ← predecessor activity of n;
28 update p with stats from a;
29 else
30 acts ← all predecessor activities of n;
31 for a ∈ acts do
32 clonedp ← clone of p;
33 update clonedp with stats from a;
34 PossCPs.append(clonedp);
35 PossCPs.remove(p);

36 criticalPathSet ← result of finalPathComparison(finalPossCPs);

 Angelidis, Naumann and Rose

ΛΙΛ ΛΙΛ ΛΙΛ

d f j

ΛΙΛ Λ|V VΙΛ

a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ V|Λ
b h i l o

ΛΙV V|Λ
m p

Figure 2: A simulated production plan, dotted lines represent not executed alternative activities and the nodes’ nota-
tions (∧/∨) represent a parallel/alternative way

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 3: Not critical path between nodes (c-d-f-j) and possible critical end activities between nodes (m-p) and (n-q)

• If the node links to more than one not marked predecessor, it duplicates the PSO for each not marked predecessor
activity and continues to proceed on these new paths (Figure 4 ).

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 4: At node n, a PSO is duplicated to investigate two incoming paths

• If the node links to more than one not marked successor, it puts the PSO on a waiting-to-compare-list and checks,
if all other paths’ PSOs, which are joining at this node are on this list. If so, it compares all PSO on this list
using our Path Comparison Method and continues to proceed all paths, which are returned (Figure 5, Figure 6 ).
If not, the paths are not yet ready for a comparison and the method proceeds other PSOs instead.
• If the node does not have any predecessor activities, the search reached a start activity. In this case the algorithm
adds the PSO to a list of possible final CP candidates (Figure 7).
By the time all paths are finished, it compares these final CP candidates to figure out, which is the CP for this product
(Figure 8).
 Angelidis, Naumann and Rose

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 5: At node i, two PSOs are compared

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 6: At node e, two PSOs are compared. The result of the previous comparison is shown in node i

3.1 Generic Comparison Methods

We recognized at the beginning of the CP research that a suitable method must support extensions according to the
production environment, the optimization goals, and the processing power. So every comparison implements a given
interface to support customization. We implemented comparison methods for all the three types of comparisons: end
activities, paths, and final paths.
• End-Activity-Comparison method
The method returns a subset of all end activities that could be part of critical paths. We implemented three
different strategies:
◦ all end activities are taken into account
◦ only the latest end activity is taken into account
◦ all end activities within a certain time frame before the latest end activity finished are taken into account.
If all end activities or a bigger time frame are chosen, the algorithm analyzes a larger part of the production

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 7: Analyzed production plan, with two paths waiting for final comparing
 Angelidis, Naumann and Rose

ΛΙΛ X ΛΙΛ X ΛΙΛ

d f j

X
ΛΙΛ Λ|V VΙΛ
a g k
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
c e n q
ΛΙΛ ΛΙΛ ΛΙΛ ΛΙΛ
b h i l

ΛΙV V|Λ
m p

Figure 8: Determined critical path

plan. Therefore it will need more iteration steps and comparison decisions, so the CPM has a longer runtime.
However, the result of the CPM may be more accurate, because there is a lower probability of falsely excluded
end activities.
If only the latest end activity or a small time frame are chosen, there is a smaller part of the plan to search
through. This means the CPM has less iteration steps and less comparison decisions. On the other hand, there
is a higher probability of falsely excluded end activities, which means the CPM may be less accurate.
• Path-Comparison Method
The method for path comparison returns the PSOs, which are used for proceeding the CPM. We evaluate paths
by their statistics, which are collected during the analysis. There are five comparison strategies:
◦ highest number of activities on the path
◦ highest sum of operation times of all activities on the path
◦ lowest sum of buffer time between all activities on the path
◦ highest ratio of sum of operation times and sum of buffer time on the path
◦ combined criteria.
For each chosen strategy, we have to decide whether it continues just with the best path or with all paths within
a certain range below or, in case of buffer time, above the best path. The combined option searches for the best
paths, considering all given statistics.
The question which comparison criterion is the best depends on the production environment. While the buffer-
time-criterion is close to the classic CPM, the number-of-activities-criterion can be relevant for products with
many setup times or the sum-of-operation-time-criterion can be important in an environment where long opera-
tion times are associated with crucial activities.
The options whether to choose one or more paths to continue with, have similar effects on the analysis like in the
comparison of end activities. The more restrictive the parameters are, the faster, but less accurate is the CPM.
More loose parameters mean, slower analysis, but better results.
• Final-Path-Comparison Method
The Final Path Comparison Method is the last step to the CP. It is always used, if there is more than one start
activity or if any path comparison returned more than one path. The comparison of final paths uses the same
set of comparison strategies as the path comparison, but it can be parameterized separately. Depending on the
chosen parameters the algorithm may find more than one critical path. These critical paths are similar in regard
to the specific comparing strategy.
We believe that multiple critical paths are necessary for our optimization work. The CPM analyzes a specific
production plan for a specific simulation run. We need this information to set up the next optimization iteration.
A minor change in mode selection or prioritization in a single product, may change all simulated production
plans and their critical paths. As a result the previous calculated CP would be useless. Therefore we rely on a
set of critical paths, which will probably be relevant for the product’s completion time, even if there are changes
in the current production plan.
 Angelidis, Naumann and Rose
4. Experiments and Results
As a first step, to test the extended Critical Path Method, we created generic scenarios with up to dozens of products
and thousands of activities, where we successfully observed the correctness of the method by using several solution
strategies. Based on earlier research, we realized the importance of the time factor in simulation-based optimization,
therefore we analyzed the algorithm’s runtime. The eCPM has linear complexity with respect to the numbers of
activities of the production plan. In all test scenarios we observed, that the CPM’s runtime was negligible compared
to the time needed for the actual simulation itself or to the average runtime of one step of the simulation-based
optimization.
In addition we used the eCPM on a new implementation of the algorithms and the manufacture models presented in [2].
We noticed that the algorithm comes to better results with even fewer steps, particularly in production environments
with a small amount of highly utilized resources.

5. Conclusion and Future Work

We are confident that the eCPM is a better tool for our problem domain compared to classic CPM. Especially in en-
vironments with negative slack, it is evident that optimization algorithms benefit from a reliable critical path method.
A well-determined critical path allows to accelerate only activities, which are directly responsible for the product’s
extended completion time. In addition, the resource consumption is kept to a moderate level, because just critical ac-
tivities are speeded up. Furthermore, our eCPM offers an opportunity for a more precise adjustment of the completion
time of the product.
The focus of future research lays in the comparison strategies. We want to improve our eCPM according to the
factory environment it is used in. Therefore we need to develop more comparison strategies, which will allow a better
evaluation of paths according to the requirements of the environments. A second field of future research is the analysis
of the optimal configuration of comparison strategies (like size of the time frames, etc.) for a given factory scenario.

6. Acknowledgement
We would like to thank Daniel Bohn for his suggestions and discussions during the project.

References
[1] Angelidis E., Pappert, M. F., and Rose, O., 2011, “A prototype simulation tool for a framework for simulation-
based optimization of assembly lines," Proc. of the Winter Simulation Conference, December 11-14, Phoenix,
Arizona, 2383-2394.
[2] Noack, D., and Rose, O., 2008, “A simulation based optimization algorithm for slack reduction and workforce
scheduling," Proc. of the Winter Simulation Conference, December 5-8, Miami, Florida, 1989-1994.
[3] Pappert, M. F., Angelidis E., and Rose, O., 2010, “Framework for simulation based scheduling of assembly lines,"
Proc. of the Winter Simulation Conference, December 5-8, Baltimore, Maryland, 1690-1698.
[4] Pinedo L. M., 2005, Planning and Scheduling in Manufacturing and Services, 2rd Edition, Springer Series in
Operations Research, New York.
[5] Walker, M. R., and Sayer S. J., 1959, “Project Planning and Scheduling," Report, E.I. du Pont de Nemours & Co.,
Inc., Wilmington, Del.

View publication stats