Optimistic Shortest Paths On Uncertain Terrains: 16th Canadian Conference On Computational Geometry, 2004

16th Canadian Conference on Computational Geometry, 2004
Optimistic Shortest Paths on Uncertain Terrains
William Evans
Department of Computer Science
University of British Columbia
Chris Gray
Department of Computer Science
University of British Columbia
X
@W
X
cW
XY
UeW
XY
UeW
T C w&c vC trQq C T C p i

u s P
f
hg@
9
Given an uncertain terrain and two points and in the plane, we want to nd a path from to that has the shortest
length among all paths from to where the path length is
measured on the consistent terrain that minimizes the length.
Cast as a decision problem, this becomes, Is there a path
from to on some consistent terrain of length at most ?
We call this the optimistic shortest path since we are measuring a path on the consistent terrain that minimizes its length.
X
cW
C
p (C F
h@
9
The decision problem is NP-hard. It is not clearly in NP

due to the algebraic complexity possible in shortest paths on
terrains [2].
To show that the problem is NP-hard, we will use a reduction similar to Canny and Reifs reduction of 3SAT to the
problem of nding shortest paths in 3D with obstacles [3].
In fact, we construct gadgets (splitters, shufers, etc.) that
perform the same functions as the corresponding Canny-Reif
gadgets, but are part of an uncertain terrain rather than being
3D objects.
Let be a 4SAT formula with variables and clauses.
Our construction uses
path splitters to create
path
classes, each corresponding to a truth assignment for the
variables. Then it directs the paths through clause boxes,
each of which has 4 literal lters. The purpose of each clause
($ "
)'&%#!
7 5 31
86420
T R P
VU C SQI H)( GFDB
( EC
A@
9
X
@W
XY
`GW
cmg@cs.ubc.ca. Supported by GEOIDE Networks of Centres of Excellence

will@cs.ubc.ca. Supported in part by a Natural Sciences and Engineering Research Council of Canada Research Grant
68

A terrain is a two-dimensional surface in three-dimensional
space obtained by triangulating the projection into the plane of a set of vertices in 3D (no two of which share
the same
coordinates) and using linear interpolation to
obtain the surface within each triangle.
Shortest path problems are a well-studied class of problems

in theoretical computer science. One particularly applicable
type of shortest path problem is to nd the geodesic shortest
path on a terrain. This type of algorithm nds the shortest
path between two points that stays on the surface of a terrain. The most popular methods for nding such a shortest
path involve a variant of Dijkstras algorithm and run in time
approximately
in the size of the terrain [5, 4].
These algorithms for calculating shortest paths on a terrain
require a precise input; any errors in measuring the terrain
translate into errors in the output of the algorithms. What
appears to be a shortest path according to the given input
may turn out to be longer than an alternate path in reality.
Uncertain terrains are a new model for acknowledging and
dealing with these errors.
In this paper, we consider one version of the shortest path
problem on uncertain terrains: the optimistic shortest path.
Essentially, we would like to nd the path whose length is
smallest over all paths and over all possible real terrains.
This seems to be a slight generalization of the traditional
geodesic shortest path problem. We show that it is, in fact,
more akin to the problem of nding the shortest path in three
dimensions that avoids polyhedral obstacles. This problem
was shown to be NP-hard by Canny and Reif [3] in 1986. It
is from their proof that our work is derived.
X
cW
We show that nding the optimistic shortest path on an uncertain terrain is NP-hard using a reduction similar to Canny
and Reifs reduction of 3SAT to 3D Euclidean shortest path.
An uncertain terrain is a set of vertices in the -plane,

each of which has an associated interval, and a triangulation of the ( -projection of the) vertices. Any terrain whose
projection into the -plane produces the same triangulation
and whose vertices have a value within the corresponding
interval is consistent with the uncertain terrain. A vertex in
an uncertain terrain with a interval that contains more than
a single value is called an uncertain vertex.
A path is a sequence of
coordinates. A path can be
placed on a terrain to create a sequence of 3D points by associating with each
coordinate the corresponding value
from the terrain. The length of a path on a terrain is the sum
of the Euclidean distances between adjacent 3D points after
placing the path on the terrain.
The shortest route from to is a sequence of terrain
edges traversed by a shortest path from to . A path class
is an equivalence class of points that all have the same
shortest route(s) from .
CCCG 2004, Montreal, Quebec, August 911, 2004
the uncertain vertex , cross either edge

or , and cross
output edge . The height of the input vertex closely determines the point at which the path crosses the output edge.
See Figure 3.
Note that splitters are like two rotators that have been
joined together.
q
cl p
p
8o
box is to lengthen each path class that does not have a setting
that satises the clause. After the clause boxes, the paths are
directed through another cascade of path splitters, this time
reversed. After the last one, there is only one path class left.
We place in this path class, and by measuring the shortest
path length from to , we can determine if any one of the
shortest routes have not been lengthened. If so, then we
know that our formula is satisable.
q
jo
7eY Y e
8t6p1 p
Input Vertex
Front View
Reversed
splitters
7 Y Y
`6&1
7 Y Y
o`6&1
7k e Y i Y Y e
4j& j&nFU1
Figure 1: Overview of the construction
We will now look at the gadgets in greater detail.
Output Edge
Back View
Clause boxes
Splitters

A#g
9
" s
C @ C R g
Figure 3: The splitter, showing the doubling of the number

of path classes
A basic operation that our gadgets must perform is to rotate

the orientation of the group of path classes by
. This is
important as a building block of other gadgets and is performed by a rotator, which is like a square of paper folded
along its diagonal. See Figure 2 for an example of a rotator
of width 1 located at
with a thickness parameter
that is essentially .
h 7tt& j68FU1
keY iY Ye
7p`Y@`Yp1
e
h
7 Y Y
p`@`@d1
e
r
C n)H(w Gi tp u
m s
h#g
9
hg#g
9
The purpose of a path splitter, as its name suggests, is to take

path classes and split them into
path classes.
To create the splitter, we construct the uncertain terrain
so that any shortest path from to must enter at or near
C @s i tp u
s$ s
g 8t6p1
7eY Y e
l
Since the shortest paths leave the rotator in a vertical line,

we have changed the orientation of the paths from horizontal
to vertical.
7 Y Y
`@`&1
Figure 2: Front and back view of a rotator showing shortest

paths across it as dotted lines
Back View
Front View

GF
Input Edge
Output Vertex
Given these gadgets, we are ready to describe the rst part

of the construction. In front of the initial point , we construct a wall that is high enough that any shortest path from
to must go around the wall to the left or right. This creates
two shortest path classes. Following this wall is a sequence
of connected rotators and splitters, joined at their common
vertex and increasing in size by a factor of 2 each time.
When we join the rotators and the splitters at the vertex ,
we leave a gap between vertex on the rotator and vertex
on the splitter. It is possible to ll this gap with a very high
wall (or a very low trench), but for now we will leave gaps
unlled and assume that no shortest path can afford to pass
over a gap.
The output edge of the last splitter may be reached from
via
shortest routes. Each of these shortest routes intersects the nal output edge at a different location along the
output edge.
We number the resulting path classes along this edge from
to
and associate them with the
possible truth
assignments to the variables of the formula .
The rest of the construction lengthens the paths in those
path classes that correspond to unsatisfying truth assignments.
The path shufer is designed to output a perfect shufe of the

path classes input to it. A perfect shufe takes the sequence
01234567 and permutes it to 04152637, for example. Notice
that this is the same as doing a circular bitwise right shift of
69
3

16th Canadian Conference on Computational Geometry, 2004
corresponding truth assignment has a 1 in the third position.

To do this, we take the paths through three shufers. At this
point we have all the paths with a 1 in their third bit in the
left half of the output edge and all the paths with a 0 in the
right half. Then we put a barrier on the side with the ones to
shufers to get the
stretch all of them. Finally, we put
paths back in their original positions.
r 3
$
p C &)w
H#g
9
A clause box consists of two cascaded splitter / rotator combinations that have the effect of splitting each path into four
copies. These are followed by a literal lter for each literal
in the clause put in parallel so that each of the copies of a
given path goes through exactly one literal lter. Finally, we
add two reversed splitter / rotator combinations.
A short path exists through this gadget if and only if
belongs to a path class that corresponds to a truth assignment
that satises at least one literal in the clause.
each number in the sequence. Thus, if 01234567 is the order

of path classes on the input edge of a shufer, 041526374 is
the order on the output vertex.
The implementation of the shufer is similar to the implementation of the splitter. The shufer can also be seen as
being composed of two rotators put together. Assume that
the paths are apart. We will split the group of paths into
two parts, so that half go to one side and half go to the other.
When the group has been split, one of the halves will be
raised a slight amount (
) and the other half will be
lowered the same amount. The paths will then be rotated so
that they interleave vertically.
In order to ensure that paths do not cross into the wrong
group, we put a gap between the two groups of paths. This
is shown in Figure 4 as the gray triangle.
The shortest paths leave the rotators through point or
point in Figure 4. They then travel along edge
or to
point . Point is the input vertex to a reversed rotator (not
shown) that rotates the group of paths back to horizontal.
zvu w v
8Fyx'e
gh
f
g
Qf
Splitters
7
3
6
2
5
1
4
0
Literal filters
7 } w {5 1
Y Y
f
h
7F8FF8tc8Fu r F8e g r j6'SnFU1
kzvu w veY zv v iY w Ye
Figure 5: The clause box
7YeY
pnFj1
7 Y Y
p`@`&1
We will now show that each gadget increases the length of

the shortest path traveling through it by a specied amount.
We will also show that a path that travels through a gadget in
a non-shortest manner will be detectable at the end.
s
tp u C !&s i ) T Q !F C R
( ( ( m P C " s
Lemma 1 For a rotator of width , there exists a path from

a point in
on the input edge to the output vertex of length less than or equal to
only if the path exits
through the output vertex in the range
where
.
7
4k w { `Y qr W ijY68Y41
W

w
82
7 Y Y WY i
pn&`k { U#W `1
7Y W
pn&nY 41
d
cW
7 WY Y
U6841
7 Y W
``Y 41
7 Y Y
pn&`@d1
g
U`

Proof. In order to nd the locally optimal path across a rotator, rst we consider the unfolding of the rotator when the
output vertex is at its lowest possible elevation and
.
The shortest path from the point
is clearly the
straight line across to
. If we refold the rotator,
this point goes to
. Since points can only exit via
the output vertex, the output vertex must move to that point.
7 Y eY
`pt&1
7 YeYe
`FjFU1
C )Hwc! C (
( w
H#g
9
The literal lter is a meta-gadget. It consists of path shufers and one barrier. We use the literal lter in order to
stretch out paths that have a truth value that is opposite the
truth value asked for in a clause. For example, if we want to
express , we stretch all paths in those path classes whose
7 ~r |r &n@d1
} { 5 Y Y
0 1 2 3
4 5 6 7
Input Edge
Figure 4: The shufer viewed from above. The circled regions contain rotators, shown projected into the
plane
above the shufer at the correct relative heights. The vertex marked is the output vertex.
#W
70
Splitters
CCCG 2004, Montreal, Quebec, August 911, 2004
representing one truth assignment to a path class that represents another by making the initial wall very wide. Since the
width of a path class only grows by
for each gadget, the
path classes are never close enough that changing from one
to the other is a locally optimal motion. Thus we have
Theorem 3 The optimistic shortest path problem is NPif

hard. That is, we can nd a path that has length
and only if we can nd a setting of variables that satises the
formula in 4-CNF form.
5 w 5
C C C P

C
[1] Tetsuo Asano, David Kirkpatrick, and Chee K. Yap. d optimal motion for a rod. In Proceedings of the Twelfth
Annual Symposium On Computational Geometry (ISG
96), pages 252263, New York, May 1996. ACM Press.
Figure 6: The unfolding has three xed vertices and the

fourth inside the shaded strip.
. Whereas the unfoldNow consider the case when

ing of the rotator was a square with edge length , it is now
a quadrilateral. In the unfolding (see Figure 6), the triangle
forms a right-angled triangle with leg lengths and
. The fourth vertex falls within a strip of width
above the
rectangle. We can construct an
approximation of the quadrilateral that is a
square with
a
strip added to the top to represent the uncertain location
of ; the output vertex must be inside this strip.
If we consider a path of length
across this quadrilateral from an arbitrary input point, the width of the region
that the output point could potentially be in is
.
, this region has a width that is less than
Therefore, if
. Thus, if the input point starts in range
, it will exit
in the range
.
[2] Chanderjit Bajaj. The algebraic degree of geometric optimization problems. Discrete Comput. Geom.,
3(2):177191, 1988.
[3] John Canny and John Reif. New lower bound techniques
for robot motion planning problems. In Ashok K. Chandra, editor, Proceedings of the 28th Annual Symposium
on Foundations of Computer Science, pages 4960, Los
Angeles, CA, October 1987. IEEE Computer Society
Press.
[4] Sanjiv Kapoor. Efcient computation of geodesic shortest paths. In ACM, editor, Proceedings of the thirty-rst
annual ACM Symposium on Theory of Computing: Atlanta, Georgia, May 14, 1999, pages 770779, New
York, NY, USA, 1999. ACM Press.
[5] Joseph S. B. Mitchell, David M. Mount, and Christos H.
Papadimitriou. The discrete geodesic problem. SIAM J.
Comput., 16(4):647668, 1987.
5 w 5
7 w 1
4
l
8
w
8
Un
k { `Y iW
W
5 w 5
ghf
k w { `Y r iW
W
Lemma 2 For a shufer of width

, there exists a
path from a point in
on the input edge to
the output vertex of length less than or equal to
only if the output vertex is either
or
in
in
,
depending whether
or not.
'
w
'
7 Y Y W i
p`@`k { UY W `1
W
e { W
7
4k w'vFwFw { WUY |r 8FuwecW nd1
z u v
zv v w iY w Y
7k e8vFu r v r { UY r z8Fu r r cW V|nd1
w z
W
v v iY w Y
7 z v u w v 1
8w 5 'FyqF4w 5
Proof. The locally optimal motion for a path entering a shufer is for the path to travel in a straight line to the rotator
across from it. Once it is at the rotator, Lemma 1 applies.
From Lemma 1 and Lemma 2, we can deduce the maximum distance possible for any locally optimal path that corresponds to a truth assignment that satises the formula .
This distance will always be less than the minimum distance
of any path that is either not locally optimal or that corresponds to a truth assignment that does not satisfy the formula . We ensure that no path crosses from a path class
71

Optimistic Shortest Paths On Uncertain Terrains: 16th Canadian Conference On Computational Geometry, 2004

Uploaded by

Copyright:

Available Formats

Optimistic Shortest Paths On Uncertain Terrains: 16th Canadian Conference On Computational Geometry, 2004

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Optimistic Shortest Paths On Uncertain Terrains: 16th Canadian Conference On Computational Geometry, 2004

Uploaded by

Copyright:

Available Formats

16th Canadian Conference on Computational Geometry, 2004

Optimistic Shortest Paths on Uncertain Terrains

The decision problem is NP-hard. It is not clearly in NP

cmg@cs.ubc.ca. Supported by GEOIDE Networks of Centres of Excellence

Shortest path problems are a well-studied class of problems

An uncertain terrain is a set of vertices in the -plane,

CCCG 2004, Montreal, Quebec, August 911, 2004

the uncertain vertex , cross either edge

Figure 1: Overview of the construction

We will now look at the gadgets in greater detail.

Figure 3: The splitter, showing the doubling of the number

A basic operation that our gadgets must perform is to rotate

The purpose of a path splitter, as its name suggests, is to take

Since the shortest paths leave the rotator in a vertical line,

Figure 2: Front and back view of a rotator showing shortest

Given these gadgets, we are ready to describe the rst part

The path shufer is designed to output a perfect shufe of the

16th Canadian Conference on Computational Geometry, 2004

corresponding truth assignment has a 1 in the third position.

each number in the sequence. Thus, if 01234567 is the order

Figure 5: The clause box

We will now show that each gadget increases the length of

Lemma 1 For a rotator of width , there exists a path from

CCCG 2004, Montreal, Quebec, August 911, 2004

Theorem 3 The optimistic shortest path problem is NPif

Figure 6: The unfolding has three xed vertices and the

. Whereas the unfoldNow consider the case when

Lemma 2 For a shufer of width

You might also like