Automated design of relocation rules for minimising energy

consumption in the container relocation problem

Marko Ðurasević Mateja Ðumić
University of Zagreb, Faculty of Electrical Engineering and Department of Mathematics, Josip Juraj Strossmayer
Computing University of Osijek
Zagreb, Croatia Osijek, Croatia
[email protected] [email protected]

Rebeka Čorić Francisco J. Gil-Gala

Department of Mathematics, Josip Juraj Strossmayer University of Oviedo. Department of Computing
arXiv:2307.01513v1 [cs.NE] 4 Jul 2023

University of Osijek Gijón, Spain

Osijek, Croatia [email protected]
[email protected]
ABSTRACT atop another and/or side by side. This way blocks are formed that
The container relocation problem is a combinatorial optimisation have stacks (width), a number of tiers (height), and a number of
problem aimed at finding a sequence of container relocations to bays (length). The objective is to retrieve and load all containers
retrieve all containers in a predetermined order by minimising a from the yard in a predetermined order. However, a container can
given objective. Relocation rules (RRs), which consist of a priority be retrieved only if it is located on the top of its stack. If there are
function and relocation scheme, are heuristics commonly used for containers on top of the one that needs to be retrieved, they first
solving the mentioned problem due to their flexibility and efficiency. need to be relocated to other stacks.
Recently, in many real-world problems it is becoming increasingly During the years, many heuristics and metaheuristics were pro-
important to consider energy consumption. However, for this vari- posed to solve this problem [4, 8, 12, 13]. These methods are com-
ant no RRs exist and would need to be designed manually. One putationally expensive and require a substantial amount of time to
possibility to circumvent this issue is by applying hyperheuristics obtain solutions for larger problem sizes. Therefore, simple heuristic
to automatically design new RRs. In this study we use genetic pro- methods, called relocation rules (RRs), are proposed in the literature
gramming to obtain priority functions used in RRs whose goal is to to solve CRP [2, 12]. RRs construct the solution incrementally by
minimise energy consumption. We compare the proposed approach determining which relocation should be performed based on the
with a genetic algorithm from the literature used to design the current system information. For that purpose, RRs use a priority
priority function. The results obtained demonstrate that the RRs function (PF) to rank all possible relocations and select the best
designed by genetic programming achieve the best performance. one. Since manually designing such PFs is difficult, certain studies
investigated the possibility of automatically designing them [6, 15].
KEYWORDS Due to the growing environmental concerns that arise today,
optimising energy related criteria is becoming increasingly impor-
Genetic Programming, Genetic Algorithm, Container Relocation
tant in various optimisation problems, such as vehicle routing [5]
Problem, Hyper-heuristics
or various scheduling problems [14]. However, in CRP the energy
ACM Reference Format: consumption criterion did not receive much attention, with only
Marko Ðurasević, Mateja Ðumić, Rebeka Čorić, and Francisco J. Gil-Gala. a few studies focusing on optimising it either directly [6], or as a
2023. Automated design of relocation rules for minimising energy consump-
part of the total cost objective [9]. Thus, there is a lack of RRs that
tion in the container relocation problem. In Proceedings of ACM Conference
could be used to efficiently optimise this criterion.
(Conference’17). ACM, New York, NY, USA, 4 pages. https://doi.org/10.1145/
nnnnnnn.nnnnnnn To close this gap, we examine the application of hyperheuris-
tics to generate RRs appropriate for optimising the total energy
1 INTRODUCTION consumption during the retrieval process of containers. Although
this problem was tackled in [6], the authors manually defined the
The container relocation problem (CRP), is a combinatorial optimisa- mathematical expression of the priority function used to rank all
tion problem with applications in warehouse and yard management the relocations, and used a GA to optimise certain parameters in
[7]. Due to the limited space, containers are usually stacked one that expression. As such, the approach is limited in a sense that the
Permission to make digital or hard copies of part or all of this work for personal or structure of the priority function still needs to be defined manually.
classroom use is granted without fee provided that copies are not made or distributed Therefore, we revisit this problem and apply GP as a hyperheuristic
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
to generate PFs of an arbitrary structure. We use the same informa-
For all other uses, contact the owner/author(s). tion as the authors in [6] to make a fair comparison between the
Conference’17, July 2017, Washington, DC, USA methods, and show that PFs designed by GP construct significantly
© 2023 Copyright held by the owner/author(s). better solutions than the previously proposed GA. The contribu-
ACM ISBN 978-x-xxxx-xxxx-x/YY/MM.
https://doi.org/10.1145/nnnnnnn.nnnnnnn tions of this paper can be outlined as follows:
Conference’17, July 2017, Washington, DC, USA Marko Ðurasević, Mateja Ðumić, Rebeka Čorić, and Francisco J. Gil-Gala

(1) develop a GP based hyperheuristic method to optimise the above it must be moved to another stack to allow retrieval. RS uses
total energy consumption for CRP; PF to decide which stacks the containers above the target container
(2) compare priority functions for relocation rules evolved by should be moved to in order to relieve the target containers. This
GA and GP. is done iteratively, one container at a time. RS determines the con-
tainer that needs to be moved next, and PF assigns a numeric value
2 CRP PROBLEM DESCRIPTION to each stack to which the given container can be moved. Depend-
We consider the single bay CRP in which the bay consists of 𝑆 stacks ing on PF, the container is moved to the stack for which the best
with 𝐻 tiers. Every stack has a height denoted with ℎ(𝑆) that has to value was determined.
be less or equal to the maximum height 𝐻 . There are 𝐶 containers RSs are simple algorithms that are defined manually. Based on
in the bay and a single crane that can move one container at a time. the moves that are allowed we distinguish between the restricted
Each container has a different priority, which denotes the order of and the unrestricted RS. In the restricted version, only containers
their retrieval from the yard. To solve the problem, two types of located above the target container may be moved, while in the
operations can be performed by the gantry crane, relocation and unrestricted version, there is no such restriction, i.e., all containers
retrieval. Relocation moves a container from the top of one stack to located on top of their stack may be moved.
another, which can be done only if the stack to which the relocation
is being made has a height smaller than 𝐻 . The second operation,
3.2 Using GA and GP for developing PFs
retrieval, picks a container from the top of the stack and moves it to
the truck used for loading, which is located at position 0 denoting Designing a good PF manually is a challenging task, because of
the beginning of the bay. which several attempts to automate this process were performed.
Each container in the bay has an ID that determines in which Partial automation was done in [6], in which the authors manually
order they need to be retrieved. The container with the smallest ID defined a general expression with a certain number of free parame-
in the yard is the one that needs to be retrieved next, and is called ters that were optimised with GA. In a more recent work the entire
the target container. If the target container is not located at the top PF was developed using GP [15], which achieved significantly better
of its stack, it is required to relocate all the containers above it to performance than several existing manually designed PFs.
different stacks. The stack from which the container is moved is In this work, we consider both ways to design PFs to optimise
called the origin stack, whereas the stack to which the container is the total energy consumption for CRP. Both GA and GP use the
moved is called the destination stack. All relocation and retrieval same evolutionary scheme, with the the main difference being the
sequences that guarantee the crane can retrieve every container in representation of individuals. The GA uses a list of floating point
a predetermined order denote feasible CRP solutions. The goal is to numbers denoting the parameters it optimises, while GP uses the
find a sequence that minimises a given objective. In this study we standard expression tree representation. The evaluation is done
optimise the total energy consumed while retrieving all containers using a fitness function that evaluates each individual on a set of
from the yard, which can be defined as [6]: problems and assigns a numerical value (fitness) to the individual.
In each iteration, a 3-tournament selection is used, the two better
𝑀
∑︁ of the selected individuals are used for crossover, and the worst
𝑇 𝐸𝐶 = 𝑊𝑚 (ℎ ∗ ℎ𝑚 + 𝑙 ∗ 𝑙𝑚 + 𝑥 ∗ 𝑥𝑚 ),
is replaced by a newly created individual to which the mutation
𝑚=1
operator is applied with a certain probability. This is repeated until
where: the maximum number of fitness function evaluations is reached.
• ℎ (𝑙) – energy consumed per ton for one tier hoisted (lowered) In [6], the authors propose the global retrieval heuristic (GRH)
• 𝑥 – energy consumed per ton when moving the crane stack for restricted CRP with container weights to optimise total en-
• ℎ𝑚 (𝑙𝑚 ) – tiers hoisted (lowered) during move 𝑚 ergy consumption. In this study, we adapt GRH to also work with
• 𝑥𝑚 – stacks crossed during move 𝑚 the unrestricted RS to test whether this can improve the results.
• 𝑊𝑚 – moving weight of move 𝑚; 𝑊𝑚 = 𝑊𝑠 + 𝑊𝑐 , where 𝑊𝑠 When deciding where to move the container, GRH uses a penalty
denotes the weight of the crane, and 𝑊𝑐 denotes the weight function and selects the stack that received the lowest value. The
of the container moves (equals to 0 if crane was empty) penalty function is given by expression (1), and the description of
• 𝑀 – number of moves required to retrieve all containers. the variables can be found in the Tables 1 and 2. Table 2 outlines the
Based on [6], values for ℎ, 𝑙, 𝑥 are set to 0.9, 0.02 and 0.08 respec- variables that the algorithm uses as inputs, while Table 1 contains
tively, while crane weight 𝑊𝑠 is equal to 0.5 tons. the free parameters that must be set before solving the problem
and whose values are between 0 and 1. The idea presented in the
3 METHODOLOGY paper [6] is to apply a GA to determine the free parameters from
Table 1. The GA uses a simple floating-point encoding, where each
3.1 Relocation rules individual consists of 12 real numbers, each denoting one of the
Relocation rules (RRs) represent simple constructive heuristics that parameters.
iteratively build the solution to CRP. They consist of two parts - GP as a hyperheuristic achieves good results in automatic de-
the relocation scheme (RS) and the priority function (PF) [15]. RS velopment of scheduling rules [1, 10] and has been successfully
takes care of problem constraints and creates a plan for container applied to the basic CRP problem [15], in which the total number of
retrieval and relocation. If the container that needs to be retrieved relocations and crane operation time were optimised. Encouraged
next is on top of its stack, it is retrieved, otherwise, the containers by this, in this paper we apply GP to generate PFs to minimise the
 Conference’17, July 2017, Washington, DC, USA

total energy consumption in CRP. To analyse how GP compares to

the GA approach of [6], GP uses the same system information to   𝐴1   𝐴1
construct the PF. This means that the terminal set of GP comprises ℎ𝑠 𝑙𝑠  𝑥  𝐴1
𝑠
𝑃𝑒𝑛𝑎𝑙𝑡𝑦𝑠 = 𝛼 +𝛽 +𝛾
of the variables given in Table 2 (except 𝐴1, 𝐴3 and 𝐴4 ). The set 𝑚𝑥𝐻𝑒𝑖𝑔ℎ𝑡 𝑚𝑥𝐻𝑒𝑖𝑔ℎ𝑡 𝑆
of functions used in the development of the penalty function con-
𝑊𝑐 10𝑃1
   𝑐 − 𝑡  𝐴3
𝑠
sists of addition, subtraction, multiplication and protected division + 𝛿𝑟𝑠 + 𝜖𝑟𝑠 + 𝜂 (1 − 𝑟𝑠 )𝑔𝑠 (1)
(returns 1 if the divisor is close to 0). 𝑊𝑚𝑎𝑥 𝐶
  𝐴4  
𝑘𝑠 𝑛𝑠
+𝜃 +𝜇
𝑆 𝑚𝑥𝐻𝑒𝑖𝑔ℎ𝑡

Table 1: GRH settings [6] 4 EXPERIMENTAL SETUP

To test the performance of GRH and GP evolved PFs for RRs, the
Parameter Description Caserta [3] and Zhu [13] datasets are used. These original instances
are used as the test set, whereas additional instances were generated
𝛼 / 𝛽 /𝛾 importance of minimising hoist- to be used for training GP and GRH. In order to be able to use
ing/lowering/trolleying the original instances from these two sets with energy criteria,
𝑃1 importance of minimising hoisting, lowering, an additional weight with an uniform distribution from 1 to 30
and trolleying of heavy (versus light) containers was generated for each container, as was done in [6]. The adapted
𝛿 /𝜖 importance of minimising/delaying rehandling problem instances can be obtained from http://www.zemris.fer.hr/
𝑃2 / 𝑃3 importance of minimising/delaying rehandling ~idurasevic/CRP/CRP.7z.
of heavy (versus light) containers Both GP and GA use a population of 1 000 individuals, mutation
𝜂 importance of tightness probability of 0.3 for the restricted and 0.1 for the unrestricted RRs,
𝜃 importance of moving containers closer to the and 50 000 function evaluations. The maximum tree depth was set to
truck lane 5 in GP. GP used the subtree, uniform, context preserving, size fair,
𝑃4 importance of moving heavy (versus light) con- and one point crossover operators, as well as the subtree, hoist, node
tainers closer to truck lane complement, node replacement, permutation, and shrink mutation
𝜇 importance of keeping stack heights low operators [11]. The GA used several well known genetic operators,
like arithmetic, SBX, BLX-𝛼, and others. For mutation, the uniform
mutation operator is used, which generates a random number from
the interval [0, 1]. In cases when several crossover or mutation
operators are defined, a random one is selected and applied for each
time the operator needs to be invoked.
To obtain a notion on the performance of the algorithms, GP
and GA were executed 30 times to evolve RRs using the training
Table 2: Variables contained in the penalty score function set. The best RR obtained in each execution is evaluated on the
whose values are calculated when the container 𝑐 with weight test set and the total consumed energy for each of these 30 rules is
𝑊𝑐 is reshuffled to the destination stack 𝑠 [6] determined. To test whether the obtained results are statistically
significant, the Kruskal-Wallis test with the Dunn post hoc test and
Bonferroni correction method was used. The obtained differences
Variable Description
are considered significant if a p-value below 0.05 was obtained.
ℎ𝑠 / 𝑙𝑠 / 𝑥 𝑠 number of tiers hoisted/lowered/trolleyed when
moving reshuffled container to stack 𝑠 5 RESULTS
𝑟𝑠 binary variable that equals 1 if the container c
must be reshuffled again if it is placed on the Figure 1 outlines the results obtained for RRs generated by GP and
stack s, otherwise it equals 0. GRH. By comparing the two methods used to automatically design
𝑡𝑠 lowest numbered container in stack 𝑠 RRs, we see that GP evolved PFs consistently achieve a better mini-
mum and median values of the results on both datasets. The first
𝑔𝑠 tightness, calculated with formula 𝑔𝑠 = 𝑡𝑠 −𝑐 −1
𝐶 thing to notice is that the restricted versions (marked in the figure
𝑘𝑠 amount of trolley movement away from truck
with -R next to the name of the approach) of the RRs consistently
lane if container c reshuffled to stack 𝑠; is equal
perform better than their unrestricted variants (marked in the fig-
to 0 if trolley moves toward truck lane when
ure with -U next to the name of the approach). The reason why
container reshuffled to stack 𝑠
this happens is that the unrestricted version introduces additional
𝑛𝑠 number of containers in stack 𝑠
moves that are performed, which ultimately increases the total
𝐴1 1 + 𝑊𝑊𝑐
𝑚𝑎𝑥
· 10 · 𝑃1 consumed energy. As we see, this increase is quite substantial and
𝐴3 𝐴3 = 1 + 𝑊𝑊 𝑐
· 10 · 𝑃3 therefore leads to significant deterioration of the results. Therefore,
𝑚𝑎𝑥
𝐴4 𝑊𝑐
1 + 𝑊𝑚𝑎𝑥 · 10 · 𝑃4 we can conclude that for this optimisation criteria the unrestricted
version of RRs is not appropriate.
Conference’17, July 2017, Washington, DC, USA Marko Ðurasević, Mateja Ðumić, Rebeka Čorić, and Francisco J. Gil-Gala

If we compare the rules obtained with GP and the GRH, we see for CRPs with non-standard criteria, especially considering that
that the rules generated by GP perform better in almost all cases. it uses the same system properties as GRH. Thus, we see that the
This is most evident in the restricted variant, where even the worst additional flexibility of GP to freely design the expression of PF gives
solution obtained with GP outperforms the best solution obtained it the ability to obtain better solutions than rules where the structure
with GRH. The only discernible advantage of GRH is that the re- is manually defined and only the corresponding coefficients are
sults are less dispersed than with GP. However, as these results optimised. Based on these results, we conclude that GP can be used
are generally worse, this has no obvious advantage. On average, to generate effective RRs for new problem variants of CRP.
the results obtained with GP are about 5% better than those ob- In the future work we will propose several new terminal nodes
tained with GRH. The statistical tests showed that restricted RRs for minimising total energy consumption. Furthermore, it is planned
evolved by GP perform significantly better than all other RR vari- to optimise the energy consumption in a multi-objective scenario
ants, thus confirming its superiority. Furthermore, the restricted RR with other criteria like the total number of relocations. Finally, the
variants always perform significantly better than their unrestricted model considering energy minimisation will be extended to other
counterparts, proving that the restricted variants are preferable for CRP variants including multiple bays and duplicate containers.
optimising this criterion. Finally, GP evolved RRs always perform
better than GA evolved rules, except in one case in the Zhu dataset, ACKNOWLEDGMENTS
where only GP-U and GRH-R perform equally well. Based on these This research has been supported by the Croatian Science Founda-
results, we can conclude that GP is more appropriate for designing tion under project IP-2019-04-4333 and the Spanish State Agency
new RRs, compared to GA coupled with GRH. for Research (AEI) under research project PID2019-106263RB-I00.

2200000 REFERENCES
2150000
[1] Jürgen Branke, Su Nguyen, Christoph W. Pickardt, and Mengjie Zhang. 2016.
Automated Design of Production Scheduling Heuristics: A Review. IEEE Transac-
2100000
tions on Evolutionary Computation 20, 1 (2016), 110–124. https://doi.org/10.1109/
TEVC.2015.2429314
2050000
[2] Marco Caserta, Silvia Schwarze, and Stefan Voß. 2012. A mathematical formula-
tion and complexity considerations for the blocks relocation problem. European
2000000 Journal of Operational Research 219, 1 (2012), 96–104. https://doi.org/10.1016/j.
ejor.2011.12.039
1950000 [3] Marco Caserta, Stefan Voß, and Moshe Sniedovich. 2011. Applying the corridor
method to a blocks relocation problem. OR Spectrum 33, 4 (2011), 915–929.
1900000 https://doi.org/10.1007/s00291-009-0176-5
[4] Camila Díaz Cifuentes and María Cristina Riff. 2020. G-CREM: A GRASP approach
GP-R GP-U GRH-R GRH-U to solve the container relocation problem for multibays. Applied Soft Computing
Journal xxxx (2020), 106721. https://doi.org/10.1016/j.asoc.2020.106721
[5] Tomislav Erdelić and Tonči Carić. 2019. A Survey on the Electric Vehicle Routing
(a) Caserta dataset Problem: Variants and Solution Approaches. Journal of Advanced Transportation
2019 (May 2019), 1–48. https://doi.org/10.1155/2019/5075671
[6] Mazen Hussein and Matthew E.H. Petering. 2012. Genetic algorithm-based
31500000
simulation optimization of stacking algorithms for yard cranes to reduce fuel
consumption at seaport container transshipment terminals. 2012 IEEE Congress
31000000 on Evolutionary Computation, CEC 2012 1 (2012), 10–15. https://doi.org/10.1109/
CEC.2012.6256471
30500000 [7] Raka Jovanovic, Shunji Tanaka, Tatsushi Nishi, and Stefan Voß. 2019. A GRASP
approach for solving the Blocks Relocation Problem with Stowage Plan. Flexible
30000000
Services and Manufacturing Journal 31, 3 (2019), 702–729. https://doi.org/10.
1007/s10696-018-9320-3
[8] Kap Hwan Kim and Gyu Pyo Hong. 2006. A heuristic rule for relocating blocks.
29500000
Computers and Operations Research 33, 4 (2006), 940–954. https://doi.org/10.
1016/j.cor.2004.08.005
29000000 [9] Israel López-Plata, Christopher Expósito-Izquierdo, and J. Marcos Moreno-Vega.
GP-R GP-U GRH-R GRH-U
2019. Minimizing the operating cost of block retrieval operations in stacking
facilities. Computers & Industrial Engineering 136 (2019), 436–452. https://doi.
org/10.1016/j.cie.2019.07.045
(b) Zhu dataset [10] Su Nguyen, Yi Mei, and Mengjie Zhang. 2017. Genetic programming for pro-
duction scheduling: a survey with a unified framework. Complex & Intelligent
Figure 1: Results of automatically designed RRs. Systems 3, 1 (Feb. 2017), 41–66. https://doi.org/10.1007/s40747-017-0036-x
[11] Riccardo Poli, William B. Langdon, and Nicholas Freitag McPhee. 2008. A field
guide to genetic programming.
[12] Kun-Chih Wu and Ching-Jung Ting. 2010. A beam search algorithm for mini-
6 CONCLUSIONS AND FUTURE WORK mizing reshuffle operations at container yards. , 703–710 pages.
[13] Wenbin Zhu, Hu Qin, Andrew Lim, and Huidong Zhang. 2012. Iterative
In this paper, the application of GP to automatically generate RRs for deepening A* algorithms for the container relocation problem. IEEE Trans-
actions on Automation Science and Engineering 9, 4 (2012), 710–722. https:
solving CRP with the aim of minimising total energy consumption //doi.org/10.1109/TASE.2012.2198642
was investigated. The method was tested on an extensive set of [14] Marko Ðurasević and Domagoj Jakobović. 2022. Heuristic and metaheuristic
experiments and compared with GRH, where the parameters for RR methods for the parallel unrelated machines scheduling problem: a survey. Artifi-
cial Intelligence Review (Aug. 2022). https://doi.org/10.1007/s10462-022-10247-9
are adjusted using a GA. The experimental results show that RRs [15] Marko Ðurasević and Mateja Ðumić. 2022. Automated design of heuristics for the
designed using GP significantly outperform those designed using container relocation problem using genetic programming. Applied Soft Computing
GRH. This shows the versatility of GP in obtaining high quality RRs 130 (2022), 109696. https://doi.org/10.1016/j.asoc.2022.109696