Professional Documents
Culture Documents
Discrete Optimization
a r t i c l e i n f o a b s t r a c t
Article history: We propose two exact algorithms for two-dimensional orthogonal packing problems whose main com-
Received 2 August 2010 ponents are simple mixed-integer linear programming models. Based on the different forms of time rep-
Accepted 2 June 2011 resentation in scheduling formulations, we extend the concept of multiple time grids into a second
Available online 13 June 2011
dimension and propose a hybrid discrete/continuous-space formulation. By relying on events to contin-
uously locate the rectangles along the strip height, we aim to reduce the size of the resulting mathemat-
Keywords: ical problem when compared to a pure discrete-space model, with hopes of achieving a better
Optimization
computational performance. Through the solution of a set of 29 test instances from the literature, we
Integer programming
Strip packing
show that this was mostly accomplished, primarily because the associated search strategy can quickly
Resource-Task Network find good feasible solutions prior to the optimum, which may be very important in real industrial envi-
Spatial grids ronments. We also provide a comprehensive comparison to seven other conceptually different
approaches that have solved the same strip packing problems.
2011 Elsevier B.V. All rights reserved.
1. Introduction strip. The three major classes of 2-D orthogonal packing problems
are strip packing, knapsack and bin packing (Amossen and Pisinger,
N-dimensional allocation problems occur in a variety of indus- 2010) and are illustrated in Fig. 1 for a simple example (white areas
tries (Westerlund, 2005). The terms scheduling, cutting, packing are waste). Strip packing problems are open dimensional (Wascher
and layout are frequently used to refer to specific classes of the et al., 2007) in the sense that all items need to be packed into a
problem, each featuring a few variants. In scheduling, one is con- strip of a given width (W) so as to minimize its height (H). In knap-
cerned with the procedure of allocating resources to tasks over sack problems, the two dimensions of the strip are fixed and the
time so as to meet a given objective such as makespan minimiza- objective is to choose the most profitable subset of items whereas
tion or profit maximization. It can be viewed as a 1-D problem, in in bin packing problems, the objective is to minimize the number
which the relevant dimension is time. One common aspect of cut- of bins required for all the items.
ting and packing problems is that small items have to be packed If one keeps differentiating based on the objective function
into, or cut from, one or more objects often called strip or bins (Ort- there are obvious similarities to simple classes of scheduling prob-
mann et al., 2010). 2-D problems occur in the paper industry where lems. The 1-D version of the strip packing problem can be associ-
the objective is cutting jumbo reels of paper into smaller reels so as ated to a single machine scheduling problem with sequence-
to minimize trim losses. Another class involves the optimal layout dependent changeovers (shown in grey) under makespan minimi-
of industrial facilities based on the minimization of the connectiv- zation. If the time horizon (H) is fixed and we can no longer exe-
ity cost between equipment units (Papageorgiou and Rotstein, cute all tasks, the best ones need to be selected by maximizing
1998; Barbosa-Póvoa et al., 2001). Moving from small rectangles profit, similarly to the knapsack problem. Finally, if we need to exe-
inside a large rectangle to boxes inside containers of a given size cute all tasks in such time horizon there is the possibility of adding
leads to 3-D problems (Tsai et al., 1993; Barbosa-Póvoa et al., machines in parallel, as few as possible to minimize cost, like in bin
2002), while in 4-D problems a number of items are to be allocated packing. This is often called a simultaneous design and scheduling
in three spatial dimensions and one time dimension (Westerlund problem (Castro et al., 2005). The three possibilities are illustrated
et al., 2007). in Fig. 2.
Frequently, certain constraints are imposed on the type of pack- There have been remarkable advances in scheduling ap-
ing that may occur. The most common is orthogonal packing where proaches by the Process Systems Engineering community in the
the edges of the items must be parallel to the edges of the bin or last 20 years. First, with the appearance of unified frameworks
for process representation that allow the systematic modeling
⇑ Corresponding author. Tel.: +351 210924643.
of a process flowsheet (e.g. continuous processes) and/or produc-
tion recipe (e.g. batch processes) in terms of the virtual entities
E-mail address: pedro.castro@lneg.pt (P.M. Castro).
0377-2217/$ - see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2011.06.001
46 P.M. Castro, J.F. Oliveira / European Journal of Operational Research 215 (2011) 45–56
that are the states, resources and tasks. In particular, the Re- straints and the disadvantage of possibly generating very large
source-Task Network (Pantelides, 1994) has brought us the con- problems. Furthermore, since the starting time of all slots is known
cept of excess resource, which is particularly useful to ensure a priori, one needs an iterative procedure for efficiently tackling
perfect packing (zero-waste patterns), as will be seen later. The makespan minimization (Maravelias and Grossmann 2003a; Castro
RTN was linked to a very tight mixed integer linear programming and Grossmann, 2005).
(MILP) discrete-time model that has been shown to be very ver- There are however alternatives to a discrete-time representa-
satile and efficient in a wide variety of industrial problems from tion, as can be seen in Fig. 3. Indeed, there have been major devel-
Dow Chemical plants (Wassick, 2009). In particular, it has been opments concerning continuous-time models and essentially three
used as the basis for solving a non-classical 2-D packing problem different concepts can be used. One alternative is to rely on the
(Wassick and Ferrio, 2011), consisting on loading a semi-trailer concept of precedence, immediate or general, with the model fea-
with packages of different sizes and weights so that the trailer turing binary sequencing variables that give the relative placement
payload weight and also the rear axle weight are kept below of tasks. This is a widely used strategy for higher dimensional prob-
highway allowable limits. lems easily identifiable by the presence of (x, y, z) positioning vari-
In discrete-time models, the time horizon is divided into a finite ables (Sawaya and Grossmann, 2005; Westerlund et al., 2007; Wu
number of uniform slots of a given duration. It is perhaps the sim- et al., 2010). The other choice is to keep track of events taking place
plest alternative with the advantage of not requiring big-M con- by referring to one or multiple time grids.
2 3 4 5 6 7 8 9 10 11 12 2 3 4
1 13 1 5
1
2
2 3 4
1
4. New discrete-space approach (DS) To consider the problem with rotations, one can duplicate the
number of original rectangles (I+) by adding to set I their 90 coun-
The assumption of integer data makes it possible to accurately terparts (the rotated item of rectangle i, is the sole element, i’, of set
model the problem with a discrete-space formulation. Let I
i , where hi = wi and wi’ = hi). Then, one needs to ensure that at
0
Xb = {1, . . . , Wb} be the set of vertical slots in the x-axis and most one orientation is selected for every rectangle, Eq. (7)
Yb = {1, . . . , Hb} the set of horizontal slots in the y-axis for bin b.
XX X XXX X
Ni;x;y;b þ Ni0 ;x;y;b 6 18i 2 Iþ : ð7Þ
We define binary variable Ni,x,y,b to identify whether the bottom- b2B x2X b y2Y b i 0
2I b2B x2X b y2Y b
i
left corner of rectangle i is assigned to the matrix element with i2Ix;y;b i0 2Ix;y;b
the rectangle. Variables R0b indicate if the bin has been selected.
In order to minimize the strip height H, where the bin index has
The no-overlap constraint is given by Eq. (1). The domain of the
been dropped for simplicity, there are basically two possibilities.
binary variables can easily be reduced by realizing that the top-
The first is to consider the variable explicitly and to use another
right corner of every rectangle must also be inside the bin. Set Ix,y,b
set of constraints to ensure that H is greater than the y-coordinate
(2), is used for this purpose. The domain of the remaining variables
of the top facet of all rectangles. The second, much more efficient,
is given in Eqs. (3)–(5)
is to use an indirect procedure (Maravelias and Grossmann, 2003a).
X X X
Rx;y;b ¼ R0b Ni;x0 ;y0 ;b 8b 2 B; x 2 X b ; y 2 Y b ; ð1Þ Starting with an obvious lower bound for the height of the packing
i2I x0 2X b y0 2Y b (Martello et al. 2003; Kenmochi et al., 2009), we can iteratively
xwi þ16x0 6x yhi þ16y0 6y
i2Ix0 ;y0 ;b
solve the problem until a feasible solution is found, which will be
a global optimal solution with respect to the strip height. While
it is sufficient to solve for feasibility, it is better to solve an optimi-
Ix;y;b ¼ i 2 I : x þ wi 6 W b þ 1 ^ y þ hi 6 Hb þ 1 8b 2 B;
zation problem featuring an objective function with a low degree
x 2 Xb; y 2 Y b; ð2Þ of degeneracy (Maravelias and Grossmann 2003a; Castro and
Grossmann, 2005). We follow the same principle and choose to
place rectangles as close as possible to the origin, see Eq. (9)
X X X
min z ¼ Ni;x;y;1 ðx þ yÞ: ð9Þ
x2X 1 y2Y 1 i2Ix;y;1
8b 2 B; x 2 X b ; e 2 Eb ; ð15Þ
X X 8b 2 B; x 2 X b ; x – jX b j; e 2 Eb ð16Þ
N i;x;y;1 6 1 8i 2 I: ð11Þ
x2X 1 y2Y 1 0 1
i2Ix;y;1
B C
B X X C
B C
Y xþ1;e;b Y x;e;b 6 M B1 Ni;x0 ;e;b C
4.1.3. Bin packing problem B C
@ i2I x0 2X b A
In the bin packing problem, the objective is to minimize the xwi þ26x0 6x
i2Ix0 ;b
number of bins required to pack all the items, see Eqs. (12) and
(13). In the case of identical bins, and to reduce solution degener- 8b 2 B; x 2 X b ; x – jX b j; e 2 Eb ; ð17Þ
acy, one can force bin assignments from first to last, Eq. (14)
X Y x;e1 ;b ¼ 0 8b 2 B; x 2 Xb; ð18Þ
min R0b ; ð12Þ
b2B X X
Ni;x0 ;e;b 6 R0b 8b 2 B; x 2 X b ; e 2 Eb : ð19Þ
XX X i2I x0 2X b
Ni;x;y;b ¼ 1 8i 2 I; ð13Þ xwi þ16x0 6x
b2B x2X b y2Y b i2Ix0 ;b
i2Ix;y;b
Fig. 8. Temporary plateau in the objective function may not correspond to a global
optimum.
found (shown on the left). However, this is not the global optimal
solution since one can move rectangle I7 into the largest empty
area, and lower rectangles I1, I2 and I4 to achieve H = 18. A degen-
erate solution is shown on the right, where elementary strips with
x = 11. . .16 have slices of four rectangles placed on it, i.e. |E| = 4.
Following the definition of the cutoff value to 17,001, the problem
becomes infeasible for |E| = 5 and the search is stopped. It is impor-
tant to highlight that consecutive events can have the same y-coor-
dinate (e.g. Y 1...2;e1 ¼ Y 1...2;e2 ¼ Y 1...2;e3 ¼ 0; Y 3...9;e3 ¼ Y 3...9;e4 ¼ 15) and
that a given event point can be located anywhere with respect to
their counterparts on different grids if they are not occupied by
the same rectangle.
XX
Ni;x;e;1 6 1 8i 2 I: ð29Þ
x2X 1 e2E1
i2Ix;1
Fig. 10. Influence of the number of event points on the optimal solution for Ex2. On the left |E| = 3 for H = 20. On the right |E| = 4 for H = 18, which is a global optimal solution.
52 P.M. Castro, J.F. Oliveira / European Journal of Operational Research 215 (2011) 45–56
Table 1
Overview of computational statistics (best performer in bold, infeasible solutions in italic).
Fig. 11. Best solution found with hybrid discrete/continuous-space algorithm for cgcut02 (H = 67).
running Windows Vista Enterprise. In Eqs. (16) and (17) we have 08); 3 in ngcut09 and ht09; 2 for cgcut02. The best found solution
used M = 300. for the latter, showing the location of the event points, can be seen
in Fig. 11.
6.1. Discrete vs. hybrid discrete/continuous The most significant advantage of DCS is the ability to always
find very good if not the optimal solutions. Furthermore, it finds
We start the analysis by comparing the performance of the two them fast, as can be seen in Table 2, which shows the progress in
approaches presented. Besides the number of rectangles, strip the value of the objective function (intermediate checkpoints at
width and known optimal solution from the literature (Hopt), Ta- 1, 10 and 60 CPUs), as well as the total computational time and
ble 2 gives the starting and ending iterations of the discrete and optimality gap up to 120 CPUs. In fact, the optimal solution is
hybrid discrete/continuous search algorithms, the best solution found in less than 2 minutes for all but one of the 21 successful
found and total computational effort. runs. Interestingly, the failure is in ht06, which can be solved to
The new hybrid discrete/continuous-space (DCS) approach global optimality in just 11.5 CPUs (see Table 1) and can be ex-
could find the optimal solution in 21 cases, proving optimality plained as follows.
62% of the time. In contrast, the discrete-space model (DS) was able The search algorithm of DCS is able to find the global solution
to solve only 19 out of the same 29 problems, and the unsolved 10 for 5 events, and then removes it from the feasible space by means
were infeasible (recall that DS works mostly in the infeasible space, of a cutoff value before resolving the problem for 6 events with the
with the first feasible solution being an optimum). The optimality purpose of proving optimality. In contrast, the results in Table 2
gap for DCS was equal to: 1 in five test cases (ht02, 04, 05, 07 and consider the complete feasible region corresponding to the number
54 P.M. Castro, J.F. Oliveira / European Journal of Operational Research 215 (2011) 45–56
Table 3 number of successful runs rises to 24. This behavior is shared with
Performance of discrete-space algorithm using a tighter lower bound (L4) instead of the discrete/continuous-space algorithm, which exhibits an even
H0 in the initialization.
higher increase in complexity between consecutive iterations,
Problem L4 H CPU s and is typical in multidimensional location problems, character-
ngcut01 23 23 0.96 ized by a huge number of feasible item allocations (high degree
ngcut02 30 30 2.38 of degeneracy).
ngcut04 20 20 0.48 The two main motivations for the replacement of discrete by
ngcut06 31 31 5.73
ngcut08 33 33 10.3
continuous models are: (i) increased data accuracy from consider-
ngcut10 80 80 122 ing real rather than integer values; (ii) reduced problem size that
ngcut11 52 52 661 may potentially lead to easier to solve mathematical problems.
ngcut12 87 87 1582 With respect to the former aspect, the hybrid model enables to
consider real data for the rectangles height and is thus a more
accurate approach in the y-dimension. To analyze the effect of
using event points instead of a discrete mesh of size one, Table 4
Table 4
gives the model statistics.
Statistics for last problem solved (DV = discrete variables, TV = total variables,
SE = single equations).
Cleary, we were very successful in reducing the number of dis-
crete variables, which can go down to one order of magnitude with
Problem Discrete-space Discrete/continuous-space
cgcut02, the most difficult problem of the lot. In terms of total
DV TV SE DV TV SE number of equations, the results are not as significant and the
Ex1 1342 1667 334 580 671 359 new approach can even lead to a higher value (e.g. ngcut02) due
Ex2 1323 1648 334 585 676 359 to the spatial constraints. Given the use of four-index binary vari-
Ex3 2485 2946 471 652 733 322 ables (Ni,x,y,b), the discrete-space model is not suitable for simulta-
Ex4 2428 2849 432 720 801 323
neously large values in terms of number of rectangles, bins, strip
Ex5 3729 3970 262 1336 1417 325
SCP16 1829 2028 213 441 484 168 height and width. With the increase in |I| W H, the discrete-
cgcut01 2409 2640 247 868 939 282 space model for strip packing (b = 1) rapidly becomes intractable,
cgcut02 63,438 67,849 4434 6425 6776 1413 and one may run out of memory just to generate the problem. This
ngcut01 1069 1300 241 420 491 276
explains why other test instances from the literature were not
ngcut02 2582 2873 308 1080 1181 397
ngcut03 3820 4101 302 1296 1377 325 included.
ngcut04 606 807 208 186 217 121 One final remark to mention that for DCS, the objective function
ngcut05 2954 3315 375 560 621 204 of the linear program resulting from the relaxation of the integral-
ngcut06 2391 2692 316 792 873 319 ity constraints is equal to the value of the continuous lower bound,
ngcut07 1603 2004 409 393 454 242
given by Eq. (31)
ngcut08 4780 5441 674 856 937 325
ngcut09 10,428 11,409 999 1578 1699 486
X
wi hi =W 1 : ð31Þ
ngcut10 12,285 14,536 2264 1512 1693 721
i2I
ngcut11 11,076 12,577 1516 1920 2101 723
ngcut12 23,813 26,334 2543 4277 4668 1556
ht01 3907 4308 417 1305 1406 406
ht02 4222 4623 418 1305 1406 407
6.2. Comparison to continuous-space models
ht03 4054 4455 417 1488 1609 484
ht04 10,691 11,292 626 4140 4341 815 The purpose of this section and the next is to discuss the perfor-
ht05 10,938 11,539 626 4230 4431 815 mance of other approaches that can tackle the 2-D strip packing
ht06 11,131 11,732 626 5010 5251 973
problem, on most of the test problems considered in this article.
ht07 35,531 37,332 1829 8394 8755 1456
ht08 37,249 39,050 1830 7400 7701 1219 We start with Ex5 that was tackled by Sawaya and Grossmann
ht09 35,415 37,216 1829 7055 7356 1218 (2005) and Westerlund et al. (2007) using identical computers and
versions of the same MILP solver. In both papers, a continuous-
space approach is used, where (xi, yi) variables define the position
of rectangle i in the strip, either the upper-left corner or the cen-
of events listed in the ninth column of Table 1, which for ht06 is troid coordinate, respectively. Note that contrary to DCS, neither
equal to 6. Together with the observed higher computational times of these approaches relies on the concept of events.
for Ex1, ngcut03, 04, 07 and ht01, it serves to show the importance The big-M reformulation of the generalized disjunctive problem
of the search algorithm in Fig. 9. formulation of Sawaya and Grossmann (2005) was able to find the
The discrete-space approach (DS) is, on the other hand, better at optimal solution unlike its convex hull counterpart and despite the
proving optimality. This is even more relevant if one recalls that slightly larger integrality gap. It then became 44 times faster with
the procedure is rigorous while the method for DCS may terminate the addition of 62 cuts, see Table 5. Westerlund et al. (2007) were
at a suboptimal solution (Fig. 8). Clearly, the success rate is greater able to do even better with a closely related big-M approach, pri-
when the lower bound (H0) is close to the optimal solution making marily due to the major reduction in integrality gap due to the
DS particularly suitable for perfect-packing problems (ht01-09). In addition of symmetry breaking constraints. In fact, the optimal va-
these problems, Eq. (6) was added to the model, while Eq. (23) for lue from the relaxed MILP (RMIP), 22.5, corresponds to the contin-
DCS was not included since it did not lead to better results. More uous lower bound (Eq. (31)), like in DCS.
specifically, since it removes all suboptimal solutions from the fea-
sible space, DCS may fail to provide a solution in a reasonable 6.3. Comparison to exact algorithms
amount of time, thus compromising one of its major strengths.
The performance of DS can be improved further if the overall Table 6 lists the results for six other methods that are heavily
lower bound (L4) from Martello et al. (2003) is used instead of based on the exact approach by Martello et al. (2003). Besides
the weaker continuous lower bound. Major increases in perfor- the original algorithm, we reproduce the results from the sources
mance can be observed whenever L4 > H0, as can be seen in Table 3. for the branch & bound (B&B) and dichotomous algorithm (DA)
In particular, the number of iterations decreases to one and the of Bekrar et al. (2007), algorithms STAIR and G-STAIR of Kenmochi
P.M. Castro, J.F. Oliveira / European Journal of Operational Research 215 (2011) 45–56 55
Table 5
Results of continuous-space models for Ex5.
Table 6
Total computational time from other sources in CPUs (success in finding the optimal solution in bold).
Name Martello et al. (2003) Bekrar et al. (2007) Kenmochi et al. (2009) Alvarez-Valdes et al. (2009) This work
Hardware Pentium 3, 800 Megahertz Pentium M, 1.7 Gigahertz Pentium 4, 3.0 Gigahertz Pentium 4, 2.0 Gigahertz T9300, 2.5 Gigahertz
Approach B&B DA STAIR G-STAIR DS DCS
SCP16 – 606 217 – – – 5.30 2.87
cgcut01 11.5 521 324 0.10 0.12 0.0 2.37 4.75
cgcut02 3600a,c 3600a,c 1219c 3600a,b 3600a,b 1200a 7463a,b 7201a,c
ngcut01 0.05 1.01 23.2 2080 0.39 2.2 81.6 461
ngcut02 11.3 3600a 1053 3600a,b 3600a,b 3.1 7202a,b 12,368a
ngcut03 27.0 1621 520 0.09 0.10 0.0 12.6 2.87
ngcut04 0.00 0.01 0.02 3.63 0.14 0.0 1.22 5.90
ngcut05 0.00 0.01 120 0.11 0.07 0.0 3.18 0.62
ngcut06 727 3579 1079 3600a,b 147.3 4.6 3742b 8194a
ngcut07 0.00 0.01 0.00 3600a,b 0.10 0.0 0.98 3.37
ngcut08 53.1 55.0 178 30.5 0.50 3.5 29.5 7363a
ngcut09 3600a 3600a 1270 3600a,b 1972 58.1 7210a,b 7200a,c
ngcut10 0.18 198 1153 3600a,b 114 2.6 8760a,b 7200a
ngcut11 483 1059 733 3600a,b 7.71 13.8 7215a,b 10,028a
ngcut12 0.00 3600a 866 3600a,b 3600a,b 0.0 10339a,b 7202a
ht01 10.8 0.00 0.00 0.10 0.07 0.0 2.08 2.01
ht02 624 7593 379 0.12 0.07 0.4 5.28 7200a,c
ht03 501 25.2 198 0.08 0.10 0.1 2.51 9.34
ht04 8.26 3600a,c 874 0.08 0.11 0.1 91.8 7202a,c
ht05 20.3 3600a 572 0.09 0.06 0.1 20.4 7200a,c
ht06 16.9 0.00 0.00 0.11 0.06 1.4 18.3 11.5
ht07 3600a,c 3600a,c 2045c 0.12 0.10 1.8 3771 7201a,c
ht08 3600a,c 3600a,c 1905c 71.4 77.0 1200a,c 7200a,b 7201a,c
ht09 0.00 0.00 0.00 0.10 0.13 8.7 7200a,b 7201a,c
a
Maximum computational time termination.
b
No feasible solution.
c
Suboptimal solution.
et al. (2009) and the also B&B algorithm of Alvarez-Valdes et al. Clearly, our new proposed algorithms are not as efficient, which
(2009). is not surprising considering that we do not go into details of
From the results in Table 6, one can see that the original algo- improving the search strategy through special purpose branch
rithm is still very much competitive, particularly in the ngcut01- and bound operations. Still, the easy to implement discrete-space
12 instances, and that the B&B algorithm by Alvarez-Valdes et al. approach, when combined with a standard MILP solver like CPLEX,
(2009) is the best overall performer. This approach only failed to can also be quite efficient in cases that do not require a large mesh
return the optimal solution in one occasion (ht08) and failed to and featuring a tight lower bound. Furthermore, the hybrid dis-
prove optimality in another (cgcut02). For the latter problem, the crete/continuous-space model is more accurate since it can handle
best found solution features H = 65, while the other methods re- instances having non-integer values for heights. With respect to
port H = 67. this aspect, Kenmochi et al. (2009) refer that an algorithm based
With respect to the work by Bekrar et al. (2007), DA outper- on a sequence pair representation was tried to allow for continu-
forms B&B to achieve a closer but still overall worse perfor- ous input data but failed to be as competitive as STAIR and G-
mance than Martello et al. (2003), considering the significant STAIR. Since our ultimate goal will be to use the events concept
difference in hardware. It is also worthwhile to mention that a to come up with a continuous-space algorithm for N-dimensional
normal termination status is indicated for cgcut02 and ht07-08 problems, such remarks and the results obtained show that we
(DA) but the reported solutions are suboptimal (H = 67, 31 and are on the right track.
31, respectively).
The algorithms by Kenmochi et al. (2009) were aimed at perfect
packing problems so it is not surprising to find out that the two 7. Conclusions
algorithms excel in the perfect packing instances (ht01-09), being
able to solve all nine instances in less than 2 minutes. The algo- This paper has presented two new mixed-integer linear pro-
rithms are also efficient in instances where the optimal values gramming approaches for two-dimensional orthogonal packing
are close to the trivial lower bounds, particularly G-STAIR. problems. Both are heavily based on existing models for scheduling
56 P.M. Castro, J.F. Oliveira / European Journal of Operational Research 215 (2011) 45–56
problems (one-dimensional) and can be distinguished in terms of Castro, P.M., Grossmann, I.E., 2005. New continuous-time MILP model for the short-
term scheduling of multistage batch plants. Industrial & Engineering Chemistry
spatial representation.
Research 44, 9175–9190.
The discrete-space approach borrows the concept of excess re- Castro, P.M., Harjunkoski, I., Grossmann, I.E., 2011. Greedy algorithm for scheduling
source from the Resource-Task Network process representation for batch plants with sequence-dependent changeovers. AIChE Journal 57, 373–
the derivation of the no-overlap constraint. It essential consists in 387.
Castro, P.M., Novais, A.Q., 2008. Short-term scheduling of multistage batch plants
two sets of constraints and one objective function easily adaptable with unlimited intermediate storage. Industrial & Engineering Chemistry
to strip, knapsack and bin packing problems. The approach has Research 47, 6126–6139.
been shown to be particularly efficient for perfect packing in- Dolatabadi, M., Lodi, A., Monaci, M., 2010. Exact algorithms for the two-dimensional
guillotine knapsack. Computers and Operations Research. doi:10.1016/
stances, with the main drawback being that the first feasible solu- j.cor.2010.12.018.
tion, which is also optimal, is only achieved at termination. This He, K., Huang, W., 2011. An efficient placement heuristic for the three-dimensional
may render it useless in problems involving a large number of rect- rectangular packing. Computers and Operations Research 38, 227–233.
Hifi, M., 1998. Exact algorithms for the guillotine strip cutting/packing problem.
angles, wide and potentially long strips, for which prohibitive Computers and Operations Research 25, 925–940.
problem sizes may result. Ierapetritou, M.G., Floudas, C.A., 1998. Effective continuous-time formulation for
The main novelty has however been the development of a hy- short-term scheduling. 1. Multipurpose batch processes. Industrial and
Engineering Chemistry Research 37, 4341.
brid discrete/continuous-space approach that involved: (i) discret- Jain, V., Grossmann, I.E., 2001. Algorithms for hybrid MILP/CP models for a class of
izing a particular bin (strip) into a set of parallel vertical strips of optimization problems. INFORMS Journal on Computing 13, 258.
width one; (ii) postulating a number of events to be optimally lo- Kenmochi, M., Imamichi, T., Nonobe, K., Yagiura, M., Nagamochi, H., 2009. Exact
algorithms for the two-dimensional strip packing problem with and without
cated anywhere along the strip height. A key aspect is that the
rotations. European Journal of Operational Research 198, 73–83.
location of event points can vary from one elementary vertical strip Li, J., Floudas, C.A., 2010. Optimal event point determination for short-term
to the next, unless a particular rectangle is shared by them. To scheduling of multipurpose batch plants via unit-specific event-based
overcome the solution dependency on the number of events, which continuous-time approaches. Industrial and Engineering Chemistry Research
49, 7446–7469.
is directly related to the maximum number of rectangles that can Maravelias, C.T., Grossmann, I.E., 2003a. Minimization of the makespan with a
be packed vertically, the model has been embedded in a search discrete-time state-task network formulation. Industrial and Engineering
algorithm that iteratively solves the problem until no improve- Chemistry Research 42, 6252.
Maravelias, C.T., Grossmann, I.E., 2003b. A new general continuous-time state task
ment is observed in the objective function. The main strength of network formulation for short term, scheduling of multipurpose batch plants.
the approach is the ability to provide very good results with few Industrial and Engineering Chemistry Research 42, 3056–3074.
computational resources from the solution of an easy to implement Martello, S., Monaci, M., Vigo, D., 2003. An exact approach to the strip-packing
problem. Informs Journal on Computing 15, 310–319.
mixed integer linear programming model. These can be viewed as Méndez, C.A., Cerdá, J., Grossmann, I.E., Harjunkoski, I., Fahl, M., 2006. State-of-the-
tight upper bounds, so the approach also has the potential to be a art review of optimization methods for short-term scheduling of batch
module of exact algorithms. processes. Computers and Chemical Engineering 30, 913–946.
Miyazawa, F.K., Wakabayashi, Y., 2009. Three-dimensional packings with rotations.
Future work will consist on developing a continuous-space Computers and Operations Research 36, 2801–2815.
model using the concept of events in both dimensions, where the Ortmann, F.G., Ntene, N., van Vuuren, J.H., 2010. New and improved level heuristics
main challenge will be to find an efficient set of spatial constraints for the rectangular strip packing and variable-sized bin packing problems.
European Journal of Operational Research 203, 306–315.
that rigorously accounts for the interaction between the different
Pantelides, C.C., 1994. Unified frameworks for the optimal process planning and
grids. scheduling. In: Proceedings of the Second Conference on Foundations of
Computer Aided Operations; Cache Publications, New York, pp. 253.
Papageorgiou, L., Rotstein, G.E., 1998. Continuous domain mathematical models for
References optimal process plant layout. Industrial and Engineering Chemistry Research
37, 3631–3639.
Sawaya, N.W., Grossmann, I.E., 2005. A cutting plane method for solving linear
Alvarez-Valdes, R., Parreño, F., Tamarit, J.M., 2008. Reactive GRASP for the strip- generalized disjunctive programming problems. Computers and Chemical
packing problem. Computers and Operations Research 35, 1065–1083. Engineering 29, 1891–1913.
Alvarez-Valdes, R., Parreño, F., Tamarit, J.M., 2009. A branch and bound algorithm Tsai, R.D., Malmstrom, E.M., Kuo, W., 1993. Three dimensional palletization of
for the strip packing problem. OR Spectrum 31, 431–459. mixed sized boxes. IIE Transactions 25 (4), 64–75.
Amossen, R.R., Pisinger, D., 2010. Multi-dimensional bin packing problems with Wascher, G., Haussner, H., Schumann, H., 2007. An improved typology of cutting
guillotine constraints. Computers and Operations Research 37, 1999–2006. and packing problems. European Journal of Operational Research 183 (3), 1109–
Barbosa-Póvoa, A.P., Mateus, R., Novais, A.Q., 2001. Optimal two-dimensional layout 1130.
of industrial facilities. International Journal of Production Research 39 (12), Wassick, J., 2009. Enterprise-wide optimization in an integrated chemical complex.
2567–2593. Computers and Chemical Engineering 33, 1950.
Barbosa-Póvoa, A.P., Mateus, R., Novais, A.Q., 2002. Optimal three-dimensional Wassick, J.M., Ferrio, J., 2011. Extending the resource task network for industrial
layout of industrial facilities. International Journal of Production Research 40 applications. Computers and Chemical Engineering, doi:10.1016/
(7), 1669–1698. j.compchemeng.2011.01.010..
Bekrar, A., Kacem, I., Chu, C., 2007. A comparative study of exact algorithms for the Wei, L., Zhang, D., Chen, Q., 2009. A least wasted first heuristic algorithm for the
two dimensional strip packing problem. Journal of Industrial and Systems rectangular packing problem. Computers and Operations Research 36, 1608–
Engineering 1 (2), 151–170. 1614.
Castro, P.M., Barbosa-Póvoa, A.P., Matos, H.A., Novais, A.Q., 2004. Simple Westerlund, J., 2005. Aspects on N-dimensional allocation. PhD Thesis. Faculty of
continuous-time formulation for short-term scheduling of batch and Chemical Engineering, Åbo Akademi University, Finland.
continuous processes. Industrial & Engineering Chemistry Research 43, 105– Westerlund, J., Papageorgiou, L.G., Westerlund, T., 2007. A MILP model for N-
118. dimensional allocation. Computers and Chemical Engineering 31, 1702–1714.
Castro, P.M., Barbosa-Póvoa, A.P., Novais, A.Q., 2005. Simultaneous design and Wu, Y., Li, Wenkai, Goh, M., Souza, R., 2010. Three-dimensional bin packing problem
scheduling of multipurpose plants using resource-task network based with variable bin height. European Journal of Operational Research 202, 347–
continuous-time formulations. Industrial & Engineering Chemistry Research 355.
44, 343–357.