You are on page 1of 6

Meta-Heuristic for No-Wait Job Shops with Makespan Minimization

Jie Zhu *, Xiaoping Li, Yi Zhang, Qian Wang


School of Computer Science &Engineering, Southeast University, Nanjing, P.R .China
Key Laboratory of Computer Network and Information Integration (Southeast University),
Ministry of Education
zhujie_hades@163.com; xpli@seu.edu.cn; zhangyiprivate@yahoo.com.cn;
qwang@seu.edu.cn

Abstract no-wait job shops with makespan minimization, among


which VNS (variable neighborhood search) [9], GASA
In the paper, the no-wait job shop problem with (hybrid simulated annealing/genetic algorithm) [9] and
makespan minimization is considered, which is CLM (complete local search with memory) [10] seem to
decomposed into the sequencing problem and the be the most effective or efficient ones. Among these
timetabling problem. Based on the non-delay approaches, the problem is usually decomposed into
timetabling procedure and the inverse timetabling two sub-problems: timetabling problem and sequencing
procedure, an enhanced timetabling procedure is problem. VNS depends on special structure of the
constructed by shifting jobs leftwards or rightwards to solved problem and its speed is very fast. GASA is a
obtain better timetables. The two sub-problems are non-deterministic hybrid algorithm, which is more
solved independently by traditional methods. However, time-consuming but more effective than VNS. Both
a meta-heuristic algorithm MCLM (modified complete VNS and GASA adopt the Non-Delay Timetabling to
local search with memory) is presented to solve the tackle the first sub-problem. For the second
sub-problems integrally in this paper. Experimental sub-problem, VNS searches the optima results
results show that MCLM outperforms all the existing iteratively in a k-insertion neighborhood until no better
effective algorithms for the considered problem with solution can be found. GASA integrates Genetic
little more computation time. Algorithm (GA) with Simulated Annealing (SA) for the
sequencing problem. CLM [10] also takes advantage of
Keywords: No-wait, Job shop, Timetabling, the non-delay timetabling procedure. Furthermore, the
Sequencing. inverse timetabling procedure is introduced, in which
the non-delay timetabling is performed on “reversed
1. Introduction jobs” (the processing route of every job is reversed). For
In this paper, no-wait job shops with makespan the sequencing problem, CLM generates neighborhood
minimization are considered, in which two consecutive by 1-insertion and all the exploited solutions are
operations of each job must be performed continuously recorded to avoid re-calculating. It seems that CLM is
without any interruption. No-wait constraints usually the best algorithm so far for the considered problem.
arise from requirements for processing environments or In this paper, algorithm MCLM (modified complete
characteristics of jobs. Typical examples are iron being local search with memory) is presented. Instead of the
immediately staunch while hot in metallurgical two sub-problems being treated separately as in CLM,
processes, unstable intermediate products or absence of VNS and GASA, they are solved by an integrated
intermediate storage capacity in chemical processes. method in MCLM. An enhanced timetabling procedure
Such job shop problems widely exist in many real-life is constructed according to the non-delay timetabling
applications, such as pharmaceutical industry [1], steel and the inverse timetabling. Then the enhanced
production [2], computer systems [3], scheduling of timetabling procedure is embedded into a complete
trains [4-6] and the flow control of packet switching local search to solve the sequencing problem.
communication networks [4]. The rest of the paper is organized as follows. The
The traditional job shop problem is NP-hard in strong considered problem is described in Section 2. Section 3
sense [5]. Detailed surveys have been conducted in [6] constructs the enhanced timetabling approach.
and [7] for the results on this problem in the past Combined with the enhanced timetabling approach, a
decades. However, little attention has been paid on the new local search approach is proposed in Section 4. In
problem considered in this paper. The review given by Section 5, the experimental results are shown, followed
Hall N.J. & Sriskandarajah C. [8] showed that the by conclusions in Section 6.
considered problem is difficult, especially for large-size
instances. 2. Problem Description
Recently, several algorithms have been presented for A no-wait job shop is a scheduling problem in which

_____________________________________
978 -1-4244-1651-6/08/$25.00 © 2008 IEEE


n jobs are processed on m machines and each job Inverse timetabling procedure generates a new
has a specific processing route. For job J i , rij sequence for the given sequence in terms of inverse
processing sequence and inverse instance. Then, the
( 0 d i d n  1 , 0 d j d m  1 ) indicates the machine
new sequence and inverse instance is applied to the
which the j th operation has to be processed on, pij non-delay timetabling procedure to obtain a timetable.
( 0 d i d n  1 , 0 d j d m  1 ) states the processing time The makespan can be calculated according to the
timetable.
of job J i on machine M j . Li ( i 0,  , n  1 )
indicates the aggregated processing times of J i , i.e. 3.2 Enhanced Timetabling
m 1 Definition 4 Left shift tries to start a job as early as
Li ¦
j 0
pij . Processing sequence S (S [ 0] ,..., S [ n1] ) , possible without ordinal constraint.
Definition 5 Minimal interval between jobs J i
in which S [i ]  ^J k | 0 d k d n  1` is the i th ^ `
and J j is min t j  ti in which J j is scheduled after
( i 1,  , n ) job in S and tS [i ] denotes the starting J i ( i.e. t j t ti ) as close as possible.
time of S [i ] . Furthermore, the constraints of sequence, Definition 6 Safe interval between J i and job J j
synchronicity and no-wait [10] should be hold. ^ `
is min t j  ti , where J i starts earlier than J j on
For the considered problem, which is NP-hard in
each machine (see Figure 1).
strong sense, effective algorithms so far usually
24 58
decompose it into two sub-problems: (1) Sequencing 1 2
problem, in which a processing sequence of an optimal 18 72
schedule is found for a given no-wait job shop problem. 1 2
32 23
(2) Timetabling problem, in which a feasible set of
1 2
starting times of the jobs is found to realize the 81 99
minimum makespan for the processing sequence 1 2
obtained from (1). 25 28
1 2
Both the sub-problems have been proved to be
99
NP-hard in strong sense [25]. It is commonly believed Safe interval
that the sequencing problem is much more difficult than Fig. 1 Safe interval of J1 and J 2
the timetabling one.
For the considered n jobs, minimal interval and
3. Enhanced timetabling approach safe interval of each pair of jobs can be predetermined
3.1 Non-Delay Timetabling and Inverse and recorded in minimal interval matrix mij nun and > @
Timetabling
Schuster and Framinan [9] described the Non-Delay safe interval matrix sij > @nun .
Timetabling, which is easy and fast to construct a For job J j and job J i , if J i precedes to J j ,
timetable.
there are three cases for the starting time distance
Definition 1 Ordinal constraint is tS[ 0 ] d ... d tS[ n 1]
t j  ti . (1) t j  ti t sij , no intersection (Gantt charts
for processing sequence (S [ 0] ,..., S [ n1] ) . overlap on some machines) exists in the timetable of
The non-delay timetabling procedure starts each job J i and J j . (2) t j  ti  mij , intersection must exist in
as early as possible without violating ordinal constraint.
Framinan and Schuster [10] develop the inverse the timetable. (3) t j  ti is called uncertain interval
timetabling to further improve the non-delay when mij  t j  ti  sij and it is hard to determine
timetabling.
whether there exists intersection or not. Therefore,
Definition 2 Inverse instance is an instance identical
uncertain interval is not taken into consideration to
to the original one which is defined by n , m , pij
avoid enormous complex cases in the paper. Actually,
and rij , but each job has the inverse processing route, minimal interval equals to safe interval in some
situations.
i.e. rij' ri , m 1 j ( j 0,  , m  1) in routing matrix Definition 7 Right shift tries to start a job as early as
>r @
'
ij num and pij' pi , m 1 j ( j 0,  , m  1) in possible with at least safe intervals to all arranged jobs.
Because it is impossible to decide the effect of every
processing time matrix > @
pij' nu m . job on the makespan minimization immediately, only
Definition 3 Inverse processing sequence partial jobs should be verified to help the above
(S '[ 0] ,..., S '[ n1] ) is obtained from processing sequence selection. In this paper, a processing sequence S is
divided into k ( 2 d k d ªn / 2º ) sub-sequences. In
(S [ 0] ,..., S [ n 1] ) by setting S '[ j ] S [ n 1 j ] .
each sub-sequence, there are no more than ªn / k º jobs


and at least 2 jobs except the last sub-sequence. The current sequence S i is computed by
first job in each sub-sequence is called head job while
the rest are member jobs. A member job can only be left C ' m max ® max
­
¯ j p ,, p  q 1
^ `
½
tS' > j @  LS > j @ , Cmax ¾ .
¿
shifted. That whether the head job is left shifted or right
shifted depends on the effects on makespan 3.5 Re-schedule the i th sub-sequence by right
minimization. The head job is left shifted if a better shifting the head job while left shifting the
makespan can be obtained than right shifted, otherwise member jobs sequentially. Calculate starting
the head job should be right shifted. By this way, the times tS' ' > p @ ,  , tS' ' > p  q 1@ . Makespan of current
right shifted jobs are selected from the k head jobs.
For a given sequence, the proposed Enhanced sequence Si is recomputed by
Timetabling Method generates timetable of the left shift
and right shift jobs. The sequence is divided into k
'' ­
C m max ®
¯j
max
p ,, p  q 1
^t ''
S > j@  LS > j @ ` ½
, Cmax ¾ .
¿
sub-sequences. The first sub-sequence is assigned as the
3.6 If C ' ! C '' , then Cmax m C '' ,
current schedule by left shifting all of its jobs. For the
'' ''
i th ( i 2,  , k ) sub-sequence, two trials are performed. ( tS > p @ ,  , tS > p  q 1@ ) m ( tS > p @ ,  , tS > p  q 1@ );
All member jobs are left shifted with the head job left
Else, Cmax m C ' ,
shifted and right shifted respectively and two schedules
' '
are generated. If the schedule made by right shifting the ( tS > p @ ,  , tS > p  q 1@ ) m ( tS > p @ ,  , tS > p  q 1@ ).
head job of the current sub-sequence is better than that 3.7 i m i  1 .
by left shifting, the head job is chosen as a right shift
job. The better of the two schedules is reassigned as the 4. Until i ! k .
current schedule. 5. C max is the final makespan and ( tS >0 @ ,  , tS >n1@ ) is

For the current schedule S [0] ,  , S [i 1] , if an the final timetable. Stop.
unscheduled job J S >i @ is right shifted, starting time Obviously, the time complexity of right shift for a job
S >i @ is 2(i ) and that of left shift for a job S >i @ is
^
tS >i @ can be calculated by tS [ i ] m max tS >k @  sS >k @S >i @ ` 2(mi ) .Therefore, the time complexity of this algorithm
(k 0,  , i  1 ).
is 2(n 2 m) .
Left shift is a little more complicated than right shift.
By calculating the maximum length of overlap between

S [i ] and S [0] ,  , S [i 1] , S [i ] is moved rightwards 4. Modified CLM
Ghosh and Sierksma [11] constructed CLM
for that length. The calculating and moving process is
(complete local search with memory) algorithm, in
repeated until the maximum overlap length is zero. S [i ] which all explored solutions are recorded in a memory
is placed to the found idle gaps and tS >i @ is obtained. with an explicit manner to avoid re-calculating.
However, the size of memory increases with the size
The enhanced timetabling procedure is described as
of the problem in CLM. For large size problems, large
follows.
memory is needed and it is very time-consuming for
1. tS >0 @ m 0 , k m c (constant c 2,  , ªn / 2º ), checking a new sequence in the memory or not.
i m 1 , Cmax m 0 , S 0 m . Therefore, the memory size of CLM should be limited
and a strategy should be introduced to refresh contents
2. Divide the sequence S into k sub-sequences. of the memory. In this paper, MCLM (modified
3. Repeat complete local search with memory) is presented by
3.1 p m (i  1) u ªn / k º . //* S [ p ] is the head job of modifying CLM.
the i th sub-sequence*// MCLM maintains three sets: (1) NEWGEN, a limited
3.2 If ik then q m ªn / k º , else storage for good enough solutions being employed by
the heuristic during the current iteration. (2)
q m n  (k  1) ªn / k º . //*Number of jobs in
NEIGHBOR, a limited storage for the neighbors of the
the i th sub-sequence*// solutions in NEWGEN. (3) DEAD, a limited storage for

3.3 S i m S i 1  S > p @ ,  , S > p q1@ . //* S i is the storing solutions that have been already explored, in
current sequence, the sub-sequence of S order to avoid re-calculating. NEWGEN and
NEIGHBOR can be neither too large (which is rather
from S [ 0] to S [ p q 1] , by appending the i th
time-consuming and space-consuming) nor too small

sub-sequence S > p @ ,  , S > p q 1@ to S i 1 *// (by which the algorithm always traps in local optimal
solutions in short time). DEAD is emptied once it is full
3.4 All jobs in the i th sub-sequence are left
shifted sequentially and their starting times in this paper. S NEW , S NEIGHBOR and S DEAD denote
the size of NEWGEN, NEIGHBOR and DEAD
tS' > p @ ,  , tS' > p q 1@ are obtained. Makespan of
respectively.
MCLM starts with a given sequence as the initial


solution and puts it into NEWGEN. DEAD and predefined constants I1 , I2 and V
NEIGHBOR are initially empty. The following process ( 0  I1  I 2  N max , V  0,1 ). N max is the
is iterated:
(1) Generating neighbors for all solutions in NEWGEN predefined maximal number of iterations without
and putting them into NEIGHBOR if NEIGHBOR improvement, I max is the maximal number of
is not full. Partial pair-wise exchange and iterations the algorithm can perform. Certain
1-insertion are performed on every solution of constant I add is added to it once improvement has
NEWGEN to generate its neighbors. Partial been made.
pair-wise exchange generates neighbors by just The above iterative process is repeated until I max
swapping the head job and its member jobs
respectively for every sub-sequence (see Figure 2). iterations have been performed, or no improvement has
The simple 1-insertion only considers k been made for N max successive iterations, or
sub-sequences, which are regarded as k units. NEWGEN is empty. MCLM is formally described as
Each unit is picked out and inserted respectively follows.
into every possible unit slot (except the original slot) 1. NEIGHBOR m I , DEAD m I . Generate an
and k  1 neighbors are produced (see Figure 3). initial sequence S 0 . NEWGEN m ^S 0 ` .
The set of neighbors for solution S generated by 2. i m 0 , S best m S 0 , I temp m 0 , I max m ªn 2 / m º ,
the above procedure is denoted as GN (S ) .
W m W max .
Sub1 … 3,2,4,5,6 … Subk 3. Calculate timetable T0 and makespan C0 by
Partial pair-wise Sub1 … 4,3,2,5,6 … Subk calling the Enhanced Timetabling Procedure.
Sub1 … 2,3,4,5,6 … Subk Tbest m T0 , Cbest m C0 .
exchange Sub1 … 5,3,4,2,6 … Subk
4. Repeat
Sub1 … 6,3,4,5,2 … Subk 4.1 flag m False .
4.2 For each S k  NEWGEN , if
Fig. 2 Partial pair-wise exchange
NEIGHBOR  S NEIGHBOR ,
Sub2 Sub1 Sub3 NEIGHBOR m NEIGHBOR ‰ GN S k .
Sub2 Sub3 Sub1 4.3 NEWGEN m I .
4.4 For each S j  NEIGHBOR
Sub2 Sub1 Sub3
1-Insertion
Sub1 Sub2 Sub3 4.4.1 If S j  DEAD , compute T j by the
Sub1 Sub3 Sub2
Enhanced Timetabling Procedure and
Sub3 Sub1 Sub2
calculate the makespan C j for S j ;
Sub1 Sub3 Sub2
else CONTINUE. //* Go to 5.4 to deal
with the next sequence in NEIGHBOR *//.
Fig. 3 1-insertion
4.4.2 If C j  Cbest , Cbest m C j ,
(2) Empting NEWGEN. Tbest m T j ,
(3) Checking solutions of NEIGHBOR belonging to
DEAD or not. Among all neighbors not in DEAD,
NEWGEN m NEWGEN ‰ S j ^ ` ,
those with makespan lower than a certain threshold S best m S j , flag m True ;
value are transplanted to NEWGEN for the next else if C j  Cbest u 1  W ,
iteration and the others are put into DEAD.
Cbest u (W  1) is set as the threshold value in this NEWGEN m NEWGEN ‰ S j ^ `
paper ( W  0, W max @ ). Cbest is the best-so-far ;
makespan and W max is predefined. I temp records else DEAD m DEAD ‰ S j . ^ `
the number of consecutive iterations without 4.4.3 If NEWGEN ! S NEWGEN , BREAK . //*
improvement. Boolean flag is assigned True Break the loop and go to 5.5 *//
once improvement has been made in the iteration. 4.5 NEIGHBOR m I . If DEAD ! S DEAD ,
(4) Empting NEIGHBOR. DEAD m I .
(5) If no improvement has been made for Rmax
4.6 If flag False , I temp m I temp  1 ; else
successive iterations ( Rmax  I temp ), the current
I temp m 0 , I max m I max  I add .
instance is reversed.
(6) W is reset according to the value of I temp and 4.7 If I1  I temp  I 2 and W / V d W max ,


W m W / V ; else if I temp t I 2 , W m W u V . Therefore, MCLM outperforms VNS, GASA and
CLM in effectiveness. Though the computation time
4.8 If I temp ! Rmax , inverse the instance.
needed by MCLM is as about 4-5 times as that by CLM
4.9 i m i  1 . and much more than that by VNS or GASA, it is
5. Until i ! I max or I temp ! N max or affordable for practical applications with the respect to
the results obtained by MCLM.
NEWGEN 0.
6. ( Cbest , S best , Tbest ) is the final result. Stop. 6. Conclusions
Because NEWGEN , NEIGHBOR and I max may In this paper, MCLM (modified complete local search
with memory) method is proposed for no-wait job shop
change during the iterating, it is hard to decide the problems with makespan minimization. Similar to
complexity of the algorithm. traditional algorithms, the considered problem is
decomposed into two sub-problems: the timetabling
5. Computational experiments problem and the sequencing problem. Based on the
To compare the proposed MCLM with the best so far non-delay timetabling procedure (adopted by VNS and
algorithms CLM [10], VNS and GASA [9], MCLM and GASA) and the inverse timetabling procedure (used by
CLM are coded in Java and ran on a PC with Pentium 4 CLM), an enhanced timetabling procedure is
processor. The algorithms are tested on 10 small and 15 constructed by shifting jobs leftwards or rightwards to
big well-known no-wait job shop instances from [12-16] manage to obtain better timetables. The enhanced
and every instance has been performed 20 times. timetabling procedure can solve the timetabling
A series of preliminary experimental results show that problem effectively and efficiently. Different from
selection of the initial solution pays little influence on CLM, the sequencing problem is solved by MCLM with
performance of MCLM. Traditionally, there are eight limited memory to improve efficiency. MCLM
initial solution generating methods: SPT heuristic, LPT outperforms all the other three compared algorithms.
heuristic, SMT heuristic, LMT heuristic, SOT heuristic, Though the computation time needed by MCLM is as
LOT heuristic and RANDOM 1/RANDOM 2. For each about 4-5 times as that by CLM and even much more
testing of an instance, one of the eight methods is than that by VNS or GASA, it is affordable for practical
randomly selected to generate the initial solution of applications with the respect to the effectiveness. In
MCLM. The corresponding parameters are initialized as other words, MCLM can always find the optimal for
in Table 1. most of the instances efficiently and it is desirable for
The experimental results on small instances and big practical no-wait job shop problems.
instances are shown in Table 2 and Table 3 respectively.
OPT represents the optimal value of an instance in
Table 2 and the best-found makespan for the problem in
Acknowledgements This work is supported by
National Natural Science Foundation of China under
Table 3. BEST denotes the best makespan of an
Grants No.60672092 and No.60504029.
instance among the 20 runs for the corresponding
algorithm. For CLM and MCLM, AVG is the average
makespan of the 20 runs. TIME is the average CPU References
time in seconds over 20 runs. Only BEST and ARPD [1] W. Raaymakers, J. Hoogeveen, Scheduling
multipurpose batch process industries with no-wait
are shown for VNS and GASA in this paper.
restrictions by simulated annealing, Eur. J. Oper. Res.,
Table 1 Parameters of MCLM 126 (2000): 131–51.
Parameter Value Parameter Value [2] D.A. Wismer, Solution of the flowshop
S DEAD 200000 Rmax 10 scheduling-problem with no intermediate queues.
Operations Research, 20 (1972): 689–97.
10000 20
[3] S. Reddi, C. Ramamoorthy, A scheduling-problem.
30000 5 Operational Research Quarterly, 24 (1973): 441–6.
0.15 10 [4] C. Arbib, G.F. Italiano, A. Panconesi, Predicting
deadlock in store-and-forward networks, Networks, 20
0.9 3 (1990): 861–881.
[5] J.K. Lenstra, A.H.G. Rinnooy Kan, Computational
complexity of discrete optimization problems, Annals of
Both Table 2 and Table 3 illustrate that MCLM can
Discrete Mathematics, 4 (1979): 121–40.
find the optimal for all instances. BESTs obtained by [6] J. Blazewicz, W. Domschke, E. Pesch, The
MCLM are better than or equal to those by the other job-shop-scheduling-problem: conventional and new
three methods. Though MCLM needs more computation solution techniques, European Journal of Operational
time than CLM does in overall, there are 3 instances out Research, 93 (1996): 1–33.
of 30 that MCLM requires less time than CLM does. In [7] A. Jain, S. Meeran, Deterministic job-shop- scheduling:
other words, computation time of MCLM is not always past, present and future, European Journal of
more than that of CLM despite that MCLM is little Operational Research, 113 (2) (1999): 390–434.
more complicated than CLM. [8] N.J. Hall, C. Sriskandarajah, A survey on machine
scheduling problems with blocking and no-wait in


process, Operations Research, 44 (3) (1996): 510–25. an experimental investigation of heuristic scheduling
[9] C. Schuster, J.M. Framinan, Approximate procedures for techniques, Technical report, Carnegie-Mellon
no-wait job shop scheduling, Operations Research University, Pittsburgh, 1984.
Letters, 31 (2003): 308–18. [14] D. Applegate, W. Cook, A computational study of the
[10] J.M. Framinan, C. Schuster, An enhanced timetabling job-shop problem, ORSA Journal on Computing, 3(2)
procedure for the no-wait job shop problem: a complete (1991): 149–56.
local search approach, Computers & Operations [15] R.H. Storer, S.D. Wu, R. Vaccari, New search spaces for
Research, 331 (2006): 1200–13. sequencing instances with application to job shop
[11] D. Ghosh, G. Sierksma, Complete local search with scheduling. Management Science, 38 (1992): 1495–509.
memory, Journal of Heuristics, 8 (2002): 571–84. [16] C. Schuster, No-wait job shop scheduling: Komplexität
[12] H. Fisher, G.L. Thompson, Probabilistic learning und local search. Dissertation Thesis, University of
combinations of local job-shop scheduling rules, In: Duisburg-Essen, Germany, 2003 [in German].
Industrial scheduling. EnglewoodClif fs, NJ:
Prentice-Hall; (1963): 225–51.
[13] S. Lawrence, Resource constrained project scheduling:

Table 2 Computational results on small-size instances


VNS GASA CLM MCLM
Instance n m OPT
BEST APRD BEST APRD BEST APRD AVG TIME(s) BEST APRD AVG TIME(s)
La01 10 5 971 1064 9.578 1037 6.797 1031 6.179 1031 8 971 0.000 972 14
La02 10 5 937 1011 7.898 990 5.656 937 0.000 937 8 937 0.000 939 33
La03 10 5 820 973 18.659 932 13.659 832 1.463 832 10 820 0.000 820 12
La04 10 5 887 1020 14.994 889 0.225 889 0.225 889 8 887 0.000 887 34
La05 10 5 777 861 10.811 817 5.148 797 2.574 801 11 777 0.000 781 8
La16 10 10 1575 1673 6.222 1637 3.937 1637 3.937 1647 22 1575 0.000 1575 41
La17 10 10 1371 1556 13.494 1430 4.303 1389 1.313 1418 15 1371 0.000 1374 53
La18 10 10 1417 1713 20.889 1555 9.739 1555 9.739 1675 1 1417 0.000 1433 51
La19 10 10 1482 1786 20.513 1610 8.637 1572 6.073 1588 20 1482 0.000 1484 16
La20 10 10 1526 1795 17.628 1693 10.944 1580 3.539 1609 20 1526 0.000 1528 6
Average 14.069 6.905 3.504 0.000

Table 3 Computational results on big-size instances


VNS GASA CLM MCLM
Instance n m OPT
BEST APRD BEST APRD BEST APRD AVG TIME(s) BEST APRD AVG TIME(s)
La11 20 5 1671 1915 14.602 1825 9.216 1714 2.573 1826 109 1671 0 1716 447
La12 20 5 1452 1694 16.667 1631 12.328 1507 3.788 1507 142 1452 0 1506 498
La13 20 5 1624 1907 17.426 1766 8.744 1668 2.709 1711 123 1624 0 1661 640
La14 20 5 1691 2313 36.783 1805 6.742 1773 4.849 1773 103 1691 0 1721 465
La15 20 5 1694 1898 12.043 1829 7.969 1771 4.545 1808 126 1694 0 1749 484
La21 15 10 2048 2496 21.875 2182 6.543 2149 4.932 2208 100 2048 0 2104 306
La22 15 10 1887 2210 17.117 1965 4.134 1979 4.875 1979 53 1887 0 1912 354
La23 15 10 2032 2397 17.963 2193 7.923 2038 0.295 2178 98 2032 0 2098 307
La24 15 10 2015 2455 21.836 2150 6.700 2133 5.856 2162 73 2015 0 2048 422
La25 15 10 1917 2344 22.274 2034 6.103 2050 6.938 2075 83 1917 0 1971 297
La26 20 10 2553 3194 25.108 2945 15.354 2868 12.338 2906 248 2553 0 2707 812
La27 20 10 2747 3286 19.621 3036 10.521 2933 6.771 2970 332 2747 0 2838 834
La28 20 10 2624 3134 19.436 2902 10.595 2789 6.288 2789 266 2624 0 2752 810
La29 20 10 2489 2838 14.022 2617 5.143 2565 3.053 2630 255 2489 0 2539 778
La30 20 10 2665 3217 20.713 2892 8.518 2835 6.379 2835 414 2665 0 2743 822
Average 19.832 8.436 5.079 0.000



You might also like