Sei sulla pagina 1di 7

International Journal of Advanced Science and Technology

Vol. 29, No. 8, (2020), pp. 1-7

Application of Hyper-heuristics in Real-World Spacecraft Trajectory


Optimization under Limited Evaluation Constraints using Hyper-heuristics

Jia Hui Ong1, Jason Teo*2


1
Faculty of Computing and Information Technology, Tunku Abdul Rahman University College
Malaysia
2
Faculty of Computing and Informatics, Universiti Malaysia Sabah, Malaysia
2
jtwteo@ums.edu.my

Abstract

Hyper-heuristics has proven to be able to solve optimization problems by utilizing its collective low-level
heuristics as its component. Although hyper-heuristics has proven to be robust and does not need
complex configurations when it is applied to a different domain, the application of hyper-heuristics has
yet to be widely used in continuous problem domains. To the best of our knowledge, hyper-heuristics has
never been tested on continuous real-world expensive optimization problems and this paper is the first
presentation of hyper-heuristics in this problem domain. Thus, hyper-heuristics performance in two real-
world spacecraft trajectory design problems is studied in this paper. The two problems can be split into
two different models which are multiple assist (MGA) missions and multiple assists with the possibility of
using deep space manoeuvres (MGADSM). The first model is a representation of spacecraft trajectory
between planets, the spacecraft is equipped with chemical propulsion only for planetocentric phase
thrust. The same trajectory model of a spacecraft is used in the second model but with the exception that
the chemical propulsion thrust can be used once anywhere between each planet trajectory. Four
algorithms DE, GA, PSO, and MVMO are tested together with hyper-heuristics tabu-search biased
ranking (HHTSBR). The results obtained for the MGA model show that HHTSBR has the best results
among all the algorithms tested while in the second model MGADSM, HHTSBR only manages to obtain
the second-best results. From the results obtained, HHTSBR again shows its ability to perform well
without major tuning towards the real-world optimization problems.

Keywords: Hyper-heuristics; expensive optimization; hyper-heuristics component selection; real-world


optimization problems

1. Introduction

Real-world expensive optimization problems require heavy computational resources for a searching
process to get a good candidate solution within a small number of function evaluation. A competition
focusing on expensive optimization problems was first held in 2014 which is the Congress on
Evolutionary Computation (CEC) 2014. A continuation of this competition was done in CEC 2015 with
harder and more complex synthetic problems were introduced. In this competition, the researcher
introduces the use of specific searching algorithms that were tailor-made to the problems given in the
competition. Their searching algorithms are specifically tailored to tackle the competition problem, hence
it is time-consuming and required a high level of expertise to tune. These specifically tailored algorithms
are too problem dependant and hence when the problem domain changes, the performance of the
algorithms drop significantly. For the tailored algorithms to be able to produce a good solution, resources
such as domain knowledge and time are needed to tune the algorithms. It is a redundant step and
resources wasting, this is more apparent when the problem is a real-world expensive optimization where
resources are limited.

ISSN: 2005-4238 IJAST 1


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

Cowling in 2001 comes up with an approach to use a pool of low-level heuristic as a searching method
(Cowling, Kendall, & Soubeiga, 2001) and name it hyper-heuristics. He then further tested the capacity of
hyper-heuristics in solving scheduling problems (Cowling & Chakhlevith, 2003). Hyper-heuristics can be
definite as a grouping of low-level heuristic approach as an approach applicable to different problem
instances. A thorough study was done by Burke (Burke et al., 2013) in reviewing the structure and
methods used for hyper-heuristics and he identify a problem of how to develop hyper-heuristics to
become a generalized approach that can be applied to all problem instances.

By referring to the challenges mentioned by Burke (Burke, et al., 2013) if hyper-heuristics can become
easy to use across problem instances then it will save the expertise and time needed to tailor an approach
that can only work well on a single problem instance.

Hyper-heuristics has not been a favourite approach to be explored in the continuous problem domain and
only a handful of research that explore the usage of hyper-heuristics (Kiraz et al, 2013; Topcuoglu et al,
2014; Maashi et al, 2014). McClymont and Keedwell (2011) explore the usage of hyper-heuristics in a
reduced number of evaluations multi-objective problem and it is the sole research that was done in an
expensive real-world problem. Hence this paper will be the first paper study the performance of hyper-
heuristics in this domain. In section three, a brief overview of the hyper-heuristics and a discussion on the
method used are presented. Experiments setup for this paper are presented in section four and at the end
of this paper, we will discuss on results obtained.

2. Hyper-Heuristics

There are two types of hyper-heuristics which selection-based or generative-based (Burke et al, 2010). In
Fig 1 the overview of the types of hyper-heuristic is shown. Selection hyper-heuristics will select and
applies low-level heuristics to modify a solution in hand (Kheiri, 2014) while generative hyper-heuristics
refer to the hyper-heuristics methodologies for generating new heuristics from the component of existing
ones (Burke et al, 2010).

FIGURE 1. Classification of Hyper-heuristics

ISSN: 2005-4238 IJAST 2


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

Breaking down heuristics selection and generation heuristic both types can be further categories into
constructive and perturbation (Hoos and Stützle, 2004). These two categories are used to differentiate the
process of solutions. A search process that produces complete solution are categories as perturbation
while the search process which produce only partial solution is categories as constructive. To enhance the
hyper-heuristic further a feedback mechanism were introduced. This feedback mechanism allows learning
to happen and it is separated into three types of feedback: learning while searching process is still running
(online), learning during pre-processing (offline) and lastly no learning included (no-learning).

2.1 Selection Hyper-heuristics

Selection hyper-heuristics are made up of two components, one is used to select heuristics to be applied
on each iteration (heuristic selection) and another component is used to filter the solutions generate (move
acceptance). Cowling (Cowling, Kendall, & Soubeiga, 2001) presented a few simple heuristic selections
in his work, they studied the performance of Simple Random (SR), Random Descent (Gradient) (RD
(G)), Random Permutation (RP), Random Permutation Descent (Gradient) (RPD (G)), and Choice
Function (CF).

• Simple random selection: selection mechanism will do a random selection from the LLH which
are based on a uniform probability.
• Random descent (gradient): selection mechanism will do a random selection from the LLH and a
continued usage of the selected heuristic until further improvement is needed.
• Random permutation: selection mechanism generate an order of the heuristics in LLH randomly
and each heuristics will be applied to the search process according to their order.
• Random permutation descent (gradient): selection mechanism generate an order of the heuristics in
LLH randomly and each heuristics will be applied to the search process according to their order but only
change to the next heuristics in line when improvement needed
• Greedy: Utilize all heuristics in LLH to generate a solution and only selects a heuristics with a
solution that has the biggest improvements.
• Choice Function: deploy scoring criteria for LLH, the scoring is based on two criteria, one the
changes in the solution quality and two duration of the heuristic usage.

Another heuristic selection method used by Cowling (Cowling & Chakhlevith, 2003) captured our
attention. Tabu search utilized it ranking systems to choose the next heuristic to be used and also to keep a
record of heuristics that did badly to be reuse. Tabu search was first introduced by Glover (1986) in an
integer programming.

FIGURE 2. Selection Hyper-heuristics Pseudocode

ISSN: 2005-4238 IJAST 3


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

The second component of selection hyper-heuristics is the move acceptance. There are three types of
move acceptance introduced by Cowling (2001), All Move (AM), Only Improving (OI) and Improving or
Equal (IE).

In AM, all solutions that were generated will be accepted without regards of if the solutions improve or
not. OI on the other hand only accept solutions will only improved performance and lastly, IE will not
accept any solutions which has a lower performance.

These move acceptance criteria can be deterministic and non-deterministic. Deterministic move
acceptance criteria will never change its decision throughout the iteration while non-deterministic move
acceptance criteria will change according to the current iteration that it is at.

There are only a few studies on hyper-heuristics that perform perturbation multi-point search and more
work has been done on single point-search (Burke, 2013). Multi-point search refers to the search that
maintains a pool of population solution while single-point search refers to search that improves and
maintains single solutions.

3. Hyper-Heuristics and Real-World Problems Setup

In this paper, Hyper-heuristics selection method used is Tabu-Search Biased Ranking (TSBR) and move
acceptance used are only improving (OI).

There are four low-level heuristics that will be used by BROI:

 Particle Swarm Optimization (PSO), initial velocity 1 and maximum velocity 3.


 Differential Evolution (DE), Cr = .9, F = .2,
 Genetic Algorithms (GA), Cr = .5, Mutation Rate = .1. Tournament selection, tournament size = 2
 Mean Variance Mapping Optimization, archive size = 3

Normal LLH that was mostly used in previous work is a simple heuristic method, in this paper meta-
heuristic which are commonly used and is considered as good solve are use as LLH. MVMO (Longatt et
al, 2012) is one of the new meta-heuristic used as LLH.

A feature from MVMO was also adapted to the LLH whereby an achieve was implemented which will
store the best three solutions this archive is used to calculate the mean and shape variable for our hyper-
heuristics to generate a new solution.

In hyper-heuristic selection, a single heuristics chosen can control the whole population for a single-
iteration. This will allow diversity in the population as different heuristics can generate a possible solution
to the population pool. In a multi-point search using island-based population was proved to be more
efficient and yield better result hence in this paper island base setup will be used.

For our experiment setup, the island-based population will be separated into two different sections. The
initial section will be used for tabu-search selection and the performance of the heuristics are then
recorded as a score. The scoring is based on whether the solution improves or not in the next iteration.

The second section will be used by the highest rank LLH based on the score from the first section. The
highest rank LLH will be able to control the whole section. These two sections are form based on a
division of the total evaluation allowed, the overview of our hyper-heuristics are shown in Fig 3.

ISSN: 2005-4238 IJAST 4


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

FIGURE 3. Selection Hyper-heuristics Overview.

4. Results

In Table 1, the results for Cassini1 is shown. HH-TSBR managed to obtain the best results with 16.48931
while coming in seconds is DE with 19.87999 and follow by MVMO with the results of 22.21619. GA
and PSO weren’t able to give good results as compared to HH-TSBR. Previously GA and PSO were able
to perform well in dynamic and noisy problems but when the problem domains changed to real-world
problems, GA and PSO weren’t able to perform well since GA and PSO need to be fine in order to obtain
a reasonable result.

In Table 2, although HH-TSBR wasn’t able to give the best results by yielding 22.5799 coming in second
behind MVMO 21.5031 while DE gave the third best results which are 25.2809 and again GA and PSO
weren’t able to give close comparable results to the previous three algorithms.

Average Functions
Algorithms
value
HH-TSBR 16.48931
DE 19.87999
MVMO 22.21619
PSO 143.1156
GA 245.7709

TABLE 1: Results for CASSINI1

ISSN: 2005-4238 IJAST 5


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

Average Functions
Algorithms
value
MVMO 21.5031
HH-TSBR 22.5799
DE 25.2809
PSO 83.1236
GA 146.722

TABLE 2: Results for CASSINI2

5. Conclusion and Future Work

Hyper-heuristics has proved to be able to perform well even it is applied to expensive real-world
problems. From this paper, we can further conclude that hyper-heuristic can be a generalized approach
that can be applied to various problem instances without heavy tuning. This is very important as it saves
time and expertise to be spent on just tuning and approaches rather than focusing on the underlying
problem instances. In real-world expensive optimization, a various factor will affect the performance of
approaches used to solve the problems and researcher might be trap in local optima due to attempts to
“perfect” an approach to suited the condition that appears in the expensive real-world problems. From the
set of compared approach with hyper-heuristics, the results show that hyper-heuristic can still achieve
good results even if it tested in different problem instances. With the usage of various heuristics collection
in hyper-heuristic, it might be an answer to No Free Lunch theorem,

For future work, more attention and work on Hyper-heuristics in the continuous problem domain and real-
world problems should be done to further verify and established hyper-heuristic as a global generalized
approach. New approaches should be explored to enhance the robustness of hyper-heuristic.

References

[1] E. K. Burke, M. Hyde, G. Kendall, G. Ochoa, E. Özcan, and J. R. Woodward. A classification of


hyper-heuristicss approaches. In Handbook of Metaheuristic,2010 vol.146 pp.449-468
[2] E.K. Burke, M. Gendreau, M. Hyde, G. Kendall, G. Ochoa, E. Özcan, and R. Qu, “Hyper-
heuristicss: a survey of the state of art,” Journal of the Operational Research Society,
64(12):1695-1724, 2013
[3] P. Cowling, G. Kendall, and E. Soubeiga, “A hyperheuristic approach to scheduling a sales
summit,” in Practice and Theory of Automaed Timetabling III, vol. 2079, pp. 176-190, 2001
[4] P. Cowling and K. Chakhlevith, Hyperheuristic for managing a large collection of low level
heuristic to schedule personnel, in IEEE Congres of Evolutionary Computation (CEC’ 03) 2003,
pp.1214-1221.
[5] F. Glover. Future paths for integer programming and links to artificial intelligence. Computers
and Operations Research 1986 .13(5):533-549
[6] H. H. Hoos, and T. Stützle. Stochastic local search: foundation and applications. Elsevier/Morga
Koufmann: San Francisco, CA. 2004
[7] S. H. Jun , 2006. A Hybrid Genetic Algorithm and New Criterion for Determining the Number of
Clusters. International Journal of Soft Computing, 1: 313-318.
A. Kheiri. Multi-stage hyper-heuristicss for optimization problems. Phd Thesis. 2014
B. Kiraz, A. S. Uyar, and E. Özcan. Selection hyper-heuristicss in dynamic
environments.Journal of the Operational Research Society , 2013. 64(12):1753–1769
[8] F. G.Longatt, J. Rueda.,I, Erlich,W, Villa. and D, Bogdano. Mean Variance Mapping
Optimization for the identification of Gaussian Mixture Model: Test case. In Proceedings IS'2012
- 2012 6th IEEE International Conference Intelligent Systems, 2012

ISSN: 2005-4238 IJAST 6


Copyright ⓒ 2020 SERSC
International Journal of Advanced Science and Technology
Vol. 29, No. 8, (2020), pp. 1-7

[9] M. Maashi, E. Ozcan, and G. Kendall. A multi-objective hyperheuristic based on choice


function. Expert Systems with Applications, 2014, 41(9):4475–4493.
[10] K. McClymont and E. C. Keedwell. Markov chain hyper-heuristicss (MCHH): an online selective
hyper-heuristics for multi-objective continuous problems.In Proceedings of the 13th Annual
Conference on Genetic and EvolutionaryComputation (GECCO '11), pages 2003–2010, New
York, NY, USA, 2011. ACM.
[11] M. Nasereddin , 2006. Using Genetic Algorithms to Find Weights for Multiple Heuristic for The
Stochastic Resource Constrained Project Scheduling Problem. International Journal of Soft
Computing, 1: 255-260.
[12] Thamarai and S. Murugavalli, 2016. Analogy Based Software Effort Estimation Based on
Differential Evolution and Hybrid Fuzzy Logic and Firefly Algorithm. Asian Journal of
Information Technology, 15: 1484-1493.
[13] H. R. Topcuoglu, A. Ucar, and L. Altin. A hyper-heuristicss based framework for dynamic
optimization problems. Applied Soft Computing, 2014, 19:236–251.

ISSN: 2005-4238 IJAST 7


Copyright ⓒ 2020 SERSC

Potrebbero piacerti anche