Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Articulo Ebs Co

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Application of genetic algorithms to assembly sequence planning

with limited resources


J. Bautista, A. Lusa, R. Surez, M. Mateo, R. Pastor, A. Corominas
Institut dOrganitzaci i Control de Sistemes Industrials (UPC)
Diagonal 647, planta 11, 08028 Barcelona , Spain. Phone: +34-934016653, Fax: +34-934016605.
Contact emails: bautista@ioc.upc.es, suarez@ioc.upc.es, mateo@ioc.upc.es

Abstract
Heuristic procedures based on priority rules are quite
frequently used to solve the multiple resource-constrained
(RCPSP),
i.e.
task
project-scheduling
problem
programming with limited resources. The rules are based
on the problem knowledge. Different local search
procedures have been proposed in order to look for
acceptable solutions in scheduling problems. In this work,
local search procedures, that define the solution
neighborhood based on greedy heuristics, are proposed to
assign assembly operations to a fixed number of robots in a
manufacturing cell. A genetic algorithm is used to generate
the solution.

1 Introduction
The multiple resource-constrained project scheduling
problem (RCPSP) has been extensively treated in the
literature (e.g. [1][2][3]). Exact solutions have been
obtained using branch-and-bound procedures as well as
dynamic programming [4][5]. Nevertheless, these
procedures are only useful for low dimension problems
due to its NP-hard complexity [6].
In order to solve realistic problems, different heuristics
have been used like, for instance, those based on priority
rules constraining the serial or parallel dispatching of tasks
[7][8]. The rules consider different aspects like, for
instance, processing times (activity duration), slacks,
number of subsequent tasks, resource requirements,
randomizing, etc. The rules are applied step by step to
choose a task among a set of them whose precedents have
already been scheduled, while taking care that the resource
requirements fit the available resources. Usually, each
heuristic of this type has been associated to only one rule
that determines the task to be dispatched at each situation
(unless random selection is used).
These heuristics often produce acceptable solutions, and,
as average, the higher the number of aspects considered in
a rule the better the solution is. Nevertheless, it cannot be
concluded that there exists one rule that works better than
This work was partially supported by the CICYT Projects TAP 96-0868,
TAP 98-0494 and TAP 98-0471.

any other for any instance of the problem. Moreover,


unless a random selection of tasks is incorporated, the
rules will always produce the same solutions.
Another type of heuristics is based on local search [9] like,
for instance, Hill Climbing (HC), Simulated Annealing
(SA), Tabu Search (TS) and Genetic Algorithms (GA).
The GA, introduced by Holland in 1975 [10], can be
applied to the optimization of several combinatory
problems [11] and, in particular, to the scheduling
problems to analyze the behavior of different heuristics
[12][13] as well as to solve the problem itself [14][15].
This second type of heuristics provides alternative ways to
look for solutions in a defined neighborhood. Nevertheless,
the particular knowledge of any scheduling problem is not
considered if the neighborhood is defined in a general way.
This does not happen with greedy heuristics.
In this work, a local search procedure is proposed
including the positive aspects of both types of heuristics:
1) the knowledge about the RCPSP offered by the priority
rules of the problem and, 2) the possibility of generate
solutions in the search space. For this purpose, the
solutions are characterized by sequences of priority rules.
Each sequence of rules generates one or more solutions
following a simple algorithm that optimize the makespan
(total time needed for the real execution of the task). A
Genetic Algorithm is applied to generate the solutions
using crossovers, mutations and regeneration of different
priority-rules sequences.

2 Local Search Heuristics


Local search methods (TS, SA, GA, etc) are used to
explore a solution neighborhood. A typical way to define a
neighborhood in a scheduling problem is the interchange
of tasks. This is a general procedure that does not use the
specific information about the problem.
Other approaches to the definition of a neighborhood use
the relation between a heuristic h and the solution s
obtained applying h to a problem p, i.e. h(p) = s [16]. This
relation allows the determination of neighborhoods in both
the problem space and the heuristics space.

In order to obtain a neighbor in the problem space the


following actions are done: 1) introduce a random
perturbation (within some range) in the data of the problem
(e.g. change in a 10% the duration of the assembly time of
each part), 2) one particular heuristic is applied to the new
data to obtain a dummy solution (i.e. a dummy sequence
to assembly all the parts), 3) the dummy solution is
evaluated (i.e. the makespan is computed) with the original
data of the problem.

51-59, 61-70, 83, 89, 91, 92, 96 and 100. The use of other
rules produces solutions between 63 and 71 seconds (the
worst case for rule #97, i.e. [k] = rule_#97 k).
Nevertheless, in this example it is easy to find an optimal
solution, like any of those in Figure 2. This simple
example does not show that the rules are inadequate but
that the way they are applied is not optimal.
Task
B1
C1
D1
E1
F1
G1
E2
F2
G2
B2
C2
D2

The definition of neighborhoods in the heuristics space is


done by developing parameterized variations of the set of
specific heuristics of the problem. This can be done in at
least two ways in the RCPSP:
1) Defining a new hybrid rule as a linear combination of
the original dispatching rules i, i.e. =i i i.
2) Dividing the dispatching into ordered subsets of rules
(e.g. the three first tasks will be dispatched by rule #2,
the next two tasks by rule #7, etc). An extreme case of
this approach is that each decision in the dispatching is
characterized by one particular rule, i.e. for a problem
with N tasks the dispatching is controlled by the vector
r = ([1], ...,[k], ..., [N]), where [k] is the rule applied in
the decision k (note that the rule.[N] is irrelevant, but
we include for homogeneity).

Solution 1
0
Rob.1

B1

00
11
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11

C1

E1

1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111

G1

F1

15
B2

Rob. 2

E2

F2
E2 1111
0000
0000
1111
1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111

G2

F2

D1

D2

C2

G2

23

16

60

54

G1
51

40

60

Solution 2
Rob.1

15
B2

Rob. 2

E2
10

F2

C1
16

D2

C2

E1
B1

23

54

40

23

33

D1

G2

F1
40

60

G1
51

60

Figure 2: Two optimal solutions (makespan=60s) for the


assembly problem of Figure 1 showing the task assigned to each
robot and the total time after each assembly operation.

D2
00
11
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11
00
11

40

F1

C1

B1
10

30

23
E1

D1

Precedence
B1
C1
E1
F1
E2
F2
B2
C2

Table 1

3 Assembly example
Figure 1 shows a set of 12 parts to be assembled on the
base A by two robots of the same type (the parts Ci, Gi and
Di i{1,2} act as fasteners). Table 1 summarizes the time
(in seconds) needed for the assembly of each part and the
precedence relations between them.

Duration
6
7
6
8
7
9
10
10
11
15
17
14

B2

C2

4 Basic Dispatching Algorithm


Given a vector of rules r = ([1], ...,[k], ..., [N]), the task
scheduling solution is directly obtained using the algorithm
A1 described below.

Figure 1: Assembly to be performed using two robots.

Nomenclature:
A first attempt to determine the assignment of tasks to each
robot was done using a procedure based on the parallel
dispatching of 100 well known rules. The set of rules,
listed in Appendix A, includes, for instance, SIO (Shortest
Imminent Operation), GRD (Greatest Resource Demand),
Weighted Resource Utilization Ratio and Precedence
(WRUP), Minimum Job Slack (MINSLK), among others.
The best solution without any subsequent local
optimization has a makespan of 62 seconds, obtained with
vectors uniquely composed by any of the rules: 4-13, 27,

N
M
i
j
k
T
P(i)
C(i)
(i)
R(i,j)

number of tasks (components to be assembled).


number of resource types (robots, pallets, etc).
task index, 1iN.
resource index, 1jM.
scheduling decision index, 1kN.
dispatching time.
duration of task i.
ending time of task i.
set of precedence tasks of task i.
number of units of resource j required by task i.

R(j)
X
Y
Z
z*
W
S
[k]
C
Cmax

number of available units of resource j (initialized


as R0(j) ).
set of tasks to be scheduled (initialized as X0).
set of tasks satisfying the precedence constraints
(YX)
set of tasks satisfying the precedence constraints
and resources availability (ZY).
first task of Z
set of tasks being executed.
set of tasks being executed with closer ending time.
rule of the scheduling decision k in vector of rules r.
maximum ending time of the scheduled tasks.
time needed to solve all the tasks (makespan).

Begin A1
1. Initialize:
T0
k1
C(i)
i (1iN)
R(j)R0(j) j (1jM)
XX0
W
2. Create Y:
Y ={yX : (C(x)T x(y) ) (y)= }.
3. Create Z:
Z ={zY : R(z,j)R(j) j}
IF Z= GO TO 6
4. Schedule the task:
Arrange Z according to rule [k]
C(z*)=T+P(z*)
R(j) R(j)-R(z*,j) j
WW+{ z*}
C = max[C(w)] with wW
YY-{ z*}
XX-{ z*}
IF X= GO TO 7
5. Increment the decision pointer:
k k+1
GO TO 3
6. Release of resources:
Search S={ s : C(s) = min[C(w)] with wW }
R(j)R(j) + sS R(s,j) j
T T+ C(s) with sS
WW S
GO TO 2
7. Determine Cmax:
Cmax = max[C(w)] with wW
End A1
The application of the algorithm A1 to the assembly
example previously described in Section 3 with the vector
of rules r = (40,40, 7, 7, 7,40,40,40,40,40,40,40) gives the
solution 1 shown in Figure 2. It is interesting to remark

that using rules #7 and #40 independently the obtained


solutions last 62 and 63 seconds respectively.
In general, given a vector of rules r and an algorithm A it is
possible to define a heuristic h from the pair (r, A), i.e. h =
h(r, A).
The algorithm A1 can be used for any local search
procedure that allows the generation of neighbor solutions
in the heuristics space. Then, the rules are altered instead
of the tasks.

5 The Genetic Algorithm


The generation of solutions in the heuristics space (i.e.
vectors r = ([1], ...,[k], ..., [N]) to be used by A1), was
done using the genetic algorithm GA1 described below.
Nomenclature:
I
L
p
r
h
s
r
h
s
r
h
s
r
ri
hi
si

number of individuals (vectors of rules) in the


population.
number of iterations (generations).
instance of the problem to be solved.
population of ancestors of the sequences of rules.
population of ancestors of the heuristics.
population of ancestors of the solutions.
population of descendants of the sequences of rules.
population of descendants of the heuristics.
population of descendants of the solutions.
population of mutated descendants of the sequences
of rules.
population of mutated descendants of the heuristics.
population of mutated descendants of the solutions.
population of eligible sequences of rules for the next
iteration (generation).
element i of the sets r , r , r and r.
element i of the sets h , h , h and h.
element i of the sets s , s , s and s.

Begin GA1
Phase A: Initialization
0. Generation of initial populations:
0.1 Generate the initial r of I as:
r ={ ri = ([1],..., [N]) : [1]= ...= [N]}
0.2 Generate the initial population of heuristics:
h = { hi = hi(ri,A1) : ri r}
0.3 Generate the population of solutions of p and
evaluate their makespan:
s = { si = hi(p) : hi h}
0.4 Save as heuristic and incumbent solution the pair
(h*,s*) with the best makespan.
0.5 Determine the fitness fj of the elements of s as:
( D j Dmin ) 1
fj = I
1
i =1 ( Di Dmin )

with:
Di
Dmax
Dmin

6 Experimental Results
makespan of the solution i
greatest makespan of the population
lowest makespan of the population
index of the population homogeneity
1 I Di D min
=
I i =1 D max D min

Phase B: Iterate through the following steps L times:


1.

Selection of ancestors:
Build I/2 pairs of elements of r according to the
fitness of the elements of s.
2. Choice of the pairs for the crossover:
2.1 Determine the probability of the current
crossover: Pc = Pc().
2.2 Assign a random number to each pair of
sequences of rules.
2.3 Decide, for each pair of sequences of rules, if a
crossover should be done according to their
random number and Pc.
3. Generation of descendants:
3.1 Crossover the selected pair of sequences of rules
to generate two descendants, creating r.
3.2 Generate h and s from r as it was done in
0.2 and 0.3 respectively.
3.3 Determine the makespan of the elements of s If
any element of s has a better makespan than
the incumbent solution, then save as heuristic
and incumbent solution the pair (h*,s*)
associated to that element.
4. Mutation of descendants:
4.1 Determine the probability of mutation of the
current generation: Pm = Pm().
4.2 Assign a random number to each element r.
4.3 Decide the elements of r to be mutated
according to their random number and Pm.
4.4 Mutate the chosen elements of r creating r.
4.5 Generate h and s from r as it was done in
0.2 and 0.3 respectively.
4.6 Determine the makespan of the elements of s.
If any element of s has a better makespan than
the incumbent solution, then save as heuristic
and incumbent solution the pair (h*,s*)
associated to that element.
5. Regeneration of the population:
5.1 Build the population of eligible elements
r r + r + r
5.2 Determine the fitness of the elements of the
populations s and s as it was indicated in 0.5.
5.3 Choose I elements from r according to the
fitness of the elements of s , s and s.
End GA1

In order to validate the proposed approach, 270 different


instances of the task sequencing problem have been solved
considering:
Number of tasks: 6 < N < 15.
Types of resources (robots, pallets, etc): 1 M 3.
Units of resource j (e.g. number of robots) 2 Ro(j) 5.
Duration of task i: 1 P(i) 16.
6 different ratios #precedence/N.
The proposed GA1 has obtained the optimum solutions of
the 270 instances of the problem in less than 12 minutes
using a Pentium II 233MHz.
The following subsections detail some particular aspects of
the implemented algorithm.
6.1 Initial Population
A population with size I =100 was used with the aim of
considering all the heuristics derived from the rules shown
in Appendix A. Then, the initial population is composed of
100 vectors of rules, each one composed by one particular
rule in all its components.
This allows the exploration of all the solutions generated
by the greedy heuristics. In order to increase the size of the
initial population it is only necessary to include new rules
or generate hybrid rules using linear combinations of
previous rules.
The random selection of the task to be dispatched has also
been incorporated as a rule (rule 26), to allow the
generation of any solution. The rule including random
selection is necessary when the set of rules does not
guarantee the generation of all the solution space.
6.2 Selection Process
The elements of r are randomly selected with larger
probabilities for those elements with better fitness.
6.3 Probabilities of Crossover and Mutation
The probability of any crossover or mutation, Pc and Pm
respectively, depends on the homogeneity index of the
current population. In this way, a population with quite
similar individuals will be modified through mutations
because the crossovers would not be effective for
diversification. The following values have been used for
the experiments: Pc =1-0.5, and Pm = 0.05+0.95.
6.4 Crossover Process
Given two vectors of rules (the ancestors) two components
are randomly selected, and the rules between these
components in both of them are interchanged to obtain two

new vectors of rules (the descendants). Figure 3 illustrates


a crossover.

Ancestors

Descendants

Figure 3: Example of a crossover.

6.5 Mutation Process


Three different types of mutation of a vector of rules have
been considered: soft, medium and hard. The probability
of each type of mutation is given a priori.
Soft mutation: two components of the vector of rules are
randomly selected and interchanged.
Medium mutation: one rule of the vector is randomly
selected, this rule is successively replaced by all the
available rules and the combination that generates the
lowest makespan is selected.
Hard mutation: it is equivalent to test all the possible soft
mutations of the vector of rules and select the one with the
lowest makespan.
6.6 Regeneration Process
The elements of r are randomly selected giving higher
priority to those sequences of rules with better fitness.

7 Conclusions
A method to look for solutions of the RCPSP with
application to the scheduling of assembly operations with
limited resources (the robots) has been presented. The
main contribution of the method is the incorporation of the
knowledge provided by the specific heuristics of the
problem in a local search procedure. In this way, the
solution is characterized by a sequence of priority rules.
The method has been implemented using a genetic
algorithm. The experiments and computational experiences
were quite satisfactory.

Appendix A: List of Rules


Nomenclature:
P(i)
R(i,j)
R0(j)
Z

duration of task i.
number of units of resource j required by task i.
number of available units of resource j
set of tasks satisfying the precedence constraints
and resources availability.
ns(i) number of direct successors
nst(i) number of successors
ih h is a direct successor of i.
ih h is a successor of i.
EST Earliest Start Time.
LST Latest Start Time.
EFT Earliest Finish Time.
LFT Latest Finish Time.
Schedule the task z* : v(z*)=maxiZ[v(i)]

6.7 Frequency of the Rules


The frequency of the 100 rules in the 270 optimum
solutions was also determined (Figure 4). The initial
frequency of each rule is 1%, but the results show that,
after the evolutionary process, some rules appear with
higher frequency; in particular, it is interesting to note the
6% of rule 26 (random selection of a task).
7

NAME
1. SIO Shortest
Imminent Operation.

v1 (i ) = P (i )

2. GRD Greatest
Resource Demand.

v2 (i ) = P (i ) R (i, j )

3. GRPW Greatest Rank


Positional Weight.

v 3 ( i ) = P (i ) P ( h )

4-14*. WRUP Weighted


Resource Utilization
Ratio and Precedence.

RULE:

j =1

ih

v4 (i ) = w p ns(i ) + wr

R ( j)

R (i , j )

j =1

v5 (i) = w p P (h) + wr

26. ALEA.

v6 (i ) = Random(i )

27. MTS Most Total


Successors

v7 (i ) = nst (i )

28-38*. WRUP3

v8 (i ) = w p P(i ) + wr

39-49*. WRUP4

v9 (i ) = w p P (i ) + v5 (i ) = w p

50-60*. WRUP5

v10 (i ) = w p nst (i ) + wr

4
[%]

R(i, j )
j =1 R0 ( j )

15-25*. WRUP2

ih

97

100

94

91

88

85

82

79

76

73

70

67

64

61

58

55

52

49

46

43

40

37

34

31

28

25

22

19

16

13

10

R (i, j )
j =1 R0 ( j )

P (h) + v8 (i )

ih

Rules

Figure 4: Final frequency of the 100 rules in the experiments.

j =1

R (i, j )
R0 ( j )

R (i, j )
j =1 R0 ( j )

61-71*. WRUP6

v11 (i ) = w p P (h) + wr

72-82*. WRUP7

v12 (i ) = w p P(i ) + v11 (i ) = w p

83. MIT Most Immediate


Successors

ih

P( h) + v8 (i )

ih

v13 (i ) = ns (i )

P(h)

84. MIT2

v14 (i ) = ns (i ) +

85. MIT3

v15 (i ) = ns(i ) P (i )

86. MIT4

v16 (i ) = ns (i) P(i ) + P(h)


i h

87. MIT5

v17 (i ) = ns (i ) + P(i ) R(i, j )

ih

j =1

88. MIT6

v18 (i ) = ns (i ) P (i) R(i, j )

89. MIT7

v19 (i ) = nst (i ) +

90. MIT8

v20 (i ) = nst (i ) P(i )

91. MIT9

v21 (i ) = nst (i ) P (i) + P (h)


i h

92. MIT10

v22 (i ) = nst (i ) + P (i ) R (i, j )

j =1

P(h)

i h

j =1

93. MIT11
94. LST Latest Start
Time
95. EST Earliest Start
Time
96. LFT Latest Finish
Time
97. EFT Earliest Finish
Time
98. MINSLK Minum Job
Slack
99. RSM Resource
Scheduling Method
100. RSM2

v23 (i ) = nst (i ) P (i ) R (i, j )


j =1

v24 (i ) = LST (i )
v25 (i ) = EST (i )
v26 (i ) = LFT (i )
v27 (i ) = EFT (i )
v28 (i ) = (LST (i ) EST (i ) )

v29 (i ) = max 0, min (EFT (i ) LST (h) )


hZ {i}

v30 (i) = max[0, (EFT (i ) LST (h) )]


hZ {i }

*A different rule is considered for each value of wr such


that wr {0, 0.1,...,0.9,1}, wp =1- wr .

References
[1] zdamar, L. & Ulusoy, G. (1995): Survey on the
resource-constrained project scheduling problem. IIE
Transactions, 27 (5), 574-586.
[2] Weglarz, J. (Ed.) (1998): Handbook on Recent
Advances in Project Scheduling. Kluwer, Amsterdam.
[3] Kim, J. Desrochers A. & Sanderson, A. (1995): Task
Planning and Project Management using Petri Nets. 1995
IEEE International Symposium on Assembly and Task
Planning, Pittsburgh, Pennsylvania, 265-271.
[4] Patterson, J.H. (1984): A comparison of exact
approaches for solving the multiple constrained resource,
project scheduling problem. Management Science, 30 (7),
854-867.

[5] Simpson, W.P. & Patterson, J.H. (1996): A multipletree search procedure for the resource-constrained project
scheduling problem. European Journal of Operational
Research, 89, 525-542.
[6] Blazewicz, J.; Lenstra, J.K. & Rinnoy Kan, A.H.G.
(1983): Scheduling projects to resource constraints:
Classification and complexity. Discrete Applied
Mathematics, 5, 11-24.
[7] lvarez-Valds, R. & Tamarit, J.M. (1989): Heuristic
algorithms for a resource constrained project scheduling: A
review and an empirical analysis, Advances in Project
Scheduling, 113-134. R. Slowinski & J. Weglarz (Ed.).
Elsevier, Amsterdam.
[8] Kolisch, R. (1996): Efficient priority rules for the
resource-constrained project scheduling problem. Journal
of Operations Management, 14, 179-192.
[9] Daz, A. & Glover, F. & Ghaziri, H. & Gonzlez, J.M.
& Laguna, M. & Moscato, P. & Tseng, F. (1996):
Optimizacin heurstica y redes neuronales. Paraninfo.
[10] Holland, J.H. (1975): Adaptation in Natural and
Artificial Systems. University of Michigan Press, Ann
Arbor, Mich.
[11] Goldberg, D.E. (1989): Genetic Algorithms in Search,
Optimization and Machine Learning. Addison-Wesley,
Reading, Mass.
[12] Padman, R. & Roehrig, S.F. (1997): A Genetic
Programming Approach for Heuristic Selection in
Constrained Project Scheduling, Interfaces in Computer
Science and Operations Research: Advances in
Metaheuristics, Optimization, and Stochastic Modeling
Technologies, 405-421. R.S. Barr & R.V. Helgason & J.L.
Kennington (Ed.). Kluwer, Norwell, MA, USA.
[13] Gyoung K. & Lee C.S.G. (1994): An Evolutionary
Approach to the Job-Shop Scheduling Problem. 1994
IEEE International Conference on Robotics and
Automation, San Diego, CA, 501-506.
[14] Mori, M. & Tseng, C.C. (1997): Genetic Algorithm
for multi-mode resource constrained project scheduling
problem. European Journal of Operational Research,
100, 134-141.
[15] Fujimoto H., Yasuda K., Tanigawa Y. & Iwahashi K.
(1995): Application of Genetic Algorithm and Simulation
to Dispatching Rule-Based FMS Scheduling. 1995 IEEE
International Conference on Robotics and Automation,
Nagoya, Japan,V1, 190-195.
[16] Storer, R.H. &Wu, S.D. & Vaccari, R. (1992): New
search spaces for sequencing problems with application to
Job Shop Scheduling. Management Science, 38 (10),
1495-1509.

You might also like