Document downloaded from:

http://hdl.handle.net/10251/36586

This paper must be cited as:

Ponz Tienda, JL.; Yepes Piqueras, V.; Pellicer Armiñana, E.; Moreno Flores, J. (2013). The
Resource Leveling Problem with multiple resources using an adaptive genetic algorithm.
Automation in Construction. 29(1):161-172. doi:10.1016/j.autcon.2012.10.003.

The final publication is available at

http://dx.doi.org/10.1016/j.autcon.2012.10.003

Copyright
Elsevier
Title Page

THE RESOURCE LEVELING PROBLEM WITH MULTIPLE

RESOURCES USING AN ADAPTIVE GENETIC ALGORITHM

Ponz-Tienda, J.L.

Yepes, V.

Pellicer, E.

Moreno-Flores, J.
*Manuscript
Click here to view linked References

THE RESOURCE LEVELING PROBLEM WITH MULTIPLE RESOURCES USING

AN ADAPTIVE GENETIC ALGORITHM

Abstract
Resources management ensures that a project is completed on time and at cost, and that its
quality is as previously defined; nevertheless, resources are scarce and their use in the activities
of the project leads to conflicts in the schedule. Resource leveling problems consider how to
make the resource consumption as efficient as possible. This paper presents a new Adaptive
Genetic Algorithm for the Resource Leveling Problem with multiple resources, and its novelty
lies in using the Weibull distribution to establish an estimation of the global optimum as a
termination condition. The extension of the project deadline with a penalty is allowed, avoiding
the increase in the project criticality punishing the shift of activities. The algorithm is tested
with the standard Project Scheduling Problem Library PSPLIB, and a complete analysis and
benchmarking test instances are presented. The proposed algorithm is implemented using VBA
for Excel 2010 in order to provide a flexible and powerful decision support system that enables
practitioners to choose between different feasible solutions to a problem, and in addition it is
easily adjustable to the constraints and particular needs of each project in realistic environments.

Keywords: Project Scheduling, Resource leveling, Genetic algorithms, Benchmarking

1. Introduction
Project management is the process of the coordination and integration of activities in an
efficient and effective manner using limited resources. It consists of linking resources to their
respective deliverables and assembling them into the whole project [1]. Resource management
is an intrinsic element of project management [2,3,4]; resource management ensures that the
project is completed on time and at cost and that the quality is as previously defined [5,6,7].
This is even more necessary for project-based companies such as contractors [3,8,9]. In fact,
project scheduling problems are one of most important problems that practitioners deal with in
scheduling, especially when they need to achieve the most efficient resource consumption
without increasing the prescribed makespan of the project.
However, because resources are scarce, the use of resources in the activities of the project
leads to conflicts in the schedule [10]. Project scheduling problems comprise not only resource-
constrained problems but also resource leveling problems, among others [11]. These two kinds
of problem consider resource consumption in two different ways: in the former it is seen as a
constraint, and in the latter the problem is to make it as efficient as possible. Even though these
two approaches may seem similar, they are conceptually different. Both have been widely
studied by researchers and applied by practitioners, although these two groups are unaware of
the differences between the approaches and the serious limitations imposed by the heuristics
used in the commercial software.
These two problems are defined as non-deterministic polynomial-time hard (NP-hard)
problems [12]. The first approach is a regular problem known as the Resource Constrained
Project Scheduling Problem; its objective is to reduce the makespan without exceeding the
constraints of resource availability [13,12]. The second, known as the Resource Leveling
Problem (from now on, RLP) is a non-regular problem; its objective is to achieve the most
efficient resource consumption without increasing the prescribed makespan of the project
[14,12]. The two problems can be combined together as a multi-objective optimization problem,
but there is always one main objective (usually the makespan); the other objective (usually the
efficient resource consumption) is secondary.
Nevertheless, conventional analytical and heuristic methods are neither flexible nor
productive when solving the RLP [15]. Some reasons for this inefficiency are, on the one hand,
that exact procedures simplify the real problems so are not useful at offering optimal solutions
with acceptable computational effort [16] and, on the other hand, that heuristics offer solutions
which are far from optimal, so that it is necessary to apply metaheuristic algorithms to complex
and realistic projects [17]. Recently, important approaches have been made by researchers to
improve the efficiency of resource consumption, proposing different heuristics which are
applicable to small projects; simple examples try to show the merits of a particular algorithm,
without establishing clear criteria for a performance comparison between the different
algorithms [18].
Following this line of work, Liao et al. [11] proposed some ideas to advance the RLP in
realistic environments; these authors made several proposals for the development or the
improvement of the RLP. Regarding resources allocation, these authors proposed the use of a
decision support system to assist project managers, as well as the development of benchmarking
tests for performance assessment and comparison [11]. Concerning resources leveling, they
suggested the use of multiple resources allowing the extension of the project deadline with a
penalty [11]. We take these proposals as challenges to be overcome in this paper, contributing a
little to the corpus of knowledge in this field.
Therefore, in this paper we present an Adaptive Genetic Algorithm (AGA) for the RLP
with multiple resources allowing the extension of the project deadline with a penalty; for this
purpose, we use the Weibull distribution as a termination condition, establishing an estimation
of the global optimum. The proposed algorithm is tested with the standard “project scheduling
problem library” (PSPLIB) [18], presenting a complete set of benchmarking tests. A decision
support system is also used in order to implement this algorithm. Without loss of generality, we
consider the classical resource leveling objective function: the total squared utilization cost for a
given schedule.
The remainder of this paper is organized as follows. Section 2 provides the classification
and formulation of the RLP. Section 3 details the different solving procedures: exact, heuristic,
and metaheuristic algorithms with the new use of the Weibull distribution as a termination
condition. Section 4 describes the algorithm proposed for the RLP with multiple resources.
Computational results and the benchmarking test are explained in Section 5. Finally,
conclusions are drawn.

2. Classification and Formulation of the Resource Leveling Problem

The general formulation of the RLP requires us to consider the following elements:
1. The set of activities, N :
N = { j1 , j2 , , jn } (1)
n being the total number of activities.
2. The set of durations, D :
D = {d1 , d2 , , dn } (2)
where di ,1 £ i £ n is the assigned duration for each activity.
3. The set of periods of time in which these activities have to be distributed:
T = {t1 , t2 , , t p } (3)
t p being the deadline of the project, from now on denoted T .
4. The set of resources, R:
R = {r1 , r2 , , rk } (4)
k being the total number of resources.
5. The set of availabilities of the resources, A :
A = {ait ,1 £ i £ k ,1 £ t £ p} (5)
where ait is the availability of the resource ri in the period t .
6. The set of costs, C :
C = {c1 , c2 , , ck } (6)
 7. The set SS : to distribute the performance of the activities along the elements of the set
T one needs to allocate a starting time for each activity, given by the ordered set, SS :
SS = {SS1 , SS2 , , SSn } (7)
SSi ,1 £ i £ n , is the starting time of the activity ji . T can be considered as the starting
time of a finish dummy activity SS finish , and then SS becomes:
SS = {SS1 , SS2 , , SSn , SS fifinish } (8)
Obviously, the schedule SS is not unique; on the contrary, there are a large number of
different possibilities, according to the logic and restrictions of the project to be
performed. Each of these schedules has significant differences in the efficiency of
resources consumption, and this is the reason for finding the values of SS which
optimize this efficiency.
8. The functions ri ( S , t ) , 1 £ i £ k : given a schedule SS , the function ri ( S , t ) is defined as
the consumption of the resource ri in the period of time t , belonging to the set T , in
such a way that the consumption of the resource ri throughout the project is given by:
ui1 = ri ( S , t1 ) , ui 2 = ri ( S , t2 ) , , uip = ri ( S , t p ) (9)
9. The function f : Given a schedule SS , the efficiency of resources consumption depends
on the layout of its use. Therefore, it becomes fundamental to establish an optimal
criterion for the distribution of the resources. This is the role we want f to play in the
development of the problem. Hence, the function f will be different for each
optimization criterion to be considered.

Once we have the elements that compose the problem, a general formulation could be:
k
Minimize å c f éër ( S , t )ùû
i =1
i i (10)

subject to:
SS finish £ T (11)
SSi + di + g ij £ SS j , for all i which are successors to j (12)
uij £ aij (13)
where g ij is the lead/lag between i and j
Having done this, the choice of the function f , which defines the criterion for the
optimization of the resources consumption, provides different ways of solving the problem. In
the case of the RLP, the optimization criterion focuses on getting the resource consumption as
level as possible. Consequently, a suitable choice of f could be:
( uit - ait )
2
T
f éë ri ( S , t ) ùû = å (14)
t =1 T
And Eq. (10) turns into:
( uit - ait )
2
k k T
Minimizeå ci × f éë ri ( S , t ) ùû =åå ci (15)
i =1 T i =1 t =1
Next, we can simplify the problem by taking into account the following:
1. ci = 1,for all 1 £ i £ k
2. ait = a,for all 1 £ i £ k and for all 1 £ t £ p
And then Eq. (15) becomes:
( uit - a )
2
k T
Minimizeåå (16)
i =1 t =1 T
 As the minimum depends neither on how far the variable is shifted nor on the constant
values, we can take a = 0 and T = 1 :
k T
Minimize åå u
r =1 t =1
rt
2
(17)

The objective function expressed in Eq. (17) is known as a minimum squares optimization,
and was introduced by Burgess and Killebrew [19] in a heuristic algorithm in which the near-
optimality is determined by the schedule with the minimum total sum of the squares of resource
consumption for each period, as illustrated in Fig. 1. Other measures for the objective function
are the Minimum Moment (MOM) proposed by Harris [20], and more recently the entropy-
maximization proposed by Christodoulou et al. [21], using the maximality and sub-additivity
properties of the entropy function.
T T
Fig. 1 Initial åu
t =1
kt
2
= 10.669; levelled åu
t =1
kt
2
= 6.477

To compare the leveling effectiveness between different projects, we use the Resource
Improvement Coefficient (RIC) developed by Robert Harris [20], a measure that is independent
of the total resource demand and is given by Eq. (18). The RIC relates the variation of a
selected resource histogram to an ideal resource histogram which is a rectangle-shaped resource
histogram (the ideal leveled schedule corresponds to a RIC value of one):
T
T × å ut 2
t =1
RIC = 2
(18)
æ ö T

ç å ut ÷
è t =1 ø
A different kind of non-regular objective function for Eq. (14) can be considered if the
objective function to be optimized represents the Net Present Value Problem. In this case, the
objective function represents the net present value of the project, which is to be maximized, and
this is used in practice when expensive resources have to be purchased.

3. Solving Procedures
The previous conceptual linear programming cannot be solved directly, because there is no
easy way to translate the set S , used in Eq. (10), into a linear programming formulation. Other
linear programming formulations have to be used in order to be able to specify the resource
constraints in a correct and solvable form.
The most efficient formulations are based on integer and binary programming, and can be of
two kinds, depending on whether the decision variables establish the period of execution or the
finishing time of the activities.
The first formulation for the Resource Leveling Problem in the multi-mode case is based on
a set of binary decision variables x jst [22] that establish the period in which the activities are
finished:
ì1, if job j is finished in mode s at the end of period t ü
x jst = í ý,
î0, otherwise þ (19)
for j Î J , s Î M j ,and t Î éë ES j + d j , , LFj ùû
The variables x jst can only be executed in one mode, and are defined over the interval
between the earliest and latest finishing times (the delimiting periods) of the activities of the
project. These limits are established using the traditional forward and backward pass
calculations for the unconstrained problem.
The vector SS of Starting Scheduled period for each activity is defined by:
Pj
æ LFj ö
SS j = å ç å t ×x jst - d js ÷ (20)
ç
s =1 è t = EF j
÷
ø
 The objective function to minimize the minimum total sum of the squares of resource
consumption for each period is:
2
æ R Pj LF j
T ö
Minimize åå ck × ç å å urst × å x jst ÷ (21)
ç jÎE(t ) s=1 ÷
k =1 t =1 è q = EF j ø
Subject to:
Pj LF j

ååx
s =1 t = EF j
jst = 1 for j Î J (22)

Pj LF j Pj LF j

å å (t - d ) × x
s =1 t = EF j
js jst -g ij - å å t ×x
s =1 t = EF j
ist ³ 0 for j Î J and i Î Pj (23)

x jst Î{0,1} for j Î J , s Î éë1, , Pj ùû and

d t Î éë ES j + d j , , LFj ùû (24)
The objective function modeled in Eq. (21) minimizes the total sum of the squares of
resource consumption for each period. Eq. (22) specifies that only one mode and one
completion time are allowed for every activity. The precedence constraints are given in Eq.
(23). Finally Eq. (24) specifies that the decision variables are binary.
Another possible formulation for the RLP is also based on binary programming, but the
decision variables x jst establish the period in which the activities are executed [23]:
ì1, if job j is processed in mode s in period t ü
x jst = í ý,
î0, otherwise þ (25)
for j Î J , s Î M j ,and t Î éë ES
S j + 1, , LFj ùû
The vector SS of Starting Scheduled period for each activity is defined by:
Pj
æ LFj ö
SS j = å ç å t × x jst × ( x jst -1 - x jst ) ÷ (26)
ç
s =1 è ES j
÷
ø
The objective function will be:
2
æ R Pj Tö
Minimize åå ck × ç å å urst × x jst ÷ (27)
ç jÎE (t ) s =1 ÷
k =1 t =1 è ø
subject to:
Pj LF j

å å
s =1 t = ES j +1
x jst = d j for j Î J (28)

t
d js × ( x jst - x j ,t +1 ) - å x jsq £ 0 for j Î J , s Î M j ,and t Î éë ES
S j + 1,, , LFj - 1ùû (29)
q = ES j +1
t -1
di × x jst - g ij - å
q = ES j +1
xist £ 0 for j Î J , i Î Pj , s Î M j ,and t Î éë ES j + 1, , LFi ùû (30)

x jst Î{0,1} for j Î J , s Î éë1, , Pj ùû and

d t Î éë ES j + 1, , LFj ùû (31)

The objective function in Eq. (27) minimizes the total sum of the squares of resource
consumption for each period. Eq. (28) ensures that each activity is executed for d j time units.
The execution of the activities without pre-emption is modeled in Eq. (29). The precedence
constraints are given in Eq. (30). Finally, Eq. (31) specifies that the decision variables are
binary. Easa [24] proposed a different binary integer formulation, preserving the precedence
constraints by limiting the shifting of the activities to the free float, which must be equal to or
greater than zero.
 The Resource Leveling Problem is NP-Hard even if only one resource is considered [25,12].
The time complexity function is of the order of O ( q n ) with q being a positive constant. The
universe of schedules for an instance is:
J Mj

ÕÕ ( Ht
j =1 s =1
js + T - mkp + 1) (32)

with Ht js being the total float of the activity j processed in mode s , T the prescribed makespan,
mkp the makespan for the resource unconstrained problem (RUPSP or resource relaxation of
the RCPSP) and M j the execution modes of activity j .
If the prescribed makespan is established to be equal to the makespan for the RUPSP and
only one execution mode is considered, Eq. (32) can be simplified to:
J

Õ ( Ht
j =1
js + 1) (33)

3.1. Exact Algorithms

Exact algorithms based upon implicit enumeration, integer programming, and dynamic
programming techniques, have been proposed to solve the RLP. Easa [24] used a mixed binary-
integer programming technique that guarantees the optimum leveling. Exhaustive enumeration
procedures were presented by Ahuja [26]. Bandelloni et al. [27] developed an optimal technique
based on dynamic programming. Recently, Rieck, Zimmermann & Gather [28] proposed a new
mixed-integer linear model formulation and domain-reducing pre-processing techniques for the
RLP based on smart discrete-time formulations.
The branch-and-bound technique [29] is probably the most widely used exact solution
technique for solving project scheduling problems, as it is the only technique which allows for
the generation of optimal solutions with acceptable computational effort. Neumann and
Zimmermann [30] describe a branch-and-bound procedure that reduces the set of all feasible
solutions by successively scheduling activities for approximately solving the problem. Recently,
a lower bound improved method (Maximum Allowable Daily Resources Method) to branch the
nodes has been developed by Mutlu [31], which may form a basis for the performance
evaluation of heuristic and metaheuristic procedures for the RLP. Recently, Gather et al. [32]
presented a new enumeration scheme embedded into a branch-and-bound framework using a
constructive lower bound as well as pre-processing techniques and Hariga and El-Sayegh [33] a
new mixed integer binary linear optimization model that allow activity splitting and minimizes
its associated costs.
The RLP as an NP-Hard problem has a phenomenon of “combinatorial explosion” [15],
especially for large-scale projects. For this reason, exact algorithms are only efficient for small
projects.
3.2. Heuristics Algorithms
To avoid the explosion problem, heuristic rules are mostly used to solve the RLP. These are
simple rules or sets of rules which aim to obtain a “good” solution (locally optimal) for a
difficult problem but do not guarantee the best solution (globally optimal). Heuristic algorithms
can be of two kinds, construction and improvement procedures. Construction procedures are
used to establish a feasible solution to the problem, and the other procedures are used to
improve it.
The first heuristic procedure for the RLP was proposed by Burgess and Killebrew [19],
establishing the minimum squares as the performance measure. The Burgess-Killebrew
algorithm is an improvement procedure that readjusts the starting time of each activity and
reduces the variability to a near optimum (local optimum). The first step of the algorithm
consists in establishing the starting time scheduled of each activity at their earliest starting time.
Later, the activities are selected one by one according to the priority rule (earliest finishing
time), selecting the best starting time, which is the one that minimizes the total sum of the
squares of resource consumption for each period.
 The Burgess-Killebrew algorithm is a single-pass algorithm that requires one to be aware of
the precedence relationships between activities because this is not considered in the formulation.
This problem was considered and solved by Burman [33] using the free float as the limit for the
activity shifting.
Harris proposed a multi-pass algorithm [20] called the MOM, which established the
minimum moment as the performance measure for minimizing the daily fluctuations in resource
use while keeping the total project duration unchanged. Later, Harris improved the MOM with
the packing method (PACK) [34], which recognizes network interactions with a more in-depth
analysis. Martinez and Ioannou [35] improved this method by introducing the Modified
Minimum Moment Method to level resources in construction projects. Hiyassat [36] devised a
heuristic to reduce the calculations needed for the minimum moment approach. Jeetendra et al.
[37] proposed the use of Petri Nets (PNs) and a P-matrix to help in token movements,
describing an algorithm for the RLP.
3.3. Metaheuristic algorithms
Given the variety of network structures and resources, no single heuristic rule can always
produce the optimal solution for all the RLP, and exact algorithms require extensive
computational effort. Therefore they are unable to deal with real large problems.
Metaheuristic methods are general purpose high level search frameworks grounded in
physical, biological and animal behavior that can be applied to any optimization problem.
Several metaheuristic methods have been proposed, such as Grasp (Greedy Randomized
Adaptive Search Procedure). Other improved methods are evolutionary algorithms (EA),
genetic algorithms (GA), scatter search (SS), simulated annealing (SA), ant colony optimization
(ACO), particle swarm optimization (PSO), and shuffled frog-leaping (SFL).
Evolutionary and genetic algorithms (EAs/GAs) [39,40] and simulated annealing (SA) [41]
are the most popular metaheuristic methods for solving the RLP. Hybrid methods are presented
by Son and Skibniewski [41], and by Alsayegh & Hariga [42], with a new approach for the RLP
which considers the cost of splitting the non-critical activities and combines the particle swarm
optimization (PSO) and simulated annealing (SA) methods. A hyperheuristic approach
implemented within commercial project management software has been proposed by Koulinas
and Anagnostopoulus [43]. Hyperheuristics chooses a sequence of “knowledge poor” heuristics
that are used to choose the most appropriate low-level heuristic from a set of heuristics during
the search.
Metaheuristic algorithms search for a better solution until a termination condition (or one of
a set of termination conditions) has been satisfied. Possible reasonable termination conditions
include: a certain amount of time, a maximum CPU time, a fixed maximum number of
iterations, a vector SS of Starting time Scheduled with a measured value less than a predefined
threshold value (Objective Bound), and a fixed maximum number of iterations without
improvements in the best solution found so far.
An Objective Bound as a termination condition should be established by computing a Lower
Bound (LB), by relaxing the restrictions of the RLP. Some LB methods are LP-relaxation,
cutting planes relaxing the integrality restrictions, Lagrangean Relaxation, removing a set of
constraints from the original problem and incorporating them into the objective function, or Set
Covering Based Approach, discarding the precedence constraints and/or allowing the pre-
emption of activities [44].
Recently, Paya-Zaforteza et al. [45] and Carbonell, Yepes and González-Vidosa [46] used
the three-parameter Weibull distribution [47] to establish an estimation of the global optimum.
They based their research on the fact, proved by Fisher and Tippet [48] that, as the number of
independent samples of size m grows, their distribution approaches a three-parameter Weibull
distribution with location parameter g as an estimation of the global optimum.

4. An Adaptive Genetic Algorithm (AGA) for the RLP with multiple resources
Genetic Algorithms (GAs) [49] are stochastic search techniques based upon the principles
of Darwinian evolution, and simulate biological evolution. Basically, a collection of candidate
solutions (initial population) is manipulated (evolved) through a number of iterations
(generations). A candidate solution, called a chromosome or individual, is represented by a set
of integer values (genes).
In the RLP, the most usual representation for encoding a candidate solution is the vector
SS of Starting time Scheduled for each activity [15]:
individuali = {SSi1 , SSi 2 , , SSij , , SSiiJ } (34)
The initial population is generated randomly by heuristic constructive algorithms, and the
quality of individuals is evaluated and ranked using Eq. (10) as a fitness function. In the RLP the
fitness function can be any of the performance measures: the minimum moment [20], the
minimum squares (Eq. 17), the RIC (Eq. 18) or the entropy function [21].
The initial population ( P ) evolves, under a set of cyclic genetic operators. First, two
parents are randomly selected, based on their fitness values (a mechanism is used to ensure that
a solution with a better fitness has a higher chance of being selected as a mate). One or more
point crossovers divides the parents into different sub-parts, generating, by recombination, two
children or offspring ( P ' ) as new candidate solutions. Some genes of the offspring can mutate
producing different offspring ( P '' ), adding them to the population and reducing the population
by selection. The previous algorithm can be structured as follows:
begin :
P ¬ Generate InitialPopulation()
Evaluate( P)
Do
P ' ¬ Recombine( P)
P '' ¬ Mutate( P ')
Evaluate( P '')
P ¬ Select ( P ''È P ')
Loop until termination condition is met
end
4.1. The formulation for the AGA
The formulation used to solve the RLP with the proposed AGA is based on an integer
programming formulation with decision variables d j that represents the shifting of the initial
early start ( ES j ) of each activity using the traditional forward and backward pass calculations
for the resource unconstrained problem.
The vector SS of Starting time Scheduled (not period) for each activity is defined by:
SS j = d j + ES j (35)
The objective function to minimize the minimum total sum of the squares of resource
consumption for each period is:
2
R T æ ö J
Minimize åå ck × ç å ukt ÷ + e × å d j + r × DT (36)
ç jÎE (t ) ÷
k =1 t =1 è ø j =1

Subject to:
d j + ES j - (d i + ESi + di + g ij ) ³ 0 for j Î J and i Î Pj (37)
ES finish £ T + DT (38)
d j ³ 0 and integral for j Î J (39)
Consequently, the individuals are encoded as follows:
{
individuali = d i1 , d i 2 , , d ij , , d iiJ } (40)
The objective function Eq. (36) minimizes the total sum of the squares of resource
consumption for each period. The precedence constraints are given in Eq. (37). The prescribed
makespan, as the maximum project duration, is preserved by Eq. (38) and, finally, Eq. (39)
specifies that the decision variables are non-negative integer values.
 The objective function in Eq. (36) has been improved by “punishing” the minimum total
sum of the squares, ensuring a better solution with the lower shifting and consequently
increasing the float of the activities. Besides this, Eq. (36) has been completed by allowing an
extension ( DT ) of the project deadline ( T ) or the prescribed makespan, with penalty ( r ). To
avoid undermining the main objective e must be very small; we recommend a value such that
the sum of the total float for all the activities multiplied by e is less than one, as expressed in
Eq. (41):
J
e × å Totalfloat j < 1 (41)
j =1

One of the major difficulties of GA is the premature convergence toward suboptimal

solutions, and the loss of genetic variability that is caused by inbreeding (crossing individuals
with common parentage). To avoid this, a GA must find an equilibrium point between
intensification and diversification strategies. Intensification produces convergence to a local
optimum with high levels of inbreeding, and diversification produces unfeasible individuals.
Static and dynamic strategies have been proposed by researchers; the static methodologies
modify the operation parameters according to a linear or logarithmic function [50]; the dynamic
strategies adjust the mutation probability [51], or crossover and mutation [52], dynamically,
with the aim of overcoming the premature problems of convergence and closing to a local
optimum.

4.2. The initial population

The initial population is randomly generated by a heuristic backward constructive procedure
to preserve the maximum diversification of the population with feasible individuals. The
procedure in pseudocode is as follows:
begin :
Topological ordering
Popsize = max(50,2 ´ Totaljobs)
For(k = 1; Popsize;1)
For( j = J ;1; -1) in a decreasing topological order
d ij = random(0, Free Float + DT )
;
;
end
Note that in the previous procedure, the shifting values d ij are randomly established
between 0 and the free float of each activity, to guarantee that all the individuals are feasible
solutions to the problem.

4.3. The evolution

Once the initial population is generated, individuals evolve by a principle inspired by
natural evolution, each transmitting its genetic material to its offspring. Only the most adapted
individuals (the elite) are able to survive and breed new individuals. Additionally, the genetic
information of the offspring is subject to small mutations.
The population is evaluated and ranked, and Eq. (42) is applied to establish the individuals
of the elite as a fraction of the total population:
0 < Elitesize < 1
Elite ¬ (Elitesize ´ Popsize) Best individuals
Based on the knowledge of the problem, the global optimum is in the neighborhood of the
best individuals. Thus, parents are selected from the elite population through the classic
roulette-wheel mechanism to probabilistically select individuals based on the fitness measure
( f i ):
 f
P ( Si ) = i
Elitesize´ Popsize
;1 £ i £ Elitesize ´ Popsize (42)
åi =1
fi

Parents ¬ random P(Si )

4.4. Recombination and mutation

Like its counterpart in nature, recombination or crossover produces offspring
( s-on and d -aughter ) that have some parts of both ( f -ather father and m-other ) parent’s genetic
material. For a one-point crossover C , the pseudo-code is:
individual f = {d f 1 , d f 2 , , d fc , d fc +1 , , d fJ }
individualm = {d m1 , d m2 , , d mc , d mcc +1 , , d mJ }
Crossover (C ) = random(1, J )
offspring s = {d f 1 , d f 2 , , d fc , d mc +1 , , d mJ }
offspringd = {d m1 , d m 2 , , d mc , d fc +1 , , d fJ }
For a multi-point crossover, the procedure is very similar. Crossover points are chosen
randomly and sorted without duplicates. Then the genes between successive crossover points
are exchanged between the parents to produce new offspring.
A mutation operator acts as natural mutation does in natural evolution, perturbing the
genome. In the mutation process, randomly selected genes of a chromosome are
probabilistically replaced by other genes from the valid domain of the problem, to produce a
new genetic structure. The mutation process is controlled by a mutation probability range
( mutprob ) that must lie between 0 and 1 ( 0 < mutprob < 1 ) and a maximum number of mutation
genes ( maxmut ).

4.5. Evaluation and Unfeasibility

Along the evolution process, offspring are potentially infeasible. Infeasible individuals are
penalized in the function that measures the quality of an individual, but are not rejected; this
guarantees the diversification of the population.
The feasibility range of individuals in the evolved population is controlled using the
feasibility lower and upper ranges ( Flr, Fur ). The feasibility range guarantees the equilibrium
between intensification and diversification strategies, changing the mutation probability range
( mutprob ) by a value of f and the maximum number of mutation genes ( maxmut ) byh ,
dynamically maintaining the genetic evolution, the diversity in the population and the
convergence capacity of the AGA. The proposed adaptive procedure in pseudocode is as
follows:
begin :
P ¬ Generate InitialPopulation()
Evaluate( P)
f = 1;h = 0
Do
P ' ¬ Recombine( P)
P '' ¬ Mutate( P ')
Evaluate( P '')
if feasibility in last g generations is out of range adjust mutprob and maxmut :
if feasibility < Flr; maxmut - = f;mutprob- = h
if feasibility >Fur; maxmut += f;mutprob+ = h
P ¬ Select ( P ''È P ')

1
 Loop until termination condition is met
end
Due to the conditions of the problem, it is very difficult to establish a general rule for the
feasibility range. The best performance values have been obtained with values for the feasibility
lower range equal to 10% ( Flr = 0.1 ) and for the feasibility upper range equal to 35%
( Fur = 0.35 ).The effect on the evolution process with different values for the feasibility lower
and upper ranges can be seen in Fig. 2:

Fig. 2 Effect of different values for the feasibility lower and upper ranges

4.6. Termination conditions; the Weibull distribution

4.6.1. Formulation of the Weibull distribution

The formulation of the Weibull distribution to compute the location parameter ( g ) or, as we
call it for this special case the Weibull Bound (WB) , to be used as termination condition, is
explained in the following lines. Let F be the distribution of a random variable X , in other
words: F ( x) = Pr ob [ X £ x ] , such that, given a sample of size n , the equality:
Pr ob [ X1 £ x1 , , X n £ xn ] = Pr ob [ X1 £ x1 ] Pr ob [ X n £ xn ] (43)
holds. Let Ln be a new random variable, given by:
Ln = min { X i } (44)
i =1,2,, n

The problem we want to solve is to determine the distribution that Ln follows, in particular
at the limit of large samples, and the asymptotic behavior of the random variable Ln . Fisher and
Tippet [48] solved this problem based on the so-called Stability Postulate: The distribution of
the largest (smallest) value in samples of size N ´ n will tend to the same asymptotic expression
as the distribution of the largest (smallest) value in samples of size n , since the largest (smallest)
of the minimums of N samples of size n is the same as the minimum of the sample of
size N ´ n . Consequently the asymptotic distribution must be the same for both cases.
Since a linear transformation does not change the form of the distribution, the Stability
Postulate can be formulated as:
F n ( x ) = F ( an x + bn ) (45)
Solving this functional equation for F , the Weibull distribution for the smallest value is
obtained as:
ì
b
æ x -g ö
-ç ÷
ï h
, if x > g
W ( x) = í1 - e
è ø
(46)
ï
î0, if x £ g
with h , b > 0 and where g is called the location parameter, h the scale parameter, and b the
shaper parameter.

4.6.2. How to compute the parameters g ,h and b when F ( x) is unknown

First: We take a sample of size n in a random way.
Second: From the sample, we calculate the density and distribution functions:
N
f ( x) = x ; N x being the number of times that the value x appears in the sample
n
F ( x) = å f ( x) (47)
X £x

Third: We approach the Weibull distribution using the experimental values of F ( x) .

b b
æ x -g ö æ x -g ö
-ç ÷ 1 -ç
è h ø h ø÷
F ( x) = 1 - e ® =e è (48)
1 - F ( x)

1
 Taking logarithms:
b
æ 1 ö æ x -g ö
log ç ÷=ç ÷ (49)
è 1 - F ( x) ø è h ø
Taking logarithms again:
æ æ 1 öö
log ç log ç ÷ ÷ = b log ( x - g ) - b log (h ) (50)
è è 1 - F ( x) ø ø
Changing the variables:
æ æ 1 öö
y = log ç log ç ÷ ÷ ; z = log ( x - g ) (51)
è è 1 - F ( x) ø ø
Eq. (49) becomes:
y = b z - b log ( x - g ) (52)
which can be seen to be a linear regression problem in the variables y and z ;
d (g ) = s yz2 (g ) / (s z2 (g )s y2 )
2
(53)
The value of g that minimizes Eq. (53), g 0 , is found using derivation techniques. In this
way we get the regression for the best correlation between y and z :
y = az + b (54)
Comparing Eq. (51) with Eq. (53) we get:
b = a; -b log (h ) = b (55)
which allows us to compute the other parameters b and h .
Finally, the termination conditions for the proposed AGA are:
ì# iteration = established number of iterations
ï
or í BB £ Objective Bound , where BB is the better bound obtained
ïW ( BB) = P ( x £ BB ) £ 0.001, where W is the Weibull distribution
î
And then the adaptive procedure in pseudocode is as follows:
begin :
P ¬ Generate InitialPopulation()
Evaluate( P)
Compute Weibull Bound (WB = ëêg ûú ) with InitialPopulation()
f = 1; e = 0
Do
P ' ¬ Recombine( P)
P '' ¬ Mutate( P ')
Evaluate( P '')
if feasibility in last g generations is out of range adjust mutprob and maxmut :
if feasibility < Flr; maxmut - = f;mutprob- = e
if feasibility >Fur; maxmut += f;mutprob+ = e
P ¬ Select ( P ''È P ')
Loop until termination condition is met
end

4.6.3. Numerical example

For a better understanding and to illustrate the application of the Weibull distribution for the
estimation of the global optimum as a termination condition, we will show a step by step
calculation, based on the first instance of the set of problems j30. The exact location
parameter (g ) of the Weibull distribution is obtained when n , the size of the sample, goes to

1
infinity. For each sample of size n, a location parameter (g n ) is obtained, in such a way
that lim g n = g . So there is a monotonous subsequence g nk ; ( k = 1, 2,3, ) , either increasing or
n ®¥
decreasing, that tends to g when k goes to infinity. Then, by performing a heuristic search for
such a monotonous subsequence of samples of size n1 = 32 , n2 = 64 , n3 = 96 , we obtain the
following results step by step:
First: We choose three samples of feasible solutions of sizes 1´ n,2 ´ n,3 ´ n , ( S1 , S2 , S3
respectively), n being the number of jobs of the instance j30_1, in a random way, for example,
the S 2 sample:
S 2 ={7863, 8115, 7905, 7931, 7983, 8151, 7969, 7957, 8113, 8077, 7803, 7903, 8055,
8163, 8193, 8127, 8145, 7963, 8051, 7959, 7733, 8151, 8005, 7767, 7883, 7937, 7939, 7877,
7945, 8001, 7745, 7811, 7949, 7865, 8043, 8161, 8033, 7763, 8039, 8259, 7881, 8129, 7993,
8187, 7957, 8019, 7921, 7965, 8057, 7985, 8137, 7989, 7981, 7999, 8003, 7775, 8171, 8081,
7969, 8193, 7855, 7953, 8075, 7975}
Second: From S1 , S2 , S3 , we calculate the density and distribution functions:
Nx
; F ( x) = å f ( x) ; (from Eq. (46) and Eq. (47)).
f ( x) =
n X £x

Third: We approach the Weibull distribution using the experimental values of F ( x) in

accordance with Eqs. (48) to (55), obtaining the location parameters (g 32 , g 64 , g 96 ) respectively,
and the other parameters ( b1 ,h1b2 ,h2 b3 ,h3 ) to compute the three Weibull distributions (see
Fig. 3):
g 32 = 7517.19 with R2 = 0.972123
g 64 = 7494.65 with R2 = 0.982858
g 96 = 7488.38 with R2 = 0.968511

Fig. 3 Approach to the Weibull distribution for W32 ,W64 and W96

Fourth: if the algorithm finds a value for BB (Better Bound) which satisfies the
inequalities Wn ( BB) = P ( x £ BB ) £ 0,001; n = 32,64,96 , then it stops.
W32 (7549) = 0.000017 £ 0.001; where BB=7549
W64 (7549) = 0.000036 £ 0.001; where BB=7549
W96 (7549) = 0.000313 £ 0.001; where BB=7549

5. Computational results and benchmarking test

As test instances, we have used the standard sets j30, j60 and j120 for the RCPSP (Resource
Constrained Project Scheduling Problem). The instances, represented as activity-on-node
networks, were generated with a full-size instance generator, named ProGen [53], for a general
class of project scheduling problems with four renewable resources. The PSPLIB library is fully
accessible in the Project Scheduling Project Library (PSPLIB) at
http://129.187.106.231/psplib/library.html [18].
The values used to compute the instances are shown in Table 1 and Table 2:
Table 1 GA parameters (1)
The instances are solved with e = 0 and Ck = 1 , the prescribed makespan as that obtained for
the RUPSP (the resource relaxation of the RCPSP), and the penalty as zero ( r = 0 ). The
termination conditions are established at a maximum number of iterations and a probability for
WB of 0.001%, as can be seen in Table 2:

1
 Table 2 GA parameters (2)
The AGA is programmed with VBA (visual Basic for applications) for Excel 2010,
applying a matrix algorithm for the RUPSP (the resource relaxation of the RCPSP), as can be
seen in Fig. 4 [54,55]. We have used a computer with an Intel Core i7 processor, 3.6Gh of
velocity and 8 GB of RAM with an average time consumed for the j30 set of 15 seconds with
1000 iterations.

Fig. 4 Decision support system application (VBA for Excel 2010)

The main statistics of the best improvement values found over the initial value of the
measure function (total sum of the squares of resource consumption for each period) are shown
in Table 3:
Table 3 Test statistics

The complete information about the benchmarking test can be downloaded from the
“Benchmarking for Resource Leveling Problem (PSPLIB j30, j60, j120)”,
(https://docs.google.com/spreadsheet/ccc?key=0AgisRN826029dHdEZmNDbmxBZDI1QmZN
SGQ4a1VhY3c), with the initial and leveled value for all the analyzed instances, the
improvements and the vector SS of Starting time Scheduled for each activity.
The correlation values obtained as a function of the logarithm of the time complexity and
the initial value of the measure function for the solved instances are shown in Table 4, Fig.5 and
Fig.6:
Table 4 Test correlation
Fig. 5 Correlation ( rxy ); Improvement over logarithm of the time complexity O ( q )
n

Fig. 6 Correlation ( rxy ); Improvement over initial sum of the squares

Finally, as there are no known results from other researchers for the PSPLIB, and to
determine how good the proposed AGA algorithm is, we have computed the same instances
with three different parallel scheduling schemes [56]: a parallel backward scheme with earliest
finishing time as the priority rule and maximum free float as the secondary one, a parallel
forward scheme with latest finishing time as the priority rule and maximum total float as the
secondary one, and a forward-backward scheme with the priority rule depending on the
direction of the scheduling. The results obtained compared with the proposed AGA are shown
in Table 5 and Fig. 7.
Table 5 Heuristic results vs. proposed AGA
Fig. 7 Heuristic results vs. proposed AGA

The results obtained in the test instances show that there is a significant positive correlation
between the improvement and the logarithm of the time complexity O ( q n ) , but that this is not
statistically significant when compared with the initial sum of the squares.
This is due to the fact that algorithmic complexity is determined by the slack in the tasks,
and, in the majority of cases, if the slacks are regularly distributed over the geometry of the
graph they will provide a greater number of feasible solutions to the problem, and consequently
a greater improvement capacity.
The heuristic does not differ significantly between the forward and backward scheduling
schemes, but a forward-backward scheme presents significant improvements compared with the
single direction scheduling scheme. The improvement of the forward-backward scheme with
respect to case of the single direction scheme is constant, and it seems not to be related to the
time complexity or the number of tasks. However, comparing the values obtained from the
application of the AGA and the analyzed heuristics, the improvement over the best heuristic
significantly increases with the number of tasks of the problem.

6. Conclusions
The Resource Leveling problem (RLP) is a non-regular problem whose objective is to
achieve a resource consumption which is the most efficient possible, without increasing the

1
prescribed makespan of the project. Conventional analytical and heuristic methods are neither
flexible nor productive when solving the RLP. Exact procedures are not useful for offering
optimal solutions with acceptable computational effort; and, on the other hand, heuristics offer
solutions which are far from the optimal so that it is necessary to apply metaheuristic algorithms
for complex and real projects in realistic environments.
In this paper:
1. We describe the complete state of the art of the RLP, proposing different binary and
integer mathematical formulations and incorporating modifications and improvements
on previous contributions.
2. We propose an Adaptive Genetic Algorithm (AGA) for the RLP with multiple resources
allowing the extension of the project deadline with a penalty.
3. We propose the use of the three-parameter Weibull distribution as a termination
condition for the metaheuristic, with location parameter g or the Weibull Bound ( WB )
as an estimation of the global optimum.
The previous contributions have been tested with the standard “project scheduling problem
library” (PSPLIB), presenting a complete set of benchmarking tests solved with only the most
usual and common parameters to provide clear criteria for comparison between the different
algorithms. To prove the merits of the proposed algorithm, the results we obtained have been
compared with the most common heuristic procedures and with a more efficient forward-
backward scheduling scheme. The proposed AGA is always better than the heuristics, especially
for the most difficult problems with 120 jobs.
The proposed AGA for the RLP has been implemented with VBA for Excel 2010 to provide
a flexible and powerful decision support system that enables practitioners to choose between
different feasible solutions to a problem, and that is in addition easily adjustable to the
constraints and particular needs of a project in a realistic environment. This contribution is a
tool that can be applied in a direct and simple way by practitioners; besides, it can serve as a
starting point for specialists in order to develop user-friendly and practical computer
applications to provide realistic and good solutions for production and project management.
The use of the Weibull distribution was applied following a heuristic process. Its
improvement is proposed as a future research field, searching for a bound ( e > 0 ) in such a way
that the inequality g n - g £ e , "n ³ N0 holds for all sample sizes n greater than a given
size N 0 ; the exact parameter would therefore be in the interval g n - e £ g £ g n + e .
Another important area for future research is the consideration of the graph density as a new
variable in the computational test, to determine its influence on the optimization and the
correlation between the other variables.

Acknowledgments
This study was partially funded by the Spanish Ministry of Science and Innovation
(research project BIA2011-23602).

References

[1] J. Turner, "Do you manage work, deliverables or resources?," International Journal of
Project Management, vol. 18, no. 2, p. 83–84, 2000.
[2] E. Goldratt, Critical Chain, North River Press, Inc., Great Barrignton, MA, 1997.
[3] I. Clark and T. Colling, "The management of human resources in project management-
led organizations," Personnel Review, vol. 34, no. 2, p. 178–191, 2005.
[4] J. Turner, "Towards a theory of project management: the nature of the project,"
International Journal of Project Management, vol. 24, no. 1, p. 1–3, 2006b.

1
[5] F. Karaa and A. Nasr, "Resource management in construction," Journal of Construction
Engineering and Management, vol. 112, no. 3, p. 346–357, 1986.
[6] M. Park, "Model-based dynamic resource management for construction projects,"
Automation in Construction, vol. 14, no. 5, p. 585–598., 2005.
[7] J. Turner, "Towards a theory of project management: the functions of project
management," International Journal of Project Management, vol. 24, no. 3, p. 187–189,
2006a.
[8] J. Söderlund and K. Bredin, "HRM in project-intensive firms: changes and challenges,"
Human Resource Management, vol. 45, no. 2, p. 249–265, 2006.
[9] M. Huemann, A. Keegan and J. Turner, "Human resource management in the project-
oriented company: a review," International Journal of Project Management, vol. 25, p.
315–323, 2007.
[10] S. Al-Jibouri, "Effects of resource management regimes on project schedule,"
International Journal of Project Management, vol. 20, p. 271–277, 2002.
[11] T.W. Liao, P.J. Egbelu, B.R. Sarker and S.S. Leu, "Metaheuristics for project and
construction management - A state-of-the-art review," Automation in Construction, vol.
20, no. 5, p. 491–505, 2011.
[12] K. Neumann, C. Schwindt and J. Zimmermann, Project Scheduling with Time Windows
and Scare Resources, Springer,Berlin, 2003.
[13] W. Herroelen, B. de Reyck and E. Demeulemeester, "Resource-constrained project
scheduling: a survey of recent developments," Computers & Operations Research, vol.
25, no. 4, p. 279–302, 1998.
[14] K. Brinkmann and K. Neumann, "Heuristic procedures for resource-constrained project
scheduling with minimal and maximal time lags: the resource-levelling and minimum
project-duration problems," Journal of Decision Systems, vol. 5, p. 129–156, 1996.
[15] S.S Leu, C.H. Yang and J.C. Huang, "Resource leveling in construction by genetic
algorithm-based optimization and its decision support system application," Automation
in Construction, vol. 10, p. 27-41, 2000.
[16] S. Hartmann, "Project scheduling with multiple modes: a genetic algorithm," Annals of
Operations Research, vol. 102, p. 111-135, 2001.
[17] P. Jaśkowski and A. Sobotka, "Scheduling construction projects using evolutionary
algorithm," Journal of Construction Engineering and Management, vol. 132, no. 8, p.
861-870, 2006.
[18] R. Kolisch and A. Sprecher, "PSPLIB - A project scheduling problem library,"
European Journal of Operational Research, vol. 96, p. 205-216, 1996.
[19] A. Burguess and J. Killebrew, "Variation in activity level on a cyclic arrow diagram,"
The Journal of Industrial Engineering, vol. 13, p. 76-83, 1962.
[20] R. Harris, Precedence and Arrow Networking Techniques for Construction, Wiley, New
York, 1978.
[21] S. E. Christodoulou, G. Ellinas and A. Michaelidou-Kamenou, "Minimum moment
method for resource leveling using entropy maximization," Journal of Construction
Engineering and Management, vol. 136, no. 5, p. 518-527, 2010.
[22] A.A.B. Prisker, L.J. Watters and P.M. Wolfe, "Multiproject scheduling with limited
resources: a zero-one programming approach," Management Science, vol. 16, p. 95-
108, 1969.
[23] L. Kaplan, "Resource-constrained project scheduling with preemption of jobs,"
Unpublished Ph. D. Thesis, University of Michigan, 1988.
[24] F. Ballestin, C. Schwindt and J. Zimmermann, "Resource leveling in make-to-order
production: modeling and heuristic solution method," International Journal of
Operations Research, vol. 4, no. 1, p. 50-62, 2007.
[25] S.M. Easa, "Resource leveling in construction by optimization," Journal of Construction
Engineering and Management, vol. 115, no. 2, p. 302-316, 1989.
[26] H. Ahuja, Construction Perfomance Control by Networks, Wiley, New York, 1976.

1
[27] M. Bandelloni, M. Tucci and R. Rinaldi, "Optimal resource leveling using non-serial
dynamic programming," European Journal of Operational Research, vol. 78, no. 2, p.
162-177, 1994.
[28] J. Rieck, J. Zimmermann and T. Gather, "Mixed-integer linear programming for
resource leveling problems," European Journal of Operational Research, vol. 221, no. 1,
p. 27-37, 2012.
[29] N. Agin, "Optimum seeking with branch and bound," Management Science, vol. 13, no.
4, p. B176-B185, 1966.
[30] K. Neumann and J. Zimmermann, "Procedures for resource leveling and net present
value problems in project scheduling with general temporal and resource constraints,"
European Journal of Operations Research, vol. 127, no. 2, p. 425-443, 2000.
[31] M. Mutlu, A branch and bound algorithm for resource leveling problem, Middle East
Technical University, Ankara (Turkey), 2010.
[32] T. Gather, J. Zimmermann and J.H. Bartels, "Exact methods for the resource levelling
problem," Journal of Scheduling, vol. 14, no. 6, p. 557–569, 2011.
[33] M. Hariga and S.M. El-Sayegh, "Cost optimization model for the multiresource leveling
problem with allowed activity splitting”. Journal of Construction Engineering and
Management, vol. 137, no.1, p. 56-64, 2011.
[34] P. Burman, Precedence Networks for Project Planning and Control, McGraw Hill,
London, 1972.
[35] R. B. Harris, "Packing method for resource leveling (PACK)," Journal of Construction
Engineering and Management, vol. 116, no. 2, p. 331-350, 1990.
[36] J. Martinez and P. Ioannou, "Resource leveling based on the modified minimum
moment heuristic," Civil and Building Engineering, p. 287-294, 1993.
[37] M. Hiyassat, "Modification of minimum moment approach in resource leveling,"
Journal of Construction Engineering and Management, vol. 126, no. 4, p. 278-284,
2000.
[38] V.A. Jeentra, O.V. Ksishnaiah Chetty and J. Prashanth Redy, "Petri nets for project
management and resource leveling," International Journal of Advanced Manufacturing
Technology, vol. 16, p. 516-520, 2000.
[39] T. Hegazy, "Optimization of resource allocation and leveling using genetic algorithms,"
Journal of Construction Engineering and Management, vol. 125, no. 3, p. 167-175,
1999.
[40] J. Roca, E. Pugnaghi and G. Libert, "Solving an extended resource leveling problem
with multiobjetive evolutionary algorithms," World Academy of Science, Engineering
and Technology, vol. 46, p. 712-723, 2008.
[41] J. Son, and M. Skibniewski, ”Multiheuristic approach for resource leveling problem in
construction engineering: hybrid approach,” Journal of Construction Engineering and
Management, vol. 125, no. 1, p. 23-31, 1999.
[42] H. Alsayegh and M. Hariga, "Hybrid meta-heuristic methods for the multi-resource
leveling problem with activity splitting," Automation in Construction, vol. 27, p. 89-98,
2012.
[43] G. K. Koulinas and K. P. Anagnostopoulus, "Construction resource allocation and
leveling using a threshold accepting based hyperheuristic algorithm," Journal of
Construction Engineering and Management, vol. 138, no. 7, p. 854-863, 2012.
[44] A. Mingozzi, V. Maniezzo, S. Ricciardelli and L. Bianco, "An exact algorithm for the
resource-constrained project scheduling based on a new mathematical formulation,"
Management Science, vol. 44, no. 5, p. 714-729, 1998.
[45] I. Paya-Zaforteza, V. Yepes, F. González-Vidosa and A. Hospitaler, "On the Weibull
cost estimation of building frames designed by simulated annealing," Meccanica, vol.
45, p. 693-704, 2010.
[46] A. Carbonell, V. Yepes and F. González-Vidosa. “Automatic design of concrete vaults
using iterated local search and extreme value estimation,” Latin American Journal of
Solids and Structures (accepted, in press).

1
[47] W. Weibull, "A statical distribution function of wide applicability," Journal of Applied
Mechanics, vol. 18, no. 3, p. 293-297, 1951.
[48] R. Fisher and L. Tippett, "Limiting forms of the frequency distribution of the largest or
smallest member of a sample," Mathematical Proceedings of the Cambridge
Philosophical Society, vol. 24, p. 180-190, 1928.
[49] J. Holland, Adaptation in Natural and Artificial Systems, University of Michigan Press,
Ann Arbor, 1975.
[50] M. Mahdavia, M. Fesangharyb and E. Damangirb, "An improved harmony search
algorithm for solving optimization problem," Applied Mathematics and Computation,
vol. 188, no. 2, p. 1567-1579, 2007.
[51] C. Lin, "An adaptive genetic algorithm based on population diversity strategy," Third
International Conference on Genetic and Evolutionary Computing, Guilin, China, 2009.
[52] M. Srinivas and L. Patnaik, "Adaptive probabilities of crossover and mutation in genetic
algorithms," IEEE Transactions on Systems, Man and Cybernetics, vol. 24, no. 4, p.
656-667, 1994.
[53] R. Kolish, A. Sprecher and A. Drexl, "Characterization and generation of a general class
of resource-constrained project scheduling problems," Management Science, vol. 41,
no. 10, p. 1693-1703, 1995.
[54] J. L. Ponz-Tienda, J. Belloch-Marco, C. Andrés-Romano and D. Gil-Senabre, "A matrix
algorithm for the RUPSP non spliting allowed," Revista de la Construcción, vol. 10, no.
2, p. 90-103, 2011.
[55] J.L. Ponz-Tienda, Gestión de Proyectos con Excel 2010, Anaya Multimedia (in
Spanish), Madrid, 2011.
[56] R. Álvarez-Valdés and J.M. Tamarit, "Heuristic algorithms for resource-constrained
project scheduling: a review and empirical analysis," in: R. Skovinski and J. Weglarz
(Eds.), Advances in Project Scheduling, p. 113-134, 1989.

1
Figure 1

T T
Fig. 1 Initial åu
t =1
kt
2
= 10.669; leveled åu
t =1
kt
2
= 6.477
Figure 2

8500
10%-15%
8300 30%-35%
8100
50%-55%
10%-35%
7900

7700

7500

7300

7100

6900

6700

6500

Fig. 2 Effect of different values for the feasibility lower and upper range
Figure 3

W32 W64 W96

Fig. 3 Approach of Weibull distribution for W32 ,W64 and W96

Figure 4

Fig. 4 Decision support system application (VBA for Excel 2010)

 Figure 5

10.00%
15.00%
20.00%
25.00%
30.00%
35.00%
40.00%
45.00%
50.00%

0.00%
5.00%
1.00E+00

1.00E+05

1.00E+10

1.00E+15

1.00E+20

1.00E+25

1.00E+30

1.00E+35

1.00E+40

1.00E+45

1.00E+50

1.00E+55

1.00E+60

1.00E+65

1.00E+70

1.00E+75

1.00E+80

1.00E+85

1.00E+90

1.00E+95

1.00E+100

1.00E+105

1.00E+110

1.00E+115

1.00E+120

1.00E+125

1.00E+130

1.00E+135

1.00E+140
Fig. 5 Correlation ( rxy ); Improvement over logarithm of the time complexity
j30

1.00E+145

1.00E+150
j60

1.00E+155
O ( qn )

1.00E+160

1.00E+165
j120

1.00E+170

1.00E+175
Figure 6

50.00%

45.00%

40.00%

35.00%

30.00%

25.00%

20.00%

15.00%

10.00%

5.00%
j30 j60 j120
Lin(j30) Lin(j60) Lin(j120)
0.00%
0

100,000

200,000

300,000

400,000

500,000

600,000

700,000

800,000

900,000

1,000,000

1,100,000

1,200,000

1,300,000

1,400,000
Fig. 6 Correlation ( rxy ); Improvement over initial sum of the squares
Figure 7

30.000% 27.858%
28.000%

26.000%
23.099%
24.000%
AGA
22.000%
20.742%
Bw Fw
20.000% 18.452% 18.014%
Forward
18.000%
14.652% Backward
16.000%

14.000%

12.000%

10.000%

Fig. 7 Heuristic results vs. proposed AGA

Table 1

Table 1
GA parameters (1)
Totaljobs Popsize Elitesize Crosspoints mutprob Fur
Flr
j 30 32 32,64,96 0.1 1 0.1 15% 35%
j 60 62 62,124,186 0.1 1 0.1 15% 35%
j120 122 122,244,366 0.1 1 0.1 15% 35%
Table 2

Table 2
GA parameters (2)
e Ck ; k = {1, 2,3, 4} T , DT r instances max iterations
j 30 0 1 RUPSP, 0 0 480 1000
j 60 0 1 RUPSP, 0 0 480 2000
j120 0 1 RUPSP, 0 0 480 3000
Table 3

Table 3
Test statistics
Average Deviation % values
m s Min value Max value m ± 2s
j 30 18.45% 0.07176 4.41% 42.33% 96.04%
j 60 23.10% 0.07483 7.57% 48.24% 95.42%
j120 27.86% 0.06737 11.25% 47.74% 96.25%
Table 4

Table 4
Test correlation
Over complexity Over initial
O ( qn ) Sum of the squares
j30 j60 j120 j30 j60 j120
Covariance 0.146 0.301 0.556 -449.025 -2295.646 -854.354
Pearson ( rxy ) 0.544 0.544 0.583 -0.213 -0.341 -0.040
rxy2 29.578% 29.624% 33.962% 4.544% 11.609% 0.160%
Table 5

Table 5
Heuristic results vs. proposed AGA
j30 j60 j120
Parallel forward scheme 12.645% 15.314% 17.208%
Parallel backward scheme 12.641% 15.530% 17.496%
Fw-Bw scheme 14.652% 18.014% 20.742%
AGA 18.452% 23.099% 27.858%