Content uploaded by Ali Rais Shaghaghi
Author content
All content in this area was uploaded by Ali Rais Shaghaghi on Jun 08, 2016
Content may be subject to copyright.
Guided Local Search for Optimal
GPON/FTTP Network Design
Ali Rais Shaghaghi, Tim Glover, Michael Kampouridis
and Edward Tsang
Abstract Fibre optic access networks are of increased popularity among network
operators. Many types of fibre network are being deployed globally to satisfy the ever
increasing users bandwidth requirements. The rate of deployments of such networks
is expected to increase in coming years, moreover this requires cost efficient, reliable
and robust network designs. Despite the relative complex structure of these networks,
designs are mostly done manually, thus design quality is not optimal. In this paper
we will introduce and propose a tree based modelling scheme and will show how the
metaheuristic search method Guided Local Search can be used to automate the design
of FTTP/GPON networks. The design optimisation will mainly focus on reducing
the deployment cost i.e finding the optimal location, type and quantity of fibre optic
equipment in order to reduce the capital expenditure (CAPEX) of such deployment
projects. Our proposed model builds a flexible optimisation framework, and results
of the GLS algorithm compared to simple local search and Simulated Annealing
show consistent optimal results.
Keywords Guided local search ·Fibre optic networks ·Network optimisation ·
Network planning
1 Introduction
With the growth of telecommunication networks and the introduction of new
applications and services, the demand for higher bandwidths is increasing rapidly.
With these increases of bandwidth demands fibre optic networks are becoming the
A. R. Shaghaghi (B
)
CCFEA, University of Essex, Colchester, UK
e-mail: araiss@essex.ac.uk
T. Glover
BT Research and Technology, Ipswich, UK
M. Kampouridis, E. Tsang
CSEE, University of Essex, Colchester, UK
N. Chaki et al. (eds.), Computer Networks & Communications (NetCom), 255
Lecture Notes in Electrical Engineering 131, DOI: 10.1007/978-1-4614-6154-8_25,
© Springer Science+Business Media New York 2013
256 A. R. Shaghaghi et al.
preferred solution. Hybrid networks with mixed use of fibre optics and copper cables
have been around for a while, however they have limited capacity. For this reason
operators are moving towards FTTP (Fibre-To-The-Premises) technologies, based on
the Gigabit Passive Optical Network(GPON), which enables them to provide services
that are ready to meet customers demand with higher bandwidths. FTTP lines are
projected to almost triple globally, from 68 million in 2011 to 198 million in 2016,
driven by uptake in China, Russia, and the US and increasing deployments from
Western European incumbent operators. Pyramid Research expects FTTH broad-
band to generate nearly $ 116 billion in service revenue by 2016 worldwide, creating
opportunities for all of the stakeholders in the value chain.1The United Kingdom
is showcasing its broadband initiative, and FTTH deployment in most of England is
growing rapidly as well.2
However, the deployment and further costs associated with FTTx.3Networks
are comparatively greater than legacy copper access networks [1]. The competitive
service-providing market motivates the network operators to design and deploy more
economic networks with relativity low capital expenditure (CAPEX). Their target
is to bring their FTTP costs down as close to the legacy copper telecommunication
networks as possible.
One of the key aspects in reducing deployment cost is to have an efficient low
cost design, meaning for a given deployment plan reducing various equipment and
labour costs associated with design.
A cost efficient design depends on positioning optical components in the under-
lying road and duct network, so as to minimise the number of components and the
length of fibre cable required. In addition to these there are several constraints that
have to be satisfied in the design and the planning of the network, such as the max-
imum outputs any specific splitter can have. All these considerations require tools
that provide efficient and robust network designs and deployment plans. In a typical
GPON deployment, an exchange area is often divided into different sections each
served by an aggregation node. Each aggregation node can be seen as root to a tree
of splitters, fibre distribution points (FDPs) and manifolds (see Fig.1). Once the
location of this equipment is determined then the planner has to design the layout of
cables from each manifold to FDPs and each FDP to a splitter respectively. These
planning tasks are time consuming and the efficiency of the design and the associ-
ated costs are dependent on the planner’s experience because of the simple fact that
they are done manually. In this paper we propose a tree-based model to represent
the network design problem and will specifically introduce a customised Guided
Local Search algorithm [2] to achieve an efficient design. The combination of the
1http://www.prnewswire.com/news-releases/ftth-lines-expected-to-triple-by-2016-finds-
pyramid-research-135128728.html
2http://www.lightwaveonline.com/business/news/Frost–Sullivan-FTTH-deployments-lend-
momentum-to-European-fiber-optic-test-equipment-markets-124229344.html
3Fiber to the x (FTTx) is a generic term for any type of access broadband network which uses fibre
optic as its main transmission medium, all starting by FTT, these variations are differentiated based
on different configurations (e.g. FTTN, FTTP, FTTH, and so on, where in the above examples “N”
denotes Node, “C” denotes Premises, “H” denotes Home).
Guided Local Search for Optimal GPON/FTTP Network Design 257
network representation model and the metaheuristic optimisation method will result
in an automated tool that enables network operators to efficiently plan and design
robust and efficient GPON/FTTP networks. The automated tool will generate solu-
tions that are considered optimal or near optimal with respect to cost and satisfying
specific design constraints. The introduction of automated planning tool in context of
GPON/FTTP networks results in these advantages: (1) Rapid generation of network
design layouts. (2) Support for exploring different scenarios by changing constraints
(3) Minimising CAPEX (4) Automatically producing bills of material (5) Providing
detailed implementation costs for techno-economic analysis.
The rest of the paper is organised as follows: Sect. 3will introduce a tree-based
representation model to solve the optimisation problem followed by description of the
metaheuristic algorithm for design automation in Sect. 4. In Sect.5we describe our
experimental design and results and Sect. 6will provide future works and conclusion.
2 Related Work
There have been several studies on optimising various types of telecommunication
networks some of which, are specifically related to fibre optic network designs.
In [3] the authors have proposed a model to optimise the design of a GPON/FTTH
network, their model considers certain green field design aspects and a mixed inte-
ger linear programming solver is used to find a near optimal solution. Their solution
promises a design with satisfactory degree of symmetry in addition to short compu-
tational time.
In [4] they propose an efficient heuristic called the Recursive Association and
Relocation Algorithm (RARA) to solve the optimization problem. Their model also
propose splitting large areas into smaller optimisation problems in order to reduce
the computation time.
The other proposed model in [4] describes an algorithm that recursively assigns
network elements to the design layout, their research provides a good theoretical
lower bound on the deployment cost for PON networks. For more complex design
cases that consider constraints such as road maps and other geographic constraints,
sub-optimal solutions can be extended from their planning approaches.
3 Model Description
When installing a new network in the access area, the majority of money has to be
spent on digging the cable ducts. Thus, minimizing the total cost is mainly a matter of
finding the shortest street paths which interconnect all optical network units (ONUs)4
4In our model we will call these points Exchange, Aggregation, FDP (Fibre Distribution Point)
and Splitter Nodes.
258 A. R. Shaghaghi et al.
Fig. 1 Logical configuration
with the optical line termination (OLT).5A city map can be represented by a graph
where the streets are the links, and the street junctions together with the ONUs and
the OLT make up the nodes. The weights of the links are set to be proportional to the
length of the respective streets. In some cases, for example, if some fibre lines exist
or if some streets are preferred to be used as duct lines, special weight values can
be assigned to theses edges. With this map representation, the optimization problem
turns into the classical minimum Steiner tree problem. This means that we want to
find a tree within a given graph which spans a designated subset of nodes in such a
way that the sum of the costs of the selected edges becomes minimal. There already
exists a number of algorithms that solve this problem exactly. Since the minimum
Steiner tree problem is NPcomplete, these algorithms have an exponential worst-
case running time. Therefore, they are not applicable in the field of network planning
where it is quite common to have a great number of nodes and edges [5].
The representation has two parts. The first maps optical components to geograph-
ical locations. Each piece of equipment is represented by a variable whose domain
ranges over possible locations. By restricting the domain we can constrain the equip-
ment to a limited geographical area.
The second part of the model describes how the optical components are connected
to each other. The optical tree is considered as a collection of clients and servers. For
example each splitter serves many FDPs, and each FDP could be served by one of
many splitters. Moving an FDP from one splitter to one another changes the connec-
tivity of the tree. This one to many relationship is represented by including a variable
for each client whose domain ranges over its possible servers. Each component may
act as both client and sever for example a splitter has FDPs as its clients and an
aggregation node as its server .
Figure 1shows one possible logical configuration and Fig.2shows how part of
this network can be laid out on a physical road network. The cables will take the
shortest path between connected points. All clients of the server passed en route to
5Known as Manifolds in our model.
Guided Local Search for Optimal GPON/FTTP Network Design 259
Fig. 2 Physical layout
the most distant client will be attached to the same cable. The cost of the network is
the cost of components plus the cost of the cables.
Design constraints are controlled by adding penalties to solutions that violate the
constraints. Design constraints include for example the capacity of junction boxes,
maximum length of cables, branch aggregation factor of splitters and so on. The total
cost to be minimised by optimisation process is : TotalCost =Componants Cost +
Cabling Cost +Penalty Cost
4 Local Search
In order to find an optimal solution to this problem we will use Local Search. Local
Search encompasses a class of algorithms which move from solution to solution in
the search space of candidate solutions by applying local changes until no further
improvements can be found or until a time limit has been exceeded. In our problem
representation a move is an assignment of a different value to one or more variables.
Table 1represents the implemented moves in this local search.
Table 1 Local search moves
Move types Description
Assignment Assign a new value to a single variable
Swap Swap the value of two variables
Join domains Moves all the clients of a server to another server
Split domain Moves half of the clients of one server to a different server
260 A. R. Shaghaghi et al.
In this section we will introduce two variants of local search; Hill Climbing and
Simulated Annealing (SA), we then describe the metaheuristic, Guided Local Search
(GLS).
4.1 Hill Climbing and Simulated Annealing
Our proposed metaheuristic will sit on top of a tailored local search scheme designed
for our proposed network optimisation model. The local search simply model solu-
tion initialization, new solution generation (neighbourhood function), and improved
solution acceptance.In Hill Climbing a move to a new solution is only accepted if it
results in an improved cost. Therefore there is a monotonic improvement in cost. The
disadvantage of this approach is that it is unable to escape from a local minimum.
Simulated Annealing attempts to overcome this by allowing an “uphill” move with
a probability that decreases over time. The allowance for “uphill” moves potentially
saves the method from becoming stuck at local optima. In our simple configuration of
SA there are 1000 iterations with the probability of move acceptance exponentially
decaying with a rate of 0.9.
4.2 Guided Local Search
Most of local search methods suffer from two disadvantages. Firstly they easily get
stuck in local minima. Secondly, in many cases we have intuition about how to guide
the search but this can not be included directly in the cost function. For example, in
the Travelling Salesman Problem, we know that long edges are undesirable though
we can not exclude them from the beginning because they may be needed to connect
remote clusters of cities in the problem. Guided Local Search (GLS) is a penalty-
based approach that sits on top of local search methods which can help solve these
problems. When the given local search algorithm is trapped in a local optimum, GLS
dynamically changes the objective function, by penalizing some selected features that
are present in this solution. This raises the cost of the solution, allowing the search
to continue. The features are chosen in such a way as to guide the search towards
promising areas by giving an incentive to remove unfavourable features. The novelty
of GLS is mainly in the way that it selects problem dependent features to penalize,
determined by two factors: the feature’s cost (i.e. influence on the objective function)
and the frequency with which it has been penalised in the search so far [2]. These
features should simply satisfy the constraint of being non trivial, meaning that they
would not appear in all solutions [6]. If Sis a set of all possible solutions the presence
of a feature fiin solution s∈Sis represented by an indicator function
Guided Local Search for Optimal GPON/FTTP Network Design 261
Ii(s)=
1s has feature fi
0otherwise
Associated with each feature fiis a cost ci, and a penalty piwhich counts the
number of times this feature has been penalised. When a local minimum is reached a
feature is chosen to be penalised by a utility function which considers the cost of the
feature and its current penalty. The utility function is defined by util(s,fi)=Ii(s)·
ci
1+pi. Features which have already been penalised are less likely to be penalised
again. This reflects the intuition that we should avoid selecting the same feature
every time. Augmenting the cost function gvia penalties on features gives us a new
objective function hdefined : h(s)=g(s)+λ· M
i=1pi·Ii(s)where Mis the number
of features defined over solutions and λis a regularization parameter. A number of
different possible features were explored for this problem and the most effective was
found to be pairs of consecutive items of equipment on a cable. The feature cost is
the cost of cable linking the two items.
5 Experimental Design and Results
The solver has to find a solution that connects all the manifolds to the exchange
whilst satisfying all the problem constraints and minimising the cost via reducing
the cabling and total equipment cost.
The selected region includes one aggregation node and one exchange, the final
solution will layout cables from the single exchange point to 85 Manifolds. The
constraints shown in Table 2describe the maximum number of connections that the
equipment could have. The number of children in this table indicates the maximum
number of clients that can be served by each item of optical equipment. Also there
is a upper bound limit for the possible number of FDPs and Splitters, which are 43
and 11 respectively.
In order to evaluate the effectiveness of GLS two solvers were compared, the first
using simulated annealing and the second using Guided Local Search with features
based on cables as described earlier. Each experiment was run 50 times over the
sample experimental data to find the optimal solution. In each case the search was
allowed to continues until no improvement had been found for more than 6 minutes.
Table 2 Equipment constraints
Connections Children Connections Children
Manifold 12 0 Exchange 100000 100
Aggregation 276 20 Splitter 4 24
FDP 24 3
262 A. R. Shaghaghi et al.
Table 3 Best cost statistics for 50 runs
GLS SA GLS SA
Mean 13958.72 14263.19 Skewness 1.015195 2.249336
Standard deviation 297.7969 690.6522 Min 13621.93 13613.12
Kurtosis 0.06251 6.583823 Max 14742.49 16977.58
Given equal execution time we are interested in the most optimal cost (minimised)
that derives from the automated design of the network. Table3depicts the statistics
for SA and GLS. The results simply imply more consistent performance from GLS
algorithm in comparison to higher standard deviation of the SA. The average cost of
the network is also smaller while using GLS. For the sake of statistical analyses we
have performed a two sample t-test with the null hypothesis that data in the vectors of
SA and GLS are independent random samples from normal distributions with equal
means and equal but unknown variances, against the alternative that the means are
not equal. The results shown in Table 3allow us to reject the null hypothesis.
For further proof of the effectiveness of our metaheuristic methods we have also
tested a simple hill climbing algorithm,6the results obtained simply shows very
poor performance i.e. the solver in this case failed to satisfy many of the problems
constrains failing to produce any acceptable solution. The nature of GLS algorithm
enables it to scape settlements in local minima therefore results prove to be more
consistent.
6 Conclusion
Designing fibre optic access networks is becoming of great interest to global tele-
com providers. In this paper we have presented an automated GPON/FTTP design
framework based on a tree-based model utilising a guided local search algorithm
to find a near optimal solution. The model structure is relatively flexible enabling
production of various network designs with different constraints and requirements.
The automated algorithm enables network designers and planners to quickly plan
GPON networks with high flexibility and near optimal solutions. We use Guided
Local Search (GLS) to eliminate the common problem of local search algorithms
getting trapped in local optimum solutions. The GLS metaheuristic tends to produce
robust results in many runs thus ensuring rapid solutions with high quality. The ability
of GLS to escape local minima provides significantly higher quality results.
6This iterative method starts with an arbitrary solution and tries to find a better solution by using
the described local search moves. If the change results in better solutions it will accept the solution
until no further improvements occur
Guided Local Search for Optimal GPON/FTTP Network Design 263
References
1. Verbrugge S, Casier K, Lannoo B, Van Ooteghem J, Meersman R, Colle D, Demeester P (2008)
FTTH deployment and its impact on network maintenance and repair costs. In: 10th anniversary
international conference on transparent optical networks ICTON (2008) vol 3 . IEEE, New York,
pp 2–5
2. Voudouris C (1999) Guided local search and its application to the traveling salesman problem.
Eur J Oper Res 113(2):469–499
3. Ouali A, Poon K (2011) Optimal design of GPON/FTTH networks using mixed integer linear
programming. In: 16th European conference on networks and optical communications (NOC),
IEEE, pp 137–140
4. Li J, Shen G (2008) Cost minimization planning for passive optical networks. In: OFC/NFOEC
2008–2008 conference on optical fiber communication/national fiber optic engineers conference,
pp 1–3, Feb 2008
5. Riedl A (1998) A versatile genetic algorithm for network planning. In: Proceedings of EUNICE,
vol 98, Citeseer, pp 97–103
6. Voudouris C, Tsang E (2003) Guided local search. Handbook of metaheuristics. In: Glover F
(ed) Handbook of Metaheuristics. Kluwer, Dordrecht, pp 185–218
