This article was downloaded by: [130.63.180.

147] On: 02 March 2015, At: 06:59

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Operations Research
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Hybrid Fiber Coaxial Network Design

Raymond A. Patterson, Erik Rolland,

To cite this article:

Raymond A. Patterson, Erik Rolland, (2002) Hybrid Fiber Coaxial Network Design. Operations Research 50(3):538-551. http://
dx.doi.org/10.1287/opre.50.3.538.7737

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

© 2002 INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 HYBRID FIBER COAXIAL NETWORK DESIGN
RAYMOND A. PATTERSON
4-30E Business Building, School of Business, University of Alberta, Edmonton, Alberta, Canada T6G 2R6, [email protected]

ERIK ROLLAND
The A. Gary Anderson Graduate School of Management, University of California,
Riverside, California 92521, [email protected]
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

(Received March 1999; revisions received December 1999, July 2000, February 2001; accepted February 2001)

Much interest exists in broadband network services to deliver a variety of products to consumers, such as Internet access, telephony, inter-
active TV, and video on demand. Due to its cost efficiency, Hybrid Fiber Coaxial (HFC) technology is currently being considered by most
Telcos and cable companies as the technology to deliver these products. The topological HFC network design problem as implemented by
several major companies is a form of the capacitated tree-star network design problem. We propose a new formulation for this problem
and present a heuristic based on hierarchical decomposition of the problem. The proposed solution methodology exploits an Adaptive
Reasoning Technique (ART), embedded as a meta-heuristic over specialized heuristics for the subproblems. In this context, we demonstrate
the dynamic use of an exact solution technique within ART. The generalizability of the proposed solution methodology is demonstrated by
applying it to a second problem, the Traveling Salesman Problem (TSP). Computational results are presented for both the HFC network
design problem and the TSP, indicating high-quality solutions expending a very modest computational effort. The proposed solution method
is found to be effective, and is shown to be easily adaptable to new problems without much crafting, and as such, has a broad appeal to
the general operations research community.

1. INTRODUCTION of the bus is approximately 750 MHz, which must be

Hybrid Fiber Coaxial (HFC) networks are currently being shared by every customer on the bus. Cox Cable Commu-
tested and implemented to deliver broadband network nications runs the coaxial cable bus to about 1,000 homes,
services (such as high-speed Internet access, telephony, whereas GTE and Southern New England Telephone and
interactive TV, and video on demand) as a bundled prod- Tele-Communications run it to between 200 and 600 homes
uct. HFC networks are utilized both by large corpora- (Watson 1997).
tions such as The Roadrunner Group of Time Warner A total cost exceeding $100 million to provide HFC
Inc., GTE, Southern New England Telephone and Tele- technology is commonplace. Cablevision Systems is spend-
Communications (Watson 1997), and The @Home Corpo- ing $160 million over three years to upgrade their Boston
ration (Watson 1997, Levine 1997); and by small cable TV HFC network to provide high-speed Internet access, digi-
operators such as Range TV Cable in Hibbing, MN (serving tal cable, and cable telephony (Coleman 1998). MediaOne
6,000 customers) and Friendship Cable in Heath, TX (serv- spent $200 million to acquire HFC-based telephony (Vit-
ing 8,000 customers) (Brown 1998). These operators of tore 1997). MediaOne and American Internet Corp. formed
broadband networks are just the beginning of a rush toward a joint venture with The Roadrunner Group of Time Warner
implementing HFC network topological designs. Gasman to provide Internet service to 100,000 Internet users (Wilder
(1997) predicts that by the year 2003, HFC will exist and Van Beaver 1998). Given an estimated cost of $1,000
as a mainstream technology for broadband networks in per user, this cost also exceeds $100 million. In November
North America. At an estimated installation cost of $1,000 of 1997, A2000 (a joint venture of US West Interna-
per customer (Brightman 1994), finding cheaper topolog- tional/United and Philips Communications) committed over
ical HFC network designs before construction begins in $100 million for HFC technology by Cablespan systems
earnest on nationwide implementation is vitally important (Vittore 1997). In June of 1988 AT&T agreed to acquire
to an industry typically characterized by very thin profit Tele-Communications Inc. and in February of 1999 AT&T
margins. agreed to a pact with Time Warner Inc. to provide cable-
In an HFC system, fiberoptic cables run from the head- based local phone service to over 12 million customers
end (or central office) to a neighborhood node Optical Net- (including New York City) (Cauley and Blumenstein 1999).
work Unit (ONU) (Gupta and Pirkul 1999, Adams 1997, On a much smaller scale, Range TV Cable in Hibbing,
Watson 1997). Coaxial cable is then run in a bus architec- Minnesota—with only 6,000 customers—spent $2.4 mil-
ture to serve up to approximately 2,000 customers (Gupta lion on the HFC infrastructure to provide video on demand
and Pirkul 1999, Adams 1997, Watson 1997). The capacity (Brown 1998).

Subject classifications: HFC network design: configuration of HFC topology. Tree-star: network architecture. Heuristic search: Solutions to problems in short time.
Area of review: Telecommunications.

Operations Research © 2002 INFORMS 0030-364X/02/5003-0538 $05.00

Vol. 50, No. 3, May–June 2002, pp. 538–551 538 1526-5463 electronic ISSN
 Patterson and Rolland / 539
Profit margins in the cable TV industry have traditionally the concentrators; and (4) the terminal clustering problem,
been very low, and there is little reason to believe that the where each demand point is assigned to a concentrator.
profit margins will improve significantly with the introduc- The topological HFC network design problem as imple-
tion of broadband services. With the significant amounts of mented by several major corporations (e.g., @Home, Time
resources being spent on HFC infrastructure, the topolog- Warner) is a form of the capacitated tree-star network
ical design of HFC networks is a timely and very impor- design (CTSD) problem (Gupta and Pirkul 1999). This
tant issue. Developing a methodology for finding least-cost paper addresses what we call the hybrid-fiber coax-
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

topological HFC network designs is vitally important for ial capacitated tree-star network design (HFC-CTSD)
the competitiveness of both cable TV and telephone com- problem—a version of the local access telecommunications
panies. network design problem. Gupta and Pirkul (1999) label this
In §2 of this paper we discuss HFC network design and problem as the HFC-CATV (Hybrid-Fiber Coaxial Com-
summarize related literature. A mathematical formulation munity Antenna Television) Network Design. The problem
for HFC design is presented in §3. Section 4 proposes the addressed in this paper, as well as by Gupta and Pirkul
problem decomposition. We present the solution procedures (1999), involves a multitap bus architecture, which is accu-
in §5, and discuss the computational results in §6. Our con- rately characterized by the CTSD topology (see Figure 1).
clusions are summarized in §7. In Appendix A we discuss Thus, we use the term HFC-CTSD to refer to this problem.
the details of the Adaptive Reasoning Technique (ART)
solution procedure as applied to HFC network design. In
Appendix B we further demonstrate the generalizability of 2.2. Research Related to Solution Approach
the ART solution methodology by applying the adaptive In this research we propose a two-phase algorithm that pro-
reasoning technique to the Traveling Salesman Problem. duces solutions to all four design subproblems. This paper
The appendices can be found at the Operations Research first presents a new formulation of the HFC-CTSD prob-
Home Page
http://or.pubs.informs.org in the Online Col-
lem. Given the intractability of this problem, we chose
lection.
to develop solution techniques for a hierarchically decom-
posed version of the problem formulation. Balakrishnan
2. RELEVANT LITERATURE et al. (1995) also used hierarchical decomposition to solve
Literature relevant to this paper will be explored in two a version of the local access telecommunications network
steps. First, research related to the HFC problem will be design problem. Adaptive memory is added to heuristic
examined. Second, research related to the solution approach solution techniques for the decomposed problems. Through
taken in this paper will be reviewed. the Adaptive Reasoning Technique (ART) (Patterson et al.
1999, Rolland et al. 1998, 1999), additional nonredundant
2.1. Research Related to the HFC Problem problem constraints are added. The ART algorithm learns
about the impact of the additional nonredundant problem
Several problems related to the HFC problem have previ-
constraints on the solutions found by the heuristic algo-
ously been studied. The local access telecommunications
rithms. ART adds and drops new nonredundant constraints
network design problem (Balakrishnan et al. 1995) is a fun-
damental problem in telecommunication network design. as the heuristic algorithm is repeatedly executed. The ART
This problem is one of deciding how to connect customer methodology is a generalized approach that externally mod-
demand points to a central switching center (central office, ifies the solution path used by one or more construc-
or headend). In general, full network connectivity can be tive heuristics. Thus, ART self-adapts to specific problem
achieved by adopting a tree or a star topology (or a combi- instances by learning from previous attempts at solving the
nation of these two), typically with a resource-minimizing problem instance at hand.
objective. Additional technology can be introduced into this Adaptive memory programming techniques, such as tabu
problem to enable the shared use of high-capacity links. search and ART, belong to the class of “generate-and-test”
That is, we often seek to use a high-capacity network struc- search techniques (Glover 1997, Patterson et al. 1999).
ture that feeds into lower-capacity areas of the network. Tabu search typically operates with classical transition
The connection points between the high-capacity links and neighborhoods (as used in local search), but can also
the remaining links are equipped with electronic equipment exploit constructive neighborhoods (Glover 1965, 1977;
often referred to as concentrators. In general, this network Patterson et al. 1999). The memory constructs of the con-
design problem can be viewed as consisting of four sepa- structive tabu search technique (Fleurent and Glover 1998)
rate subproblems (Gouveia and Lopes 1997, Gavish 1991): and ART are similar.
(1) the terminal layout problem, where the actual connec-
tions between demand points are determined; (2) the con-
3. PROBLEM DEFINITION AND FORMULATION
centrator quantity problem, where the task is to determine
the required number of concentrators in order to design an The HFC-CTSD problem is one of connecting customers
(cost) efficient network structure; (3) the concentrator loca- in capacitated tree architecture networks to a wide area
tion problem, where one selects the actual locations for network through an Optical Network Unit (ONU) to a
 540 / Patterson and Rolland
Figure 1. HFC-CTSD network topology (adapted from Gupta and Pirkul 1999).
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

central office (or headend). An HFC-CTSD network can be rooted in i. Hence, the total demand of the subtree branch
described in graph theoretic terms as follows: rooted in i does not exceed the capacity of the ONU placed
1. The HFC-CTSD network is represented by a directed at site i. Note that each subtree branch of T rooted in i
graph with a designated root vertex V ∗ representing the contains at most K nodes when uniform demand is used
central office. (where K depends on the ONU size). For each node i con-
2. A link (fiberoptic or coaxial) between site i and the taining an ONU, the destination of the arc from i must be
central office is represented by an arc between nodes i and V ∗ . The selected ONUs may have varying capacities and
V ∗ in the graph. The capacity of arc (i V ∗ ) corresponds to capacity-dependent costs.
the capacity of the ONU installed at site i. As stated previously, this problem is graphically rep-
3. A coaxial link between sites i and j is represented by resented in Figure 1, and as discussed above, numerous
an arc i j or j i connecting nodes i j ∈ V \V ∗ . cable TV companies currently experience a version of this
Given this, the input to the problem is a directed graph problem in practice as they are augmenting their services
G = V  A with a designated root vertex V ∗ , demand vec- with Internet and telephony access. Further, the HFC-CTSD
tor D where dv is the demand at node v, a vector of arc problem was shown in Gupta and Pirkul (1999) to be NP-
capacities, and a vector of arc costs. The arcs in A repre- complete (by reduction).
sent potential links in an HFC network. Because there are We now formulate the HFC-CTSD problem. We assume
a variety of ONU sizes, there are parallel arcs between a that an ONU of any capacity can be located at any demand
node location. Thus, every demand node location also acts
node i ∈ V \V ∗ and V ∗ , each with a corresponding cost and
as a potential ONU site. To formulate the HFC-CTSD, we
capacity. The output of the problem is a network topology
introduce the following notation:
represented by a subset of the arcs A ⊂ A that induce a
directed spanning tree T of V rooted in V ∗ , such that for I set of demand nodes, i = 2     n
each node i containing an ONU, the following condition is J  the set consisting of the headend (or central office) ∪ I,
met: The capacity of arc i V ∗  ∈ A must be at least as j = 1     n
large as the total demand of the nodes in the subtree of T K set of ONU types, k = 1     m.
 Patterson and Rolland / 541
ONU type k = 1 does not use a concentrator, but rather objective function then implicitly minimizes the total cost
is a coaxial connection. ONU types 2 through m do use of the communication arcs and ONUs that are used. The
concentrators. We further assume that the following data is flows emanating at the demand nodes are sent to the head-
available: end. Constraint set (2) implies that the flow from every
demand node, except the headend, is sent to one and only
di  demand at node i. one other node, using at most one ONU. In this formula-
cijk  cost of linking node i to node j using ONU of type k tion, node 1 is designated as the central office (or headend).
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

k  capacity of ONU type k. Constraint set (3) implies that the cumulative flow coming
out of every demand node is equal to the local demand
Our decision variables are:
 plus the cumulative load going into that node. Constraint

1 if arc i j using ONU type k is included set (4) assures that the cumulative flow coming out of any
xijk = in the solution node will not exceed the capacity of the outgoing arc, over

 all ONU types. When demand is uniform, then the num-
0 otherwise
ber of nodes in the subtree is limited by constraint set (4).
Auxiliary decision variables are: Otherwise, the number of nodes allowed in the subtree is
implied by constraint set (4). In constraint set (5) we make
fij = the flow on the arc i j sure that ONUs are only used if the flow coming out of the
node is going directly to the headend. Note that ONU type
The flow variables capture the flow on arcs (i j). The flow k = 1 is the same as not using a concentrator, equivalent to
must be tracked to determine the appropriate concentrator using a coaxial connection. Constraint set (6) ensures the
capacity. A unit of flow from a demand point to node 1 binary properties of this problem. Note that this mathemati-
traces out the physical path of links that will be used to cal representation models an access network where there is
serve that demand. These demand flows are aggregated in no direct communication between noncentral nodes. This is
fij . As such, the flow variable is considered to be an aux- typical of local access networks, and currently popularized
iliary decision variable. by the expansion of the cable TV network to accommodate
The HFC-CTSD problem formulation is as follows: Internet services.
Problem HFC-CTSD: This formulation of the problem is new. Utilizing a deci-
  sion variable with three indices (xijk ) and then adding con-
 n  n  m  straint set (5) to eliminate the unwanted variables is a
Minimize Z = cijk xijk (1) unique contribution. Without constraint set (5), the model is
 i=2 j=1 k=1 
j =i very general with the possibility of locating multiple ONUs
Subject to: on a single subtree. Constraint set (5) prohibits all ONUs
except those which would connect directly to the headend.

n 
m
Various related versions of this problem have been sug-
xijk = 1 ∀ i ∈ I (2)
j=1 k=1
gested in the literature. Gupta (1996) and Gupta and
j =i
Pirkul (1999) proposed a Lagrangian relaxation of the prob-

n 
n lem using additional flow constraints. The results obtained
fij − fji = di ∀ i ∈ I (3) in this paper indicate potential for significant improve-
j=1 j=2
j =i j =i ments, as the average gaps between the obtained upper and

m lower bound are approximately 10%. Lee et al. (1996) pre-
fij
k xijk ∀ i ∈ I j ∈ J  and i = j (4) sented a generalized problem formulation for capacitated
k=1
networks with a tree configuration. They test a branch-and-

n 
n 
m
cut algorithm for networks composed of only 5, 10, and
xijk = 0 (5)
i=2 j=2 k=2
20 nodes. Research on a multicenter version of the capaci-
j =i
tated minimum spanning tree (CMST) is also related to the
xijk ∈ 0 1 ∀ i ∈ I j ∈ J  k ∈ K (6) HFC-CTSD, particularly when the centers are connected
in starlike fashion (McGregor and Shen 1977, Narasimhan
We assume fixed capacity on each arc of k, but addi- 1990).
tional capacity from any node directly to the headend can
be bought by opening an ONU of type k. The costs of these
ONU alternatives differ. The formulation accommodates 4. HIERARCHICAL PROBLEM DECOMPOSITION
nonuniform load requirements (demand) at each node (but In our investigation, problem HFC-CTSD proposed above
experiments in this paper will focus on uniform demand). could only be solved to optimality (using CPLEX version
The formulation also accommodates nonuniform arc costs 4.0, a standard mixed-integer linear-programming tool, on a
of cijk for establishing a link between demand nodes i and Sun Ultra as described below) for small problem instances
j using an ONU of type k. We here assume that k = 1 (up to 20 nodes). However, the HFC-CTSD problem can be
denotes a demand node that is NOT using an ONU. The solved suboptimally by hierarchically splitting the problem
 542 / Patterson and Rolland
into two well-known subproblems, and utilizing heuristics Because the subproblem solution found to the CMST is
for the subproblems. The first subproblem is the CMST integrated into the CSSC subproblem, both the CMST and
problem with uniform demand and uniform capacities (see, CSSC affect the solutions to the four separate subprob-
for example, Patterson et al. 1999 or Rolland et al. 1999 lems related to using capacitated trees for designing the
for a formal definition of the CMST). The CMST plays topology of local access networks (terminal layout prob-
an important role in the design of telecommunications net- lem, concentrator (or ONU) quantity problem, concentrator
works. The CMST problem is NP-complete (Papadimitriou (or ONU) location problem, and terminal clustering prob-
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

1978). The CMST has in the past been solved using deter- lem (Gouveia and Lopes 1997, Gavish 1991)). However,
ministic heuristics (Kershenbaum and Chou 1974, Esau the CMST primarily impacts the terminal layout problem
and Williams 1966, Altinkemer and Gavish 1986, 1988), and the CSSC primarily impacts the ONU quantity prob-
memory-based heuristics (Patterson et al. 1999, Rolland et lem, ONU location problem, and the terminal clustering
al. 1999, Sharaiha et al. 1997, Amberg et al. 1996), cutting- problem.
plane methods (Gouveia and Martins 1995, Hall 1996), and The formulation for the HFC-CTSD problem above
integer programming formulations (Gavish 1983, Gouveia incorporates both the CMST and the CSSC subprob-
1993, 1995). The heuristic approaches are often faster than lems. The first problem is the classical CMST problem
the cutting-plane and integer programming approaches, but as described above. This is extracted from the formulation
above by limiting k to a single choice (no concentrators
the quality of the solutions are typically not as good.
are used). An illustration of the final solution to the CMST
The second subproblem is the capacitated star-star with
is depicted in Figure 2. For this illustration we have three
concentrators (CSSC) problem. In keeping with the cable
capacity levels and set 1 = 5 2 = 10, and 3 = 15.
TV nomenclature, concentrators are hereafter referred to
Next, the CSSC problem can be superimposed on the
as ONUs. Both demand and capacity are variable (nonuni-
solution configuration given in Figure 2. This is done by
form). This problem is referred to by Soltys et al. (1997) as
linking each subtree to the headend node either directly, or
the Local Access Optimization problem. This name is a bit
via a potential ONU site. Thus, for each subtree, the node
confusing because Balakrishnan et al. (1995) use the term linking the subtree to the headend node is deemed to be a
local access to imply a hierarchical tree structure. Given potential ONU site. In Figure 3, the triangles mark potential
this confusion over the precise definition of local access, ONU sites.
we use CSSC to more accurately denote this subproblem. Each subtree is then represented by a single node, which
From the CMST subproblem, each subtree branch from is the potential ONU linking the subtree to the headend
node V ∗ is considered to be a metanode. The CSSC prob- (see Figure 4).
lem is one of connecting the metanodes to a wide area net- The cost of connecting one subtree to the ONU of
work either directly or through an Optical Network Unit another subtree is equal to the minimum cost arc from the
(ONU) to a central office (or headend). The resulting con- first subtree to the ONU. Of course, capacity must be avail-
figuration will be a star-star network. The location of a able in order to connect a subtree to an ONU. For the
potential ONU is deemed to be the location of the node on CSSC, the potential ONU sites are now limited to those
each subtree branch that connects directly to the headend selected by the CMST. Note that there is nothing in the
(node V ∗ ). The cost of connecting each metanode (sub-tree formulation that prevents the solution from reverting to a
branch) to a potential ONU or headend is the minimum CMST solution itself as long as ONU nodes are not needed.
cost of connecting any node in the subtree branch in ques- However, if capacity were available and it were cheaper
tion to the ONU or headend. to link one subtree to another without an ONU, then this
Graph notation can be used to accurately describe the already would have been done by the CMST solution. Thus,
CSSC problem. Given a graph, GV  A, where V is the a constraint prohibiting tree formation among non-ONU
set of customer demand metanodes (with the associated subtrees is unnecessary. This can be proven as follows:
demand vector Dv ), A is the set of possible arcs in the graph Proof. Let the flow into a potential ONU A be denoted as
(with the associated arc cost vector Ca , an arc capacity of dA , and the flow into another potential ONU B be denoted
K for each subtree branch rooted from an ONU node, the as dB . Further, assume that triangular inequality holds.

associated ONU arc cost vector cak and arc capacity of k Now, if dA + dB  > k, where k is equal to the coaxial
for the demand of all metanodes rooted in a particular ONU line capacity (1 from the problem formulation) without
and routed via a single arc a to the headend at capacity using a higher capacity ONU, then A and B cannot be
level k), and V ∗ is a designated headend node, the objec- connected at either A or B without a higher capacity ONU.
tive is to find a minimum-cost spanning tree that is rooted With the inclusion of a higher capacity ONU at either A
in node V ∗ through any selected ONUs at any capacity k or B, then it follows from constraint set (5) that the flow
where the total demand of each subtree branch rooted in must go directly from the ONU to node 1 (the headend).
node V ∗ does not exceed the capacity of the selected ONU On the other hand, if dA + dB  <= k, then A and B would
(k ) and a maximum of two arc connections from each have been connected into a single subtree in the optimal
metanode to node V ∗ are allowed. The selected ONUs may CMST solution. Otherwise, the solution would not have
have varying capacities and capacity-dependent costs. been optimal. 
 Patterson and Rolland / 543
Figure 2. Illustration of CMST problem solution.
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

The final solution to the CSSC problem is depicted in the headend (we assume for simplicity that the cost of con-
Figure 5. From this figure, we see that those subtrees that structing these lines is included in the ONU cost).
are not directly connected to the headend are reconnected By recombining the CMST and the CSSC solutions, we
to an ONU. The capacity of an ONU (and the load com- arrive at a feasible solution configuration for the HFC-
ing through it) may vary between ONUs, as is denoted by CTSD problem. This solution, which may be suboptimal, is
the differently sized trunk lines going from the ONUs to depicted in Figure 6. We see that the HFC-CTSD solution

Figure 3. CMST problem with addition of potential ONU sites.

 544 / Patterson and Rolland
Figure 4. Illustration of initial CSSC problem with potential ONU sites.
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

has individual demand nodes linked together in subtrees in practice only makes a handful of mistakes. These mis-
(found in the CMST solution), and these trees are con- takes can be divided into two categories: primary mistakes,
nected to either an ONU or to the headend directly (found that are due to the greedy selection criteria; and secondary
in the CSSC solution). mistakes, that are due to limited available choices imposed
The next section introduces heuristic solution procedures by earlier primary mistakes. The primary mistakes are the
for solving the HFC-CTSD problem utilizing the partition- crucial ones, and, given an opportunity, these are the mis-
ing mechanism described above. takes that we would seek to prevent. A similar idea has
previously been proposed for network optimization and
5. THE ADAPTIVE MEMORY the CMST problem (Karnaugh 1976, Gouveia and Lopes
SOLUTION PROCEDURE 1997). These approaches, called second-order heuristics,
For most combinatorial optimization problems there exists share some constructs similar to those proposed below, but
a greedy heuristic that produces solutions for the problem. lack much of the memory functions we use herein. The
We know that for Matroid problems (such as the uncapac- second-order heuristics were found to be successful for the
itated version of the Minimum Spanning Tree problem), CMST problem (Kershenbaum et al. 1980), but required
a greedy heuristic always produces the optimal solution. more computational effort than the simpler heuristics (i.e.,
Many greedy heuristics are constructive; that is, they build the Esau-Williams’ (1966) procedure).
up complete solutions from scratch, one solution compo- The Adaptive Reasoning Technique (ART) attempts to
nent at the time. While studying the behavior of a greedy “learn” about the primary mistakes of a heuristic. In the
heuristic, such as the Esau and Williams (1966) procedure ART framework, the greedy heuristic is executed repeatedly,
for the CMST, one will often see that the greedy heuristic and for each new execution we probabilistically introduce

Figure 5. CSSC solution with selected ONU sites.

 Patterson and Rolland / 545
Figure 6. Illustration of HFC-CTSD solution with selected ONU sites.
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

constraints that may prohibit certain solution elements from certain number of time periods. Additional nonredundant
being considered by the greedy heuristic. The prohibitions constraints whose time duration has expired are removed.
may last for more than one iteration, and as such, we may (4) Repeat from step 1 until we have reached a preset
at any time have a collection of active constraints due to maximum number of iterations.
the prohibited solution elements. The length of the prohi- A depiction of ART as applied to the HFC-CTSD prob-
bition is a function of the cost of the solution element and lem is given in Figure 7. A detailed description of the com-
a random component. The active constraints are held in a plete ART heuristic as implemented for the HFC-CTSD
short-term memory. A long-term memory holds information problem is given in Appendix A.
regarding which constraints were in the active memory for
the best set of solutions (for example, for the best 10 solu-
tions found so far). We can now augment these two memory Figure 7. ART components and execution cycle.
functions with some basic principles of learning. For exam-
ple, we can impose a certain degree of memory loss dur-
ing the execution of the algorithm (some amount of mem-
ory loss is often important, even in human memory, since
it enables us to explore “old territory” in light of newly
discovered evidence). We can also impose a “propensity to
learn” control variable, which can be varied over the exe-
cution time of ART (this is similar to a principle observed
in humans: We are typically less “willing” to learn as we
grow older). The end result of the ART algorithm is a set of
prohibitions that, when used in conjunction with a greedy
algorithm, would enable us to find an optimal, or close to
optimal, solution.
The execution cycle of ART is as follows:
(1) EW heuristic is used to solve the CMST problem,
subject to any additional nonredundant problem constraints
created by ART (the ART method as applied to the CMST
subproblem alone is fully described by Patterson et al.
(1999) and Rolland et al. (1999)).
(2) A solution procedure is applied to solve the CSSC
problem (see below).
(3) The combined HFC-CTSD problem solution found
in steps 1 and 2 is evaluated. Additional nonredundant
constraints are added through the ART mechanism for a
 546 / Patterson and Rolland
In the context of the HFC-CTSD problem, we need at CPLEX to find an optimal solution to the CSSC subprob-
least two primary heuristics: one for the CMST and one for lem. When used within the ART solution framework, the
the CSSC problem. We elected to use the Esau-Williams AMPL data file is modified at the beginning of each iter-
(EW) heuristic for the CMST. For the CSSC subproblem ation to include all nodes connecting directly to the head-
we compared the effect of two methods: a simple greedy end as potential concentrators as given by the ART/Esau-
procedure, as well as an exact solution approach (AMPL Williams CMST solution.
with CPLEX). In the CSSC subproblem, the set of poten- The AMPL with CPLEX solution procedure is based on
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

tial ONU sites is given by the CMST heuristic. All nodes the following formulation of the CSSC problem. Let
(in the CMST solution) that are directly connected to the
headend are deemed to be potential ONU sites. I set of subtrees identified in the CMST solution,
For the greedy procedure, a cost approximation is used i = 2     n
to decide whether or not to open an ONU. The cost of con- J  the set consisting of the headend ∪ I j = 1     n
necting a subtree to an ONU is approximated by the sum K set of ONU types, k = 1     m.
of the minimum cost arc between the subtree and the ONU
plus a fixed-cost allocation (F ). The fixed-cost allocation We further assume that the following data is available:
is based on full-capacity utilization for the largest available
capacity of the ONU: di  demand emanating from subtree i (known from the
CMST solution).
F = fixed cost of ONU/maximum ONU capacity ∗ cijk  minimum cost of linking subtree i to node j using
demand on subtree ONU of type k.
k  capacity of ONU type k.
When computing the objective function for the HFC-
CTSD, actual costs utilizing minimum ONU necessary to Our decision variables are:
handle actual demand are used. The greedy CSSC proce- 
dure (GCSSC) can now be described as follows: 
1 if arc i j using ONU type k is included
xijk = in the solution
Procedure GCSSC 

0 otherwise
1. For the connections between all valid source subtrees
(a valid source subtree is a subtree whose connection to Problem CSSC
the headend has not yet been designated as an ONU site,  
nor has this subtree previously been connected to another  n  n  m 
ONU) and all available ONUs (an available ONU is a node Minimize Z = cijk xijk (7)
connected directly to the headend node in the CMST sub-  i=2 j=1 k=1 
j =i
problem which has either been declared to be an ONU site
or whose subtree has not yet been connected to another Subject to:
ONU), do:

n 
m
2. Find the minimum cost connection subject to avail- xijk = 1 ∀ i ∈ I (8)
able capacity at the largest capacity level (using adjusted j=1
j =i
k=1
GCSSC costs as described above, only connections to
ONUs with cost savings over connecting directly to the 
m 
n 
m
M xj1k  xijk ∀ j ∈ I (9)
headend are considered). k=1 i=2 k=1
i=j
3. Record the lowest cost connection. Prohibit this
source subtree from being a potential ONU site and declare 
n 
m 
n 
m
di + dj xjik
kik xijk ∀ i ∈ I (10)
it to be a nonvalid source tree. Declare the subtree with the
j=2 k=1 j=1 k=1
“winning” ONU to be an ONU site and a nonvalid source
tree. Repeat from step 1. 
n 
n 
m
xijk = 0 (11)
4. If no such subtrees are found, then go to step 5. i=2 j=2 k=2
5. Compute the cost of the new solution, and output j =i

solution configuration. xiik = 0 ∀ i ∈ I k ∈ K (12)

The GCSSC procedure is fairly simple, and numer-
xijk ∈ 0 1 ∀ i ∈ I j ∈ J  k ∈ K (13)
ous alternative implementations for solving the CSSC are
possible. One alternative is to optimally solve the CSSC where M is a sufficiently large number.
subproblem. The approach is reasonable when the CSSC Constraint set (8) ensures that each subtree will be con-
subproblem is fairly small. An exact solution method can nected to either an ONU or to the headend. Constraint set
be stated as follows: (9) ensures that only open ONUs accept demand from other
Procedure EXACT subtrees. Constraint set (10) ensures that the cumulative
1. For the connections between all valid source subtrees demand into an ONU plus local demand is less than the
(defined above) and all available ONUs, use AMPL with capacity of the selected ONU. Constraint set (11) ensures
 Patterson and Rolland / 547
that ONUs are only connected directly to the headend. Con- (labeled CPLEX). Optimal solutions were obtained using
straint set (12) simply ensures that no node or ONU can be AMPL with CPLEX, using the HFC-CTSD formulation
linked to itself. The binary nature of the decision variable presented above. The Esau-William CMST solution is a
is captured in constraint set (13). It should be noted that feasible solution to the HFC-CTSD problem (using no
the decision analog of this problem has been proven to be ONUs), and these optimality gaps are reported in column
NP-complete (Gupta and Pirkul 1999). 2 (marked EW-CMST). The EW-CMST solution in col-
In the next section we present some computational expe- umn 2 is a baseline solution without ONUs, and must be
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

riences with these algorithms, and in Appendix B we show greater than or equal to an optimal ONU-based solution.
the generalizability of ART by applying our algorithm to The CPU times associated with these solutions are always
the Traveling Salesman Problem. less than one second. The solution in column 3 is the
ART solution using EW-CMST plus the GCSSC heuris-
6. COMPUTATIONAL EXPERIMENTS tic as the heuristic engine to assign ONUs (marked ART
EW+GCSSC). The associated CPU times are found in col-
To analyze the model and algorithms presented in this
umn 7. In column 4 we report the upper bound solution
paper, we generated 80 test problems as follows: X and
found using the Gupta and Pirkul (1999) Lagrangian relax-
Y coordinates for each demand node were generated ran-
ation with the associated computational times in column 8.
domly in a 100 by 100 square with the headend node in the
In column 7 we report the optimality gaps from the solu-
center, and the arc cost without use of ONUs (cijk , where
tion constructed by EW-CMST using procedure EXACT to
k = 1) is equal to the Euclidean distance between the nodes.
find optimal ONU assignments for the CSSC sub-problem
The variable arc cost for arcs using ONUs (of level k > 1)
(EW + EXACT). The associated computational times are
are set as follows:
found in column 9. The EW + EXACT heuristic is used as
cijk = cijk ∗ 1 + k − 1∗ 01 for k > 1 the solution engine for the ART heuristic using procedure
EXACT to find improved solutions; results are reported in
This function is commensurate with recent pricing of con- column 6 (marked ART EW + EXACT) and the associated
centrators, where additional capacity (ports) typically cost CPU times are in column 10.
10% more than the smaller capacity. All nodes have unit Tables 2, 3, and 4 contain upper and lower bounds for
demands. Three ONU levels were used, where the lowest the problems (labeled CPLEX Bounds in columns 1 and
level can be viewed as having no ONU (k = 1). Capac- 2), and were obtained using AMPL with CPLEX (version
ities are four for level one, 10 for level two, and 25 for 4), with the HFC-CTSD formulation stated in §1. We show
level three. For each arc, the fixed-cost allocation for the the details for 10 problems each from node sizes 40, 60,
ONUs are as follows: zero for level one (no ONU), 10 for and 80. The CPLEX upper bounds, lower bounds, and CPU
level two, and 19 for level three. Four groups of problems times are provided to demonstrate just how rapidly these
were generated in the first data set: 20 problems with 20 problem instances become intractable (the memory capac-
nodes (19 demand nodes and 1 headend node), 20 problems ity of the Sun Ultra prevented us from getting further results
with 40 nodes (39 demand nodes and 1 headend node), 20 for these problems). No optimal solutions are known for the
with 60 nodes (59 demand nodes and 1 headend node), 40, 60, and 80 node problems. We report the CPU times in
and 20 with 80 nodes (79 demand nodes and 1 headend column 13, indicating the time CPLEX ran before exceed-
node). Optimal solutions were obtained for all the 20-node ing its memory limitations. The Gupta and Pirkul (1999)
problems using CPLEX (version 6). Larger problems could Lagrangian lower bound is shown in column 3. All other
not be solved to optimality using mixed linear integer- columns in Tables 2, 3, and 4 are identical to those of
programming tools (see below), but a Lagrangian relaxation Table 1 except that the percentage gaps are based on the
technique (Gupta and Pirkul 1999) was used to provide Lagrangian lower bound. We mark the best solution results
lower bounds as well as a comparative upper bound. for each problem in the tables in italics. The legend for the
A single processor (300 MHz) SUN Ultra 30 computer tables is presented separately.
with 128 M of RAM and 1 Gb of swap space, running From Tables 1, 2, 3, and 4 we see that using the
Solaris 2.6 CDE version 1.2 was used for all computa- EW+EXACT heuristic procedure produces better solutions
tions, and all computational times are reported in seconds. in significantly less time than the ART with EW + GCSSC
We heuristically solved all 80 problems using the ART and Lagrangian upper bound heuristics. We also see that
procedure, using the two alternative primary heuristics for ART with EW + EXACT substantially improves the solu-
the CSSC subproblem (GCSSC and EXACT) as described tions (as compared to all other methods). Results for the
above. CPLEX version 6 was used to solve the EXACT EW-CMST heuristic (without ONUs) are provided as a
portion of the heuristic procedures. Details of the compu- baseline measure. We stress here that the CSTD is a design
tational results are provided in Tables 1 to 4. problem, and as such the CPU times are not as crucial as for
In Table 1 we have reported optimality gaps from five a real-time optimization problem. However, telecommuni-
different solution approaches. The optimal solutions are cations network planners often solve these problems as part
reported in column 1, and the CPU times associated with of case studies where they have to solve a series of related
obtaining these solutions are reported in the last column problems with varying demand patterns and costs. In such
 548 / Patterson and Rolland
Table 1. Computational results for 20-node problems.
Optimality Gaps CPU Times

Optimal ART ART ART ART CPLEX

Solution EW-CMST EW + GCSSC LAG UB EW + EXACT EW + EXACT EW + GCSSC LAG EW + EXACT EW + EXACT (optimal)

317 1009% 2.21% 2.52% 0.32% 0.32% 3.00 24.50 <1 4660 15940
380 474% 4.21% 3.95% 0.79% 0.79% 2.90 21.10 <1 4300 29280
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

323 991% 6.81% 2.48% 1.86% 0.31% 3.00 27.90 <1 4760 16440
314 255% 1.59% 1.59% 1.27% 0.32% 3.00 30.50 <1 4340 45090
380 026% 0.26% 0.53% 0.00% 0.00% 2.90 29.50 <1 4750 90720
365 466% 3.84% 3.84% 2.74% 0.82% 2.90 28.70 <1 5050 53160
292 377% 1.37% 1.37% 2.40% 0.00% 2.90 25.20 <1 4750 37140
314 1115% 3.18% 1.59% 1.27% 1.27% 3.00 23.11 <1 4530 3380
329 1033% 2.74% 4.86% 7.29% 1.22% 3.10 30.95 <1 4890 529530
302 166% 1.66% 1.32% 0.33% 0.33% 2.90 26.39 <1 4660 440900
365 548% 3.56% 0.82% 5.48% 1.10% 3.00 21.45 <1 4480 317370
295 847% 8.47% 3.05% 6.44% 5.76% 2.90 25.18 <1 4990 630
358 447% 1.68% 1.96% 3.63% 0.00% 3.00 33.24 <1 4690 102140
328 732% 5.49% 1.22% 4.88% 2.13% 3.00 21.38 <1 5160 4870
323 217% 2.17% 2.17% 2.17% 0.62% 2.90 29.79 <1 4990 111480
316 949% 4.75% 4.75% 7.28% 2.85% 3.10 26.83 <1 4680 330
372 1129% 2.42% 2.96% 6.18% 0.27% 3.00 26.03 <1 4500 64830
330 545% 3.64% 3.03% 1.82% 0.00% 3.10 27.26 <1 5050 1194750
338 621% 0.59% 0.00% 3.25% 0.00% 3.00 30.47 <1 4600 40630
261 536% 4.60% 4.60% 0.77% 0.00% 2.90 23.10 <1 4980 2200
Average: 624% 3.26% 2.43% 3.01% 0.91% 2.98 26.63 <1 4741 155041

Legend for Tables

EW-CMST: Results from the Esau-Williams CMST heuristics

ART EW + GCSSC: Adaptive Reasoning Technique with EW for CMST and procedure GCSSC for CSSC
LAG UB: Best feasible solution generated by Lagrangian relaxation (Gupta and Pirkul 1999)
LAG LB: Best lower bound generated by Lagrangian relaxation (Gupta and Pirkul 1999)
EW + EXACT: EW for CMST and procedure EXACT for CSSC
ART EW + EXACT: Adaptive Reasoning Technique with EW for CMST and procedure EXACT for CSSC

planning, the trade-off between solution quality and CPU However, the solution quality is also severely affected by
time becomes more complicated, and the EW + EXACT not using an optimal solution method for the CSSC. More-
solution method is an excellent choice because it is much over, comparing the ART EW + EXACT solution times to
faster than the Lagrangian heuristic with comparable (and those of CPLEX (for solving the entire CTSD problem)
often slightly better) solution values. we see that the heuristic requires only a fraction of the
As seen in the tables above regarding the ART heuris- computational effort required by CPLEX. Note also that
tic, if we do not use an exact solution technique for the this substantial effort by CPLEX does not even guarantee
CSSC subproblem, the CPU times are severely reduced. that an optimal solution is found with that methodology

Table 2. Computational results for 40-node problems.

Solution Values Lagrangian Lower Bound Gaps CPU Times

CPLEX Bounds LAG ART LAG ART ART ART

Upper Lower LB EW-CMST EW + GCSSC UB EW + EXACT EW + EXACT EW + GCSSC LAG EW + EXACT EW + EXACT CPLEX

565 448 49870 1610% 1229% 1229% 1189% 728% 1720 21790 <1 10290 1823780
936 462 53671 1440% 1235% 937% 807% 639% 1820 24430 <1 10700 1861470
1004 424 52398 1432% 897% 840% 764% 573% 1680 26060 <1 9830 1691940
584 480 51396 1421% 1013% 1129% 682% 546% 1710 25400 <1 9260 2450020
588 437 50435 1381% 885% 1183% 865% 746% 1730 21920 <1 9950 2541190
847 489 57278 1610% 1174% 877% 1104% 563% 1930 25890 <1 10070 2565590
841 470 52531 2754% 1707% 1422% 1860% 1041% 1690 29140 <1 11050 2037040
681 440 49778 1873% 1190% 1230% 1049% 788% 1640 27120 <1 9530 3042988
586 520 54869 1445% 862% 789% 844% 480% 1720 28860 <1 11150 3477084
651 485 53231 1272% 577% 708% 671% 182% 1840 25680 <1 11030 3454928

Average: 1624% 1077% 1034% 983% 629% 1748 25629 <1 10286 2494603
 Patterson and Rolland / 549
Table 3. Computational results for 60-node problems.
Solution Values Lagrangian Lower Bound Gaps CPU Times

CPLEX Bounds LAG ART LAG ART ART ART

Upper Lower LB EW-CMST EW + GCSSC UB EW + EXACT EW + EXACT EW + GCSSC LAG EW + EXACT EW + EXACT CPLEX

742 610 652025 23.31% 1303% 1319% 951% 766% 4970 105360 <1 18990 8162974
1154 518 624143 24.49% 1600% 1456% 1376% 1247% 5020 110490 <1 19280 5853411
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

734 573 622472 31.41% 1278% 1438% 1229% 860% 4920 109040 <1 23510 5084951
903 609 664555 28.36% 1376% 1391% 1241% 1000% 4910 132780 <1 20700 3116187
743 631 666892 24.01% 616% 901% 856% 496% 5070 113990 <1 19710 2308807
1162 579 65452 28.19% 1337% 863% 893% 649% 4910 98840 <1 17470 4305454
794 577 604498 25.89% 1762% 1299% 1497% 1365% 5020 96950 <1 21830 3773778
1208 617 675951 28.12% 1554% 1066% 1125% 1022% 5210 122370 <1 17660 4312363
1021 646 677118 26.12% 1519% 899% 1150% 500% 5090 105770 <1 18620 4169509
720 585 610072 27.53% 1540% 1031% 1195% 1015% 5260 109600 <1 21440 7057445

Average: 26.74% 1388% 1166% 1151% 892% 5038 110519 <1 19921 4814488

(CPLEX stops due to memory restrictions before converg- among the heuristic procedures tested in 71 of the 80 test
ing upon an optimal solution). problems in Table 5 (17 of the 20-node problems, 19 of
In Table 5 we present a summary of all the compu- the 40-node problems, 19 of the 60-node problems, and 16
tational results. The experimental data sets from Tables of the 80-node problems). ART with EW + EXACT takes
1 through 4 are included, as well as 30 additional prob- more time than the Lagrangian method for the 20-node
lems. A total of 80 experimental data sets are summarized problems, but substantially less time for the larger prob-
in Table 5—20 data sets for each of the 20, 40, 60, and lems. ART with EW + EXACT takes more time compared
80-node problems. Average percentages for each method to ART with EW + GCSSC and also EW + EXACT. The
(over the CPLEX Optimal for the 20-node problems and computational time of ART with EW + EXACT also grows
the Lagrangian lower bound solution for the 40, 60, and with the problem size, but not unreasonably.
80-node problems) are reported by problem size in Table 5.
EW + EXACT is a high-quality technique in terms of 7. SUMMARY AND CONCLUSIONS
both solution quality and computational time. We see that We have proposed a new formulation for the HFC-CTSD
EW + EXACT on average produces better solutions than problem (a capacitated tree-star concentrator design prob-
EW-CMST, ART with EW + GCSSC, and the Lagrangian lem). The HFC-CTSD allows a choice among a variety of
upper bound. The subtrees identified by the EW-CMST ONU options to connect demand nodes through subtrees
heuristic contain a very good set of potential ONU sites, with uniform capacities to the headend of a hybrid fiber
but the simplistic GCSSC heuristic is unable to effectively coaxial cable TV network. Given the complexity of the
select the optimal ONU sites. Even though ART with EW+ HFC-CTSD problem, we proposed a heuristic based on hier-
GCSSC substantially improves upon the EW + GCSSC archically decomposing the problem into well-known sub-
solutions (not shown in the tables), the ART technique is problems: the CMST and the CSSC. Every demand node is
unable to bring the GCSSC heuristic to optimality. a potential ONU site, but ONUs are only allowed on arcs
ART with EW + EXACT finds the best solutions, and connecting directly to the headend from the CMST subprob-
the improvements in the solutions are substantial. ART lem solution. This formulation can easily be accommodated
with EW + EXACT produces the minimum solution value to allow ONUs of various capacities to be connected in

Table 4. Computational results for 80-node problems.

Solution Values Lagrangian Lower Bound Gaps CPU Times

CPLEX Bounds LAG ART LAG ART ART ART

Upper Lower LB EW-CMST EW + GCSSC UB EW + EXACT EW + EXACT EW + GCSSC LAG EW + EXACT EW + EXACT CPLEX

1151 687 78339 3352% 1897% 1220% 1195% 978% 12050 297850 <1 40280 4656000
1300 684 74802 3181% 1818% 1337% 1109% 1069% 12280 188460 <1 40900 6018700
1073 728 773325 3642% 2078% 1237% 1328% 966% 12440 203340 <1 37790 3134500
984 735 787508 3663% 1632% 1670% 1225% 1136% 11620 365910 <1 41880 5765500
1016 693 734577 4035% 2334% 1190% 1912% 1599% 12040 256490 <1 37930 6413000
1199 749 791573 3972% 2115% 1243% 2140% 1218% 12510 296630 <1 40800 7456800
1074 728 802755 3317% 1635% 1523% 1211% 1162% 11770 296150 <1 39100 4519200
1357 668 788003 3579% 1726% 1650% 1535% 914% 12350 311040 <1 46660 4849300
1270 724 794633 3717% 1553% 1389% 1691% 1099% 12280 281090 <1 36090 2529000
1216 762 819893 3453% 1709% 1392% 1331% 1014% 12180 292070 <1 36880 12814900

Average: 3591% 1850% 1385% 1468% 1115% 12152 278903 <1 39831 5815690
 550 / Patterson and Rolland
Table 5. Summary of expanded computational results.
Percentage Over Best Known Lower Bound CPU Times

Problem ART LAG ART ART ART

Size EW-CMST EW + GCSSC UB EW + EXACT EW + EXACT EW + GCSSC LAG EW + EXACT EW + EXACT

20 624% 326% 243% 301% 091% 298 2663 <1 4741

40 1728% 1168% 1035% 1067% 637% 1730 25074 <1 10372
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

60 2719% 1469% 1276% 1254% 928% 5093 112453 <1 20501

80 3531% 1859% 1338% 1534% 1127% 12387 298854 <1 40509
Average: 2151% 1205% 973% 1039% 696% 4877 109761 <1 19031

a treelike fashion by reducing the number of arcs that are that such a procedure without ART can easily match per-
currently prohibited through constraint set (5) in the HFC- formance with Lagrangian relaxation in a fraction of the
CTSD problem formulation. computational time. And such a procedure with ART sig-
The best solution technique found in this research to nificantly outperforms Lagrangian relaxation and all other
solve this problem is the ART methodology, utilizing the heuristic methods tested in terms of solution quality.
Esau-Williams heuristic for solving the CMST subproblem. Fourth, we demonstrated the applicability of the ART
Both a greedy swapping heuristic and an exact solution solution method to the well-known Traveling Salesman
procedure for solving the CSSC subproblem were tested. Problem (see Appendix B for the TSP implementation
Computational results of this approach are very favorable. of ART). In Appendix B we showed that with a simple
The contributions of this research are as follows: First, exchange of the primary heuristic with very minimal modi-
we present a new formulation for the HFC-CTSD prob- fications to the algorithm, the ART methodology efficiently
lem. This problem is a realistic version of local access net- solves a large set of well-known TSP instances. For these
work design, and is motivated by the redesign of current problems, we achieved solutions with an average optimality
cable TV networks in response to providing Internet and gap of less than 0.25% (see Table B.1). Optimal solutions
telephony services over such (largely coaxial) networks. were found for 40% of the test problems. By demonstrating
The formulation and solution techniques presented in this the ease with which the ART methodology can be modi-
paper address all four subproblems identified in previous fied to effectively solve the TSP (in addition to solving the
research (Gouveia and Lopes 1997) as the major compo- HFC-CTSD problem), this research should serve as a cat-
nents of local access network design: concentrator quantity alyst for designing new memory-based constructive search
problem, concentrator location problem, terminal clustering heuristics.
problem, and terminal layout problem.
Second, we propose a method by which we hierarchi- APPENDIX
cally decompose the HFC-CTSD problem, and solve it
The appendix can be found at the Operations Research
using the ART memory-based heuristic solution technique.
Home Page
http://or.pubs.informs.org/pages/collect.html.
ART modifies the CMST subproblem solution. The modi-
fied solution to the CMST subproblem then determines the
potential ONU sites for the CSSC subproblem, which is
then solved. Because ART creates new solutions for the REFERENCES
CMST subproblem by essentially reconfiguring the sub-
Adams, R. 1997. Switched digital video on the rise. Telephony
trees at each iteration, new potential concentrators are pro- 232(17) 18–24.
posed as well as new (closest) connections from the new Altinkemer, K., B. Gavish. 1986. Parallel savings heuristics for the
subtree configurations to the potential concentrator sites. topological design of local access tree networks. Proc. IEEE
Thus, even though ART is directly modifying the CMST INFOCOM ’86. Fifth Annual Conference on Computers and
solution, it is implicitly modifying the CSSC solution as Communications Integration Design, Analysis, Management,
well by manipulating the input to the CSSC subproblem. Miami, FL. IEEE, New York, 130–139.
This paper contributes to the ART literature by clearly , . 1988. Heuristics with constant error guarantees for
demonstrating new and inventive ways to utilize the ART the design of tree networks. Management Sci. 34(3) 331–341.
Amberg, A., W. Domschke, S. Voss. 1996. Capacitated minimum
methodology in a hybrid manner with other solution tech-
spanning trees: Algorithms using intelligent search. Combi-
niques.
natorial Optimization: Theory and Practice 1(1) 9–39.
Third, we demonstrate the dynamic use of an exact Balakrishnan, A., T. Magnanti, R. T. Wong, 1995. A decomposi-
solution technique (AMPL and CPLEX) within a memory- tion algorithm for local access telecommunications network
based heuristic framework. This is implemented by dynam- expansion planning. Oper. Res. 43(1) 58–76.
ically rewriting the AMPL CSSC data files for each iter- Brightman, J. 1994. Hybrid fiber/coax: Front runner in the broad-
ation of the ART heuristic. We have clearly demonstrated band transmission race. Telephony 227(2) 42–50.
 Patterson and Rolland / 551
Brown, P. 1998. Size doesn’t matter. Broadcasting & Cable Hall, L. 1996. Experience with a cutting plane algorithm for the
128(19, May 4) 76–78. capacitated spanning tree problem. INFORMS J. Comput.
Cauley, L., R. Blumenstein. 1999. AT&T, Time Warner in cable- 8(3) 219–234.
TV accord. The Wall Street Journal (February 2) p. A3 Karnaugh, M. 1976. A new class of algorithms for multipoint
columns 1, 2, 3, and p. A6 column 2. network optimization. IEEE Trans. Commun. 24(5) 500–505.
Coleman, P. 1998. Cablevision re-ups in Boston. Broadcasting & Kershenbaum, A., R. Boorstyn, R. Oppenheim. 1980. Second-
Cable (May 18), 50. order greedy algorithms for centralized network design. IEEE
Esau, L. R., K. C. Williams. 1966. On teleprocessing systems Trans. Comm. Com-22(11) 1835–1838.
Downloaded from informs.org by [130.63.180.147] on 02 March 2015, at 06:59 . For personal use only, all rights reserved.

design, Part II—A method for approximating the optimal net- , W. Chou. 1974. A unified algorithm for designing mul-
work. IBM Systems J. 5(3) 142–147. tidrop teleprocessing networks. IEEE Trans. Comm. Com-
Fleurent, C., F. Glover. 1999. Improved constructive multistart 22(11) 1762–1772.
strategies for the quadratic assignment problem using adap- Lee, K., K. Park, S. Park. 1996. Design of capacitated networks
tive memory. INFORMS J. Comput. 11(2) 198–204. with tree configurations. Telecomm. Systems 6 1–19.
Gasman, L. 1997. High-speed access assessed. Telephony Levine, S. 1997. @Home heads to work. Telephony (March 3), 8.
(November 24), 233(21), 20–28. McGregor, P. V., D. Shen. 1977. Network design: An algorithm
Gavish, B. 1983. Formulations and algorithms for the capacitated for the access facility location problem. IEEE Trans. Comm.
minimal directed tree problem. J. Assoc. Comput. Machinery Com-25(1) 61–73.
30 118–132. Narasimhan, S. 1990. The concentrator location problem with vari-
. 1991. Topological design of telecommunications able coverage. Comput. Networks and ISDN Systems 19 1–10.
networks—Survey of local access network design methods. Papadimitriou, C. H. 1978. The complexity of the capacitated tree
Ann. Oper. Res. 33 17–71. problem. Networks 4 217–230.
Glover, F. 1965. A Multiphase-dual algorithm for the zero-one Patterson, R., E. Rolland, H. Pirkul. 1999. A memory adaptive
integer programming problem. Oper. Res. 13 879–919. reasoning technique for solving the capacitated minimum
. 1977. Heuristics for integer programming using surrogate spanning tree problem. J. Heuristics 5 159–180.
constraints. Decision Sci. 8 156–166. Rolland, E., R. Patterson, B. Dodin. 1998. A memory adaptive
. 1997. Tabu search and adaptive memory programming— reasoning technique for solving the audit scheduling prob-
Advances, applications and challenges. R. S. Barr, R. V. lem. Working paper No. WP1998-002, Center for Advanced
Helgason, J. L. Kennington, eds., Advances in Metaheuris- Information and Telecommunication Technology Applica-
tics, Optimization, and Stochastic Modeling Technologies. tions, School of Management, University of Texas at Dallas,
Kluwer Academic Publisher, Boston, MA. Dallas, TX.
Gouveia, L. 1993. A comparison of directed formulations for the , , H. Pirkul. 1999. Memory adaptive reasoning and
capacitated minimum spanning tree problem. Telecomm. Sys- greedy assignment techniques for the CMST. S. Voss,
tems 1 51–76. S. Martello, I. Osman, C. Roucairol, eds., Meta-Heuristics:
. 1995. A 2n constraint formulation for the capacitated min- Advances and Trends in Local Search Paradigms for
imum spanning tree problem. Oper. Res. 43 130–141. Optimization. Kluwer Academic Publishers, Norwell, MA,
, M. J. Lopes. 1997. Using generalized capacitated trees for 487–498.
designing the topology of local access networks. Telecomm. Sharaiha, Y. M., M. Gendreau, G. Laporte, I. H. Osman. 1997. A
Systems 7 315–337. tabu search algorithm for the capacitated shortest spanning
, P. Martins. 1995. An extended flow based formulation for tree problem. Networks 29 161–171.
the capacitated minimal spanning tree. Paper presented at the Soltys, J. R., M. J. Fischer, B. D. Roth. 1997. FTS2000 access
Third ORSA Telecomm. Conf., Boca Raton, FL. optimization. Fifth Internat. Conf. on Telecomm. Systems,
Gupta, R. 1996. Problems in communication networks design Modeling and Analysis, Nashville, TN.
and location planning: New solution procedures. Disserta- Vittore, V. 1997. Cable telephony rebounds. Telephony (Decem-
tion, Fisher College of Business, The Ohio State University, ber 15), 6.
Columbus, OH. Watson, S. 1997. Bandwidth booster. Telephony 233(17, Octo-
, H. Pirkul. 2000. Hybrid fiber co-axial CATV network ber 6) 24–34.
design with variable capacity optical network units. Euro. Wilder, T., S. Van Beaver. 1998. Mixed nuts: Automated provision-
J. Oper. Res. 123(1) 73–85. ing, HFC and IP. America’s Network (September 1), 30–34.