HiTi Graph Model of Topographical Road Maps in Navigation

Systems

Sungwon Jung Sakti Pramanik

Department of Computer Science Department of Computer Science

Michigan State University Michigan State University
East Lansing, MI 48824 East Lansing, MI 48824
Email: [email protected] Email: [email protected]

Abstract when larger maps are considered. Minimum cost route

In navigation systems, a prim.ary task is to compute computa.tion wit,11 this large amount of road map data
the minimum cost route from the current locatzon to requires a significant amount of computation time.
the destination. One of the major problems for naviga- Since navigation systems are real time systems, it is
tion systems is that a significant amount of computa- critical to compute a minimum cost route satisfying a
tion time is required to find a mznimum cost path when time constraint.
the topographical road map is large. Since navigation Previously suggested transitive closure or graph
systems are real time systems, it is critical that the traversal algorithms [l, 5, 6, 7, 9, 14, 15, 181 are
path be computed while satisfying a time constraint. not directly applicable to topographical-road-map
In this paper, we propose a new graph model named (source? destination, cost) for the computation of
HiTi (Hierarchical mulTi graph model), for eficiently a minimum cost path due to the very large volume of
computing an optim.al minimum cost path. Based on data they ha.ve to search. Thus, we need an efficient
HiTi graph model, we propose a new single pair mrn- database organization method for structuring the to-
imum cost path algorithm. We empirically show that pographical road map to speed up the computation of
our proposed algorithm performs far better than the a minimum cost path. In this regard, two different ap-
traditional A * algorithm. Further, we empirically an- proaches have been studied in the past. One approach
alyze our algorithm by varying both edge cost distribu- is to develop a database structure which gives a sub-
tion and hierarchical level number of HiTi graphs. optimal minimum cost path quickly. The other is to
develop the database structure which gives an optimal
1 Introduction shortest path.
In navigation systems, a primary function is to find For the suboptimal shortest path generation,
possible routes from the current location (e.g. where a Ishikawa et al, Shapiro et al, and Liu et al used road hi-
driver is currently positioned) to the destination (e.g. erarchies (i.e. Freeways, Highways, ., Side roads) to
where the driver wants to go) wit.11 a minimum ex- speed up minimum cost routes [8, 16, 171. They used
pected cost. For this purpose, they use a topograph- multiple levels of hierarchical details of road maps to
ical road map which is in the form of the following cut down the unnecessary search space. For the opti-
recursive relation : mal shortest path generation, Agrawal and Jagadish
topographical-road-map(source,destination,cost) recently studied a data organization technique which
where the cost attribute indicates, for example, a min- precomputes and stores some partial path informa-
imum expected time of travel from point source to tion [2]. They use the precomputed partial path in-
point destination. Another applicable cost can be the formation to prune the search space when computing
shortest distance between the two end points. a minimumcost path. While their approach speeds up
One of the major difficulties of naviga,tion systems the computation of a minimum cost path, it does not
is the size of the topographical road map data. It re- optimize processing minimum cost path queries based
quires about 2.4 Gbytes of storage to store a small on road map navigation.
100 mi x 100 mi map discretized at 100 feet inter- The rea,son is tha,t the shortest path obtained from
vals [a, 1 I]. Thus, the size of data involved is very large their approach still includes a.11the detail intermediate

76
1063-6382/96 $5.00 0 1996 IEEE

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 nodes between the source and destination, which nav- between the nodes in V. Then, regional boundaries
igators do not frequently need to know. This problem partition a road map G(V, E) into a set of Compo-
would be very serious if we need to compute a shortest nent RQad Maps (CROM) exclusively. Each CROM
path on a large topographical road map where its dis- can be viewed as a subgraph with its boundary nodes
cretized interval is fine-grained and the two end nodes defining the boundary of the CROM. Connectivity be-
on the path are far apart. To cope with this prob- tween CROMs is represented by the connectivity of
lem, we have developed a HiTi (Hierarchical mulTi) their boundary nodes. One CROM has direct connec-
graph model of very large recursive relations (e.g. to- tions with another, if boundary nodes of the former
pographical road maps). are directly connected to boundary nodes of the lat-
HiTi graph provides a novel approach to abstract ter. We call this kind of connectivity as between con-
and structure road maps in a hierarchical fashion. It nections of CROMs. For each CROM, we precompute
models after the mechanism people use to select a the cost for each pair of comlected boundary nodes. If
route on a road map. That is, when a person wants two boundary nodes of the same CROM alre connected
to find a minimum cost route from a current loca- by a pa.th solely contained within the CROM, we call
tion in one state to a destination point in another, this a within connection. By using the between and
hle/she usually reads road maps in the decreasing map within connections, we ca.n partit,ion the entire road
scale order. For example, a driver who wants to find map. Examples of between and within connections of
a minimum cost path from a current location in East CROMs are shown in Figure 1.1 where dotted lines
Lansing to a destination point in L.A., will consult the represent regional bounda.ries.
highway road map for the entire U.S.A. That person
will then find a more detailed highway path connec-
tion by reading road maps at a state level (i.e., Michi-
gsn, and California). Finally, city road maps of Ea.st
Lansing and L.A. will be consulted to find the very
detailed path information around the start and des-
tination points. By having road maps ma.inta.ined in
this hierarchical structure, HiTi graphs provide the
following three features:

It allows the computation of a minimumcost path

without having to search through all the detailed Figure 1.1: Examples of within and between connec-
maps of the intermediate st,ates and cities. tion of CROMs
It provides a basis for storing road maps at var- HiTi graph is a graph whose nodes are the bound-
ious levels of hierarchical abstractions. Thus, ary nodes of the CROMs and edges are the within and
a more controlled storage management suitable between connections of CROMs. Note that a CROM
for navigation systems (e.g. automobile naviga- can be defined to contain a set of CROM.s, thus cre-
tion system) with limited a.va,ila.ble storage can be ating a. mult,i level hierarchy. Thus, we first need to
built. determine a set of the lowest level CROM:s which ex-
It is able to support various types of database clusively partition an entire road map. We call these
queries on hierarchical abstractions of a road CROMs a,s level 1 CROMs. Then, we can recursively
map. construct a set of level k CROMs by grouping a set of
geographically adjacent level k-l CROMs where k 2 2.
HiTi graphs, therefore, provide a powerful framework These sets of level 1, L?, . . . , k CROMs form1 a complete
for implementing road map queries as well as for stor- balanced tree structure where the root node of the tree
ing very large topographical road maps. In the fol- is a whole road map.
lowing paragraphs, we will first describe HiTi graphs A HiTi graph is named the level k HiTa graph if its
informally followed by a more formal description of nodes a,nd edges represent boundary nodes, between
the model. and within connections of level 1, 2, . . . ,Ic CROMs.
Consider a topographical roa,d map viewed as a The Hi.Ti graph constructed for the United States ex-
directed graph G(V, E). Nodes in V correspond to ample a.bove is a level 4 HiTi gra.ph.
di,scretized grid points representing map objects in a The rest of the paper is organized as follows. In
road map. Edges in E correspond to the connections section 2, a formal framework and description of HiTi

77

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 ~_________ c___-____--_----. --___- ------\

graphs are discussed. In section 3, we propose and ,

I
I
I

empirically analyzed a new single pair shortest path I

I
I
I

algorit,hm that takes advantage of HiTi graphs. Fi- I

I
1
I
,
nally, section 4 gives concluding remarks. I
I
I
I
I

2 FIormal Framework and Description I

!
.________d .__________..__--a .------------’
I

of HiTi Graphs SG:

SG; SG;
Recursive relations have the generic form R (attl,
attz, atts) where attributes satisfy the following 3 con-
Figure 2.1: A digraph G and its level 1 subgraph SG:,
ditions:
SC; and SG$
1. ai.tl and attz are the key att.ributes of R and share
the same domain
Definition 2.2 Let { SGj (‘vj’,Ejl,J 1 j = 1, n1 } be
2. att1 and attz are the recursive join attributes and a set of level 1 subgraphs of graph G(V, E) with the
their corresponding values are related by some corresponding Nj. Then, for each SG! where 1 < i 5
transitive relationship III, B; = { (z,y+ z) / (x,y, 2) E P E’ A x E K1 }
and Wi’ = { (x, Y, fi(x:, Y)) I (x,Y) E (N; x N:) A
3. al!t3 describes the transitive relationship between
dtl and attz (x fz*) y in SGf ) A x # y }.

Based on the above generic form, a recursive rela- Each set Bf contains an edge (2, y, Z) if there is an
tion can be viewed as a directed gra.ph G(V, E). Each edge (x, y, Z) in P E’ where x is in SG,’ and y is in
node in V is represented by the value of attl. Each a different level 1 subgraph. In other words, set B!
edge in E is represented by values of (attl, attz, atts). contains connectivity information between level 1 sub-
Suppose tha.t a recursive relation G(V, E) is part,i- graph SG: and other level 1 subgraphs of G(V, E).
tionedl into a set of subgraphs (i.e. SC! (V>, E:), SC: Then, the following condition holds: PE1 = UyJ.lBj
(V:, I?;), ., SGf,, (Vi’, , EA1) ) such that,: From the above, it can be noted that edges in UTAIBj
exclusively partition graph G(V, E) into a set of
v,luv2~u.. .uvl n1 = v , und E~UE~U...UE~~ c E subgraphs SG: (V:, E:), SG: (V,‘,Ei), .., SGA1
(Vi’, J E,1,J
Vi1 17 Vj’ = 0 and Et n Ej’ = 0
Each set W: contains an edge (x, y, fZ(z, y)) if there
where l<i,jsnl and i#j exists a directed path from the boundary nodes 2 to y
of SG,l. Function fz (x, y) gives an aggregated cost
We name SG: (vi’, I?,!) a level 1 subgraph i and the (i.e. from node x to y) with respect to z within
connections between all level 1 subgraphs are captured SG!. In other words, each edge in W: represents
in P E’ = E-(E:UEiU...UEAl). Then, each (~,y,z) additional precomputed connectivity information of
E PE1 has the property tha.t 2 E vi’ and y E Vj’ boundary nodes within SC!. Thus,
where i # j. Note that (;u,y,~) corresponds t,o (aiLl,
att2, gattg) where 2 = attl, y = a,ttz, z = attg.

Definition 2.1 Let N: denotes a set of nodes in SC: As a.n example, Ta.ble 2.1 shows the corresponding
(Vi’, 17:) connected to/from nodes in other subgraphs values of N: , B! and Wil of the three subgraphs shown
SG; ~(v,‘,Ejl) w h erei#jandl<i,j<nl. Specif- in Figure 2.1. In the following example, fZ(2, y) gives
the cost of a minimumexpected time of travel between
ically, N/ = {z 1 (3, y, z) E PE1 A z E K’} U
node x and y.
{y 1 (:c,y, z) E P E’ A y E I/;:‘}. Set N,’ is called
leuel 1 boundary nodes of SC:.

For example, consider Figure 2.1 where a digraph G

and ilts subgraphs SG:, SG; and SG: are shown. Set,
Ni of subgraph SG; is { G. H, h4, 0 }. Each level
1 sub’graph is described and ident.ified by its bound- Table 2.1: N!, B! , and W: of SG,! in Figure 2.1
ary nodes, since they exclusively belong to one level 1
subgraph. Based on boundary nodes of SG;, defini-
tion :!.2 gives formal definitions of level 1 between and so far, we have introduced level 1 subgraphs and
withiin edge sets Bt and W: of SG,‘. their associated information such as level 1 Bound-

78

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 ary node sets, and between and within edge sets, In
general, based on these level 1 subgraphs and their as-
sociated information, we can recursively define level k
subgraphs SGC and their corresponding N,b, Bf, and
kVik for any k 2 2. Their details are discussed in the
following definitions and propositions.

Definition 2.3 Let \kk-l = { SGf-1 (ykA1, E,“-l) Figure 2.2: An example of level 3 subgraph tree ST
1 i’ = l,nk-1 } be the set of all level k-l subgraphs of
graph G(V, E) w h ere k 2 2. Then, a level k subgraph nodes, level k between and within edge sets are same
57;: (E/;:“, E;) as d ef; ne d a s a subgraph induced by all as that of level 1 boundary nodes, level 1 between and
nodes of level k-l subgraphs in 0 5 eke1 where I@/ 2 within edge sets. They are formally defined in defini-
2. tion 2.4.
After all level k subgraphs in \k” are defined, the fol- Definition 2.4 Let PE” = E-(E~uE~LJ...UE~,).
lowing three requirements must be satisfied:
Then, N! = { x 1 (z,y,z) E PE” A t E Vi” } U { y (
Each level k-l subgraph in QkP1 is a subgraph of (x:, Y, 2) E PE” A Y E K” 1, Bik = { (x, Y, ;!I I (x, Y, z)
some level k subgraph in 9”. E PEk A z E Kk } and WF = { (~,y,fi(~,y)) I
@C,Y) E (Ni” xN?, A ( x !‘(z,\y) y in SGk) A z # y }.
All level k subgraphs in Q” do not overlap among Function fi (CC,y) g ives an aggregated cost (i.e. from
themselves. That is, node x to y) with respect to z within SC!.
Vl”uV;u.. .uV,k, = V and EfuE;u.. AJE;, c E As ca.n be noticed from the a.bove definition, PE”
= Uj”klB’j”. Note tha.t Uj”i;lBj”-’ = U~~;‘B~-’ -
L$,”n vj” = 0 and EF n Ej” = 0
Uy!~~Bjk. In order to show how N/, Bt , and lVt are
where 1 <i,j<nk and i#j obtained, we use the digraph in Figure 2.3 as an ex-
ample.
All edges belonging to level k between edge sets
are removed from level k-l between edges sets. As
a result, (UjnL;lBJ-l n UTklBt) = 0.

By satisfying the preceding first two requirements,

all subgraphs in Ur$i\kj are related to each other in a
complete balanced tree structure where the root node
of the tree symbolizes G(V, E). This complete bal-
anced tree is named a subgraph tree (ST). The root
node of ST is considered as a level k+l subgraph and
it has all level k subgraphs as child nodes. Recursively,
eac:h node symbolizing a level k subgraph has a set of Figure 2.3: An example of a digraph G(V, E) parti-
level k-l subgraphs as child nodes. All leaf nodes of tioned according to ST in Figure 2.2
ST symbolizes level 1 subgraphs. This subgraph tree is
named level k+l ST if the root of ST symbolizes level From Figure 2.3, we can obtain boundary no’des, within
k+l subgraph (i.e. G(V, E)). The following Figure and between edges of level 1 and 2 subgraphs. These
2.2 shows an example of level 3 subgraph tree. This are shown in Table 2.2.
tree structure shows a level 3 subgraph SC; induced We have given basic definitions necessary to define
by the nodes in level 2 subgraphs (i.e. SGT, SC: and a BiTi graph. However, these definitions (i.e. N,!,
SG;). It also shows a level 2 subgraph SG; induced Bk, and WF) do not provide an efficient way of ob-
by the nodes in the two level 1 subgraphs SC: and taining t,heir corresponding values. A primary reason
SG;. for this inefficiency is that their definitions require ob-
For each level k subgraph SG! of ST where 1 5 i I: taining their values by using v;” and EC, which is com-
nk, there exist corresponding level k boundary nodes putationally expensive. By using the values of N*b-i,
(i.e. N:) and level k between and within edge sets (i.e. BF-‘, and W/-l for all level k - 1 subgraph.s in Qk-‘.
Bb and Wik). The semantics of the level k boundary we can compute the values of NF, B;k, and IV: in gk.

79

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 It is a simple bookkeeping problem to keep track of
nodes on a minimum cost path; hence we will focus
on an algorithm for simply finding a minimum cost of
the path. We first introduce a set of basic notations
which will be used in the rest of this paper.

Definition 3.1 Let ST represent a leuel Ic subgraph

tree. Assume set X consist of a set of subgraphs (i.e.
nodes) of ST. Then, S;(X) = { y 1 y is a level 1 sub-
graph which is an ancestor of each subgraph in X },
sA(x) = &,s;(x,, and SC(X) = { Y 1 Y is a di-
Table 2.2: Boundary nodes, between and within edges
rect child node of a subgruph in X }. Note that SC(X)
of level 1, 2, and 3 subgraphs in Figure 2.3 = X for all leaf node subgraphs SG: in X.

Definition 3.2 Assume set X consist of a set of sub-

The detail method for computing Nf, Bk, and PV/ graphs. Then, SE(X) and SW(X) consist of between
from Nk-l,
z Bf-r, and IV!-’ is given in [lo]. Based and within edge sets associated with all the subgraphs
on within and between edge sets, the formal definit,ion in X respectively. SN(X) consists of boundary node
of HiTi graph is given as below: sets associated with all subgraphs in X.

Definition 2.5 A HiTi graph is a directed labeled Definition 3.3 Let SGt and SC,! be two distinct
graph defined in terms of between and within edges level 1 subgraphs defined in a level I; subgraph tree ST.
associated with all nodes of subgraph tree ST. A Then, LU B~G(SG~, SGj) is the least leveled common
HiTi graph defined over level kt-1 subgraph tree ST ancestor subgraph of SGf and SG;.
is called level k HiTi graph. It is represented by
Hk(Pk,Ak) where Pk = { x 1 (CC,y, z) E Uf=:=,Uj”&,$ The examples of the definition 3.1, 3.2 and 3.3 are
1 U { y ( (2, y, z) E Ufz=,Ujn&lBj } and Ak = { illustrated through the level 4 subgraph tree shown
in Figure 3.1. Assume that X = { SC:, SG:, }.
(~,Y,P,W,fz(X,Y)l) I (X,Ytfz(X,Y)) E qPy for 1
Then, S;(X) = ( SG:, SG:, }, S:(X) = { SG:,
= l,k 1 u { (x,Y,[l,B,rl) I (X,Y,f) E q&B; for 1 SG; }, sA(x) = { SG;, SG;, SC;, SG;, SG;,,
Z l,k 1. SG;}, SC(X) = { SG;, SG;, }, S&S;(X)) = { SG:,,
Note that the symbols “1” and “lV”( ‘B”) in the def- SG;, SG:,, SG;,, SG:s }, sB(x) = { B:, B:, 1,
inition of an edge in A” represent the edge is level SW(X) = { Wi, W,ll }, SN(X) = { N,1, N,‘, }, and
1 within(resp. between) edge. An example of level 2 LUBsc:(SGi. SG:,) is SGZ.
HiTi graph corresponding to the digraph in Figure 2.3
is shown in Figure 2.4.

Figure 2.4: A level 2 HiTi graph corresponding t.o the Figure 3.1: An example of level 4 subgraph tree ST
digraph in Figure 2.3
3.1 Shortest Path Algorithm based on
HiTi Graph
3 Use of HiTi Graphs for Computing We now describe a Shortest Path Algorithm based
Minimum Cost Route on HiTi graph (SPAH). SPAH takes advantage of
In this section, we discuss the use of HiTi graphs for HiTi graphs to speed up the computation of a mini-
computing a minimum cost path (i.e. shortest path). mum cost path. Our algorithm explores at most E S

80

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 search space necessary for computing a shortest cost proposition 3.1 provides the theoretical basis of the
path. The size of ES is shown in the the following detailed ascending method given in SPAH.
theorem.
Proposition Xl Given a level k HiTi graph, assume
Theorem 3.1 Let ST represent a level k+l sub-
that 1 be the level number of LUB,sc(SG! :,SGjl) where
g:mph tree where level k HiTi graph H”(P”, Ak)
node START E Vi1 and DEST E Vj’. .Assume that
is defined. Assume that we want to compute a
level ,u to r edges ure incident on node x such that
minimum cost path from START node in SC:
x E (K’ U SN(SC(SA(SG~)))) and 0 5 h 5 r. In
to DEST node in SG; Then, the explored search
order to find the shortest path from node x to any node
space ES is at most the union of Ei, Ej and
in sj$~(sk~(s($)), a‘t as enough to follow only level p
Aik E A” where A’” = ~E(SC(SA({SG~,SG~~)))) U edges from node I where min(1 - 1, r) < p 5 T.
SW(SC(SA({SG:, s’+)))).
Proof: The proof is given in [lo]. 0
Proof: The proof is given in [lo]. 0
We take Figure 2.2, 2.3, and 2.4 to exemplify propo-
From theorem 3.1, it is easy to see that how HiTi
sition 3.1. Assume that we wa.nt to find the short-
graphs can significantly reduce the search space. That
est pa,th from node 5 in SG: to node 25 in SC;.
is, without using HiTi graph, ES would be UyAIEj
Since LUB~G(SG~, SG;) = 3, 1 = 3. Th.us, we have
where n1 represents a total number of level 1 edges
SN(S~(SG,$)) = N: = {19,20,25}. From Figure 2.3
sets. Furthermore, theorem 3.1 allows more controlled
and 2.4, we ca.n see that t.hree edges are incident on
storage management that is suitable for navigation
node 5. They are: one level 1 between edge (5,7,3),
systems (e.g. automobile navigation system) where
one level 2 within edge (5, 10,6), and one level 2 be-
available storage is limited. This is possible because
tween edge (5,20,7). Then, by the a,bove proposition,
we do not need to have all level 1 edge sets to compute
in order to find the shortest path from node 5 to any
a minimum cost path. Instead, we only need a part
node in { 19,20,25}, we do not need to follow the level
of the level k HiTi graph and two level 1 edge sets
1 between edge (5,7,3) from node 5. The formal de-
(i.e. one for START node and the other for DEST
scription of SPAH is given in Figure 3.2.
node). Note that the major storage overhead of a to-
pographical road map comes from the size of level 1

1
edge sets.
We take a level 4 subgraph tree in Figure 3.1 to give
an example of theorem 3.1. Assume that we wants to
find a minimum cost path from node START in SG!
to1node DEST in SC:,. Then, our search space ES
consists of E: U Ets together with the corresponding
between and within edge sets of the subgraphs marked
w:ith * in Figure 3.1.
Within the search space ES, SPAH traverses edges
from node START in SG: to node DEST in SGj’ to
find a minimum cost path. Its edge traversal consists
of two phases, the ascending and descending phases.
Tlhe ascending phase corresponds to the period of
edge traversal from node START to the bounda.ry
nodes in S~(Sy~(sGjl)) where 1 is a level number of
LlJB,,(SG;, SGj). S imilarly, the descending phase
corresponds to the period of edge traversal from the
nodes in SN(S~’ (SC:)) t o node DEST. During the
ascending and descending phases, SPAH traverses
edges in a non-decreasing and non-increasing edge 1
level order respectively. Note that each of the con-
sidered paths has its own thread of processing. Some Figure 3.2: SPAH: Finding a Minimum Cost Path
of the paths are still in the ascending phase, whereas from START to DEST
the others are in the descending phase. The following

81

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 3.2 Performance Analysis of SPAH SPAH in the following subsect,ions.
The algorithm SPAH can be classified as a vari-
ation of A* algorithm in that it uses the function 3.2.1 Comparison between SPAH and A* Al-
f(u, DEST) to estimate the cost of the shortest path gorithm
from node u to DEST. In the domain of road
maps, the function f(u, DEST) computes the Eu- To show the search space savings of SPAH over the
clidean distance between the node u and DEST. traditional A* algorithm (presented in [17]), we create
This is possible because the coordinates (i.e. lon- 5 level 3 HiTi graphs where 10 2 IN:] < 20 and where
gitude and latitude) of all nodes on a road map the edge cost is generated based on a uniform distri-
are assumed to be available. Assuming (UX, u.y) bution [100,120] with 5 different seeds. We represent
and (DEST.x, DEST.y) are the corresponding coor- IN:] as the total number of boundary nodes defined
dinates of the nodes u and DEST, f(u, DEST) com- on level 1 subgraph SG,’ . Next, we create 5 plain grid
graphs which are simply unions of PE1 and level 1
putes (DEST.z - u.z)~ + (DEST,y - u.Y)~. Since subgraphs used in the above 5 level 3 HiTi graphs. For
Euclidean distance is not an over-estimated cost be- each level 3 HiTi graph and plain grid graph created
tween node u and DEST, our a.lgorithm finds the op- above, we compute 20 different shortest paths ran-
timal shortest path. The detailed proof for t,he op- domly prefixed pa.irs of source and destination nodes.
timality of A* algorithm is found in 141. A* shortest Let H,, and A,, be the total number of edges visited
path algorithm are shown to be more efficient t,han the by SPAH and A* respectively. Note that throughout
breadth-first search single pair shortest path algorithm this paper, we multiply the estimation of f(u, DEST)
when the database can fit in main memory [ll, 131. by 100 to normalize the estimation with respect to the
For algorithm SPAH, this is the case since we showed edge cost. The number 100 is used for this normaliza-
in Theorem 3.1 that its explored search space is at tion because the Euc1idea.n distance between two ad-
most ES which is small enough to fit in main mem- jacent nodes is 1 and the edge cost between two nodes
ory. is at least 100. We compare SPAH and A* by observ-
For our empirical analysis of SPAH, we create ing the ratio An/H,. these values are then averaged
two dimensional grid graphs G(V, E) with 4 adjacent over the 5 different level 3 HiTi and plain grid graphs
nodes. Two dimensiona. grid graphs are considered with the same source a,nd destination nodes. They are
as typical examples of road maps [l 1. 171. In grid shown in Figure 3.4 where the numbers on CCaxis rep-
graph G, IV\ and [El are equal to 800 x 800 nodes and resent 20 ordered < source,destinntion > pairs with
4 x 800 x 799 directed edges. From G(V, E), we create path length increasing from 1 to 20.
a level 4 subgraph tree ST where each level 1 subgraph
SGf(K’,Et) has IV,‘] = 100x100, IEtl = 4x100~99. 35
The level 4 subgraph tree ST consists of 64 level 1, 16
level 2, 4 level 3, and 1 level 4 subgraphs, which a.re 30
shown in Figure 3.3. 25
20
An/H,
15
10
5
0
0 2 4 6 8 10 12 14 16 18 20
\!iijo 01in
._______’0::
_‘;[gweQj
.______. Q& the number of shortest paths

Figure 3.4: Performance comparison between A* and

___~~=yu=.;.I
.-.-.
;_.~_.I._
L’I.2r
~_
i ~I~~T=.-
.~.~.~.l’, SPAN

Figure 3.3: A 800 x 800 grid graph partitioned a.ccord- Figure 3.4 clearly shows how effectively SPAH
ing to the level 4 ST cuts down the search space over the traditional A*
algorithm. It is interesting to observe that the ra-
We use the above level 4 subgraph tree ST to gen- tio An/H, increases rapidly as the path lengths from
erate level 3 HiTi graph for t.he empirica. ana.lysis of source t,o dest,ination increase. This occurs because

82

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 the search space A* needs to explore grows exponen- 3.2.3 Effects of the Number of Hierarchical
tially whereas that of SPAH grows very slowly due Levels
to the hierarchical structure of HiTi graphs.
We examined the effects of different levels of HiTi
graphs on the performance of SPAH (i.e. MA and
3.2.2 Effects of Edge Cost Distribution MD). We constructed different levels of HiTi graph out
of the same set of level 1 subgraphs. For this analysis,
In this section, we studied the effects of edge cost we create 2 (i.e. 4 < IN:] < 8 and 10 5 IN;‘] 5 20) x
d.istributions on the performance of SPAH. For 5 (i.e. an edge cost distribution [100,200] >with 5 seeds)
this study, we generate 4 x 5 (i.e. 4 uniform dis- level 3 HiTi graphs. Similarly, we create 2 x 5 level 1
tributions with 5 seeds) level 3 HiTi graphs where and level 2 HiTi graphs. Then, we measure average
10 < IN,‘] < 20. Note that the 4 uniform distri- H, of A4A and MD the same way as we did in section
btutions [100,120], [100,200], [100,300], and [100,500] 3.2.2. They are shown in Table 3.1 and 3.2.
correspond to 20%, lOO%, 200%, and 400% variations
of edge costs respectively. We apply SPAH to each
level 3 HiTi graph by randomly creating 50 different
<: source, destination > pairs and then averaging the
c’ost. Let MA and MD symbolize SPAH when the
e:stimator f(u, DEST) g ives Euclidean distance and Table 3.1: Effects of levels of HiTi graphs on H, when
zlero (i.e. no optimization based on Euclidean dis- 10 5 INi’] 5 20
tance) respectively. Figure 3.5 shows the effect of
edge cost distributions on the performance of SPAH
in terms of H, (i.e. the total number of visited edges)
of MA and MD. Note that the values of H, are av-
eraged over 5 seeds.

72000 Table 3.2: Effects of levels of HiTi graphs on H, when

70000 - ’ y...;... . . . . ‘...!.-
41 IN;‘l<S
68000 -+* - -
’ MA U -
66000 - MD + - -
64000 - As we can see from ta.ble 3.1 and 3.2, higher level
62000 - >
H, HiTi gra.phs do not necessarily guarantele the better
60000 - performance when using SPAH than the lower level
HiTi gra.phs. It depends on the t,he entire isearch space
for SPAH. The entire search space for SPAH on a
level k HiTi graph, represented by T”, is formulated
a.s follows:
nk-1
+c“k,w:,+2- “k-2
,wfc-‘
c,+l-
0 50 100 150 200 250 300 350 400
Edge cost variation in % -rk= 2.IE1J “kc “k-1 Iw;-*l+
Figure 3.5: Effect of edge cost on the performance of
MA and MD

When f(u, DEST) g ives Euclidean distance esti-

mation, the performance of SPAH deteriorates as the Note that IE’( represent t~he total number of edges of
variation of edge cost is increased. The primary reason a level 1 subgraph and n/ represent t&he tota,l number
is that increasing the varia,tion degrades the quality of of level 1 subgraphs where 1 < I < k. lf ‘P < ‘P
Euclidean distance estimation t.o shortest path. This where p > q, then SPAH is likely to perform better
degradation of Euclidean dista.nce estimation seems to on the higher level p HiTi graph than on the lower
have more severe impact on the performance of MA level q HiTi graph. Otherwise, the higher level HiTi
fosr the first pa.rt of edge cost variation (i.e. between graph is not likely to provide a performance advan-
20% and 200%) than for the rest (i.e. between 200% tage over the lower level HiTi graph. Table 3.3 shows
and 400%). Unlike MA, MD gives a very stable per- the values of T’, Tz, and ‘I3 corresponding to level
formance with varying edge cost distribution. 1,2, and 3 HiTi gra.phs used in the tables 3.1 and 3.2.

83

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE
 -rl -r 2 -P
PI Y. Ioannnidis and R. Ramakirishnan, “Efficient Transitive
Closure Algorithms”, Proc. of the 14th VLDB Conference,
10 5 IN:1 5 20 99000 98737 98487 pp. 382-394, 1988.
451h’;158 82272 83364 83332
171 Y. Ioannnidis, R. Ramakirishnan, and L. Winger, “Transi-
Table 3.3: Effects of levels of HiTi graphs on Tk tive Closure Aieorithms Based on Graph Traversal”, ACM
Transactions oi Database Systems, Vol. 18, No.3, pp. 513-
576, September 1993.

Table 3.3 verifies our conjecture on the performance of

PI K. Ishikawa, M. Ogawa, S. Azume, and T. Ito, “Map Navi-
gation Software of the Electra Multivision of the ‘91 Toyota
SPAH for different levels of HiTi graphs. An inter- Soarer”, Int. Conj. on Vehicle Navigation and Information
esting thing to note from the tables 3.1 and 3.2 is that System(VhrIS IVHS), IEEE, (1991) 463-473.
the search space (i.e. H, ) does not vary significantly [9] B. Jiang, “A Suitable Algorithm for Computing Partial
going beyond level 1 HiTi graph. This is because our Transitive Closures in Databases”, Proc. IEEE 6th Int’1
experiments were done with the grid graphs having the Conj. Data engineering, pp. 264-271, 1990.
following property. That is, the difference between Ti [lo] S. Jung, “Recursive Query Processing in Large Databases”
and T’+l does not vary significantly as i increases. Ph.D. Thesis. Dept. of Computer Science, Michigan State
As a result, creating higher level HiTi graphs does University, East Lansing, MI 48824, U.S.A., 1995.
not contribute to the reduction in computation time. [ll] R. Kung, E. Hanson, Y. Ioannnidis, T. Sellis, L. Shapiro,
From this observation, we can conclude that, level i and M. St.onebraker, “Heuristic Search in Data Base Sys-
HiTi graph is good enough for road map applications tem”. Proc. 1st Int. Workshop Expert Database Systems,
pp. 96-107, Oct. 19S4.
where the difference between Ti and Ti+’ is small.
[12] B. Liu, S. Choo. S. Lok, S. Leong, S. Lee, F. Peon, and
4 Conclusions H. Tan, “Inlegrating Case-Based Reasoning, Knowledge-
In this paper, we developed a. new graph, called a Based Approach and Dijkstra Aigorit.hm for Route Find-
HiTi graph, to model very large topographical road ing”, Proc. Tenth Conj. Arti&ial Intelligence for Applica-
tions (CAIA ‘941, (1994) 149-155.
maps. HiTi graphs provides a powerful formal frame-
work for structuring topographical road map data in a [13] J. Pearl, Heuristics: Intelligent Search Strategies for Gom-
hierarchical fashion. We have empirically shown that puter Problem Solving, Addison Wesley, Reading, Mass.,
1984.
our proposed shortest path algorithm, based on HiTi
graphs, significantly reduces the search space for com- [14] G. Qadah, L. Henschen, and 3. Kim, “Efficient Algo-
puting the minimum cost path over a very large to- rithms for the Instantiated Transitive Closure Queries“
pographical road map. Our algorithm is also empiri- IEEE Transactions on Software Engineering, Vol. 17, No.
3, pp. 296-309, March 1991.
cally analyzed by va.rying edge cost dist,ributions and
the number of levels of HiTi graphs. The applica- [15] A. Rosenthal. S. Heiler, U. Dayal, and F. Manola, “Traver-
tions of HiTi graph structure go beyond the domain sal Recursion: A practical approach t,o supporting recursive
applications’ Proc. IEEE 3rd I&‘/ Conf. Data Engineering,
of topographical roa.d maps. It can be applied to any pp. 580-590, 1987.
very large recursive rela,tions where hierarchical ab-
stractions of data are useful. [16] J. Shapiro, J. Waxman, and D. Nir, “Level Graphs and
Approximate Shortest Path Algorithms”, Networks, Vol.
References 22, pp. 691-717, 1992.
[I] R. Agrawal, S. Dar, and H. Jagadish, “Direct Transitive
[17] S. Shekhar, A. Kohli, and M. Coyle, “Path Computa-
Closure Algorithms: Design and Performance Evaluation”, tion Algorithms for Advanced Traveler Information System
A C M Transactions on Database Systems, Vol. 15. No.3, pp, (ATIS)“, P~oc. IEEE 9th Int’l Conj. Data engineering, pp.
427-458, September 1990.
31-39, 1993.
[2] R. Agrawal and H. Jagadish, “Algorithms for Searching
Massive Graphs”, IEEE Transactions on Iinowiedge and [IS] I. Toroslu and G. Qadah “The Efficient Computation of
Data Engineering, Vol. 6, No.2, pp. 225-238, April 1994. Strong Part,ial
Transit.ive-Closures”, Proc. IEEE 9th Int’1
C’,nj. Data Enyinrrriny, pp. 530.537, 1993.
[3] W. Dijkstra, “A Note on Two Problems in Connection with
Graphs”, Numer. Math., \rol. 1, pp. 269.271, 1959

[4] D. Galperin, “On the optimality of A*“, Artificial Intelli-

gerace, Vol. 8, No. 1, pp. 69-76, 1977.

[5] I<. Hua, 3. Su, and C. Hua, “Efficient Evaluation of Traver-

sal Recursive Queries Using Connectivity Index”, Proc.
IEEE 9th Int’l Conj. Data Engineering, pp. 549-558, 1993.

84

Proceedings of the 12th International Conference on Data Engineering (ICDE '96)

1063-6382/96 $10.00 © 1996 IEEE