Aarhus Univ Uri Zwick Tel Aviv Univ RandomEdge is slower than RandomFacet on abstract cubes Hebrew University April 11 2016 Maximize a ID: 806532
Download The PPT/PDF document "Thomas Dueholm Hansen" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Thomas Dueholm Hansen – Aarhus Univ.Uri Zwick – Tel Aviv Univ.
Random-Edge is slower than Random-Facet on abstract cubes
Hebrew University
April 11, 2016
Slide2Maximize a linear objective function subject to a set of linear equalities and inequalities
Linear ProgrammingFind the highest point in a polytope/polyhedron
Slide3Move
up, from vertex to vertex, along edges, until reaching the top.The Simplex Algorithm[Dantzig (1947)]
Slide4Largest improvementLargest slopeDantzig’s rule – Largest modified costBland’s rule – avoids cycling
Lexicographic rule – also avoids cyclingZadeh’s rule – “least recently entered”Deterministic pivoting rulesAll known to require an exponential number of steps, in the worst-case
[Klee-Minty (1972)] ,
… ,
[
Amenta
-Ziegler
(1996
)]
[
Friedmann
(2011)]
Slide5Random-EdgeChoose a random improving edgeRandomized pivoting rules
Random-Facet is sub-exponential! Random-Facet[Kalai (1992)] [Matoušek-Sharir-Welzl (1992)]To be described shortlyIs
Random-Edge
sub-exponential
???
Is
Random-Edge
faster than
Random-Facet
?
Slide6Random-FacetAssume
[
Kalai
(1992)]
[
MSW (1992)]
= number of inequalities
= number of variables = dimension
[Hansen-Z (2015)]
(following
[
Kalai
(1992)]
)
Fastest known
(randomized)
pivoting rule
Slide7Random-FacetAssume
[
Kalai
(1992)]
[
MSW (1992)]
= number of inequalities
= number of variables = dimension
[Hansen-Z (2015)]
(following
[
Kalai
(1992)]
)
Fastest known
randomized
pivoting rule
Slide8The diameter of a d-dimensional, n-faceted polytope is at most
n−dRefuted by [Santos (2010)].Diameter is still believed to be polynomial.
Quasi-polynomial
upper bound
[
Kalai-Kleitman
(1992)] ([Todd (2014)])
Hirsch
Conjecture
(1957)
Slide9Abstract objective functions (AOFs)
Every
face
has a unique sink
Acyclic Unique Sink Orientations (AUSOs)
Slide10Klee-Minty cubes (1972)Figure taken from a paper by Gärtner-Henk-Ziegler
Slide11AUSOs of n-cubesThe diameter
is exactly n
Williamon
-Hoke
(1988)
Kalai
(1998)
Szabó
,
Welzl
(2001)
Gärtner
(2002)
AUSOs
No diameter issue!
2
n
facets
2
n
vertices
Every
subcube
has a unique
sink
Slide12Klee-Minty
cubes (1972)
000
100
001
010
110
…
has a
Hamiltonian path
(Gray code)
.
Random-Edge
on
takes
,
w.h.p
.
[
Balogh-Pemantle
(2007)]
011
111
101
Slide13Turn-based
2-Player
0-Sum
Stochastic Games
[Shapley ’53] [Gillette ’57] … [Condon ’92]
Both players have optimal
positional
strategies.
Can optimal strategies be found in
polynomial
time?
Limiting average
cost
Discounted
cost
Total cost
Game (two actions per state)
AUSO (of a cube)
Random-Facet
is the fastest known algorithm.
Slide14Parity Games (PGs) A simple example214
132
EVEN
wins if largest priority
seen
infinitely often
is
even
Priorities
Slide15Random-Facet on the
-cube[Ludwig (1995)] [Gärtner (2002)] Start at some vertex of the -cube
.
Split the
-cube into
two
-cubes
along a
random
coordinate
.
Recursively find the sink
of the
-cube
containing
.
If
is not the global sink,
move to
.
Recursively find the sink
of the
-cube
containing
,
starting from
.
…
-
th
coordinate
Exponential?
Sub-exponential!
The starting point
of the
second recursive call contains
valuable information
.
All correct !
Would never be switched !There is a hidden order of the indices under which the sink returned by the first recursive call correctly fixes the first i bits.
(After reordering the coordinates according
to the
hidden
order.)
Random-Facet
on the
-cube
[Ludwig (1995)] [
Gärtner
(2002)]
The algorithm does not know the
hidden
order.
But, choosing a coordinate that was
fixed
has no effect.
Thus, the second recursive call is effectively on an
-cube.
All correct !
Would never be switched !
Random-Facet
on the
-cube
[Ludwig (1995)] [
Gärtner
(2002)]
(After reordering the coordinates according
to the
hidden
order.)
Slide18Primal Random-Facet Non-recursive versionChoose a random permutation of the facets
containing the current vertex v.Find the first facet that is beneficial to leave and move to a new vertex contained in a new facet .
Choose a new random ordering of
. Keep the ordering
of
. Repeat.
Upper boundLower boundAlgorithmRANDOM EDGERANDOM FACET
Randomized Pivoting Rules
[
Kalai
’92]
[
Matousek-Sharir-Welzl
’92]
[
Friedmann
-Hansen-Z ’11]
Slide20Lower bounds for Random-Edge
for AUSOs [Matoušek-Szabó (2006)]
for
LP
s
[
Friedmann
-Hansen-Z (2011)]
for
AUSOs
[here]
The new lower bound is a
simplification
of the lower
bound of
Matoušek
and
Szabó
obtained by replacing the
Klee-Minty
cube, used as a building block, by a
path AUSO
.
Two main building blocks:
Product
of AUSOs
,
Hypersink
replacement
Main part of the technical analysis:
Random Walk with reshuffles
(on a path AUSO)
Slide21A
product
(
blowup
) construction
Slide taken from a presentation by
Tibor
Szabó
.
Slide22A
product
(
blowup
) construction
Adaptation of a slide by
Tibor
Szabó
.
(
y
))
Hypersink
replacement
Slide
by
Tibor
Szabó
.
Slide24A
A
A
A
rand A
A simpler construction
Slide25R
andomized
product
+
hypersink
replacement
obtained from
by
randomly
permuting
the indices.
Adaptation of a slide by
Tibor
Szabó
.
The copy of
corresponding
to
is a
hypersink
.
Replace it by a
random
translation
of
.
11
Slide26Random-Edge
on a random product[Matoušek-Szabó (2006)]
Each step in
is either
a step in
, i.e.,
, or a step in
, i.e.,
A step
in
, does not change
,
but changes the cube
is in from
to
.
The induced process on
, until
, is exactly
Random-Edge
on
.
T
he induced process on
, is a
Random Walk with Reshuffles
on
.
As
is a random permutation of
,
this corresponds to
randomly
permuting
, a
reshuffle
.
Once
, the orientation on
changes from
to
.
As
is a
random
translation of
,
is a completely
random
in
.
If
reaches
before
reaches
,
we get
two
Random-Edge
walks on
.
The probability of reshuffle in each step of Random-Reshuffle on depends on
. (To be discussed later.) Lemma: If the probability that Random-Edge on makes less than steps is at most
, and the probability that
Random-Reshuffle
on
makes less than
steps is at most
,
both starting from a random vertex, then the probability that
Random-Edge,
starting at a
random
vertex
,
makes
less than
steps on
is at most
.
Random-Edge
on a
random
product
[
Matoušek-
Szabó
(2006
)]
Slide28Lower bound Construction
[Matoušek-Szabó (2006)]
Let
be an
-AUSO. Compute randomized
powers
of
.
,
, where
, for some
.
If
Random-Reshuffle
on
is
slow enough
, then
Random-Edge
on
, which is an
-AUSO,
makes at least
steps, with high probability.
In the last step, we need
Random-Reshuffle
on
to make at least
steps, with high probability.
Matoušek-Szabó
take
to be the
Klee-Minty
cube.
Random-Reshuffle
on Klee-Minty
cubes is not slow enough…
Further complications. Lower bound becomes
.
Our modification:
Take
to be the
path AUSO
.
A
path AUSO
A
shortest paths
problem gives rise to an AUSO.
A vertex of the AUSO corresponds to a choice of
an outgoing edge from each vertex of the graph.
In this particular case, the
-
th
bit wants to
stay or become a 1
iff
all preceding bits are 1.
Random-Reshuffle
on a path AUSO
1111
0
0
1
0
1
in
– number of leading 1s
– number of
non-leading
1s
–
outdegree
of
in
in
Outdegree =
With probability
,
reshuffle
.
Otherwise, with probability
, change the first
0
to 1,
and probability
, change a random non-leading
1
to 0.
The
reshuffle
probability
chosen by an
adversary
.
Random-Reshuffle
on a path AUSO1111
0
0
1
0
1
– number of leading 1s
– number of
non-leading
1s
in
All states of
type
are obtained with equal probabilities.
type
=
weight
=
A
reshuffle
on a state of
weight
creates a state
of
type
, where
,
with probability
The probability of increasing or decreasing
depends only on
.
We thus obtain induced random walks on
types
, and then on
weights
.
Slide32-cube
Types
0
Constant
drift
to the left
Exponential time to reach
.
Weights
Induced
random walk on the lineSlight complication:
From
, the probability of
+1
is greater than
½
.
Solution: Consider
two
steps together.
Basic step: Zero or more
reshuffles
, followed by a step in
.
Several
reshuffles
in a row are equivalent to a single
reshuffle.
The worst type of
weight
is
.
Induced
random walk on the lineProbability of weight increase in one step from
is the reshuffle probability
,
Induced
random walk on the lineProbability of weight decreasing by 2:
Concluding remarks and open problemsWe presented an
lower bound for Random-Edge. The bound can be slightly improved to
,
thus
Random-Edge
is
slower
than
Random-Facet
.
This is the best that can be obtained
using
hierarchical
constructions.
Best upper bound known for
Random-Edge
is
.
Is
Random-Edge
sub-exponential?
Is there an algorithm that beats
Random-Facet
?
Is there a
polynomial
time algorithm?
(The algorithm does not necessarily
have to follow a path. It may
jump
around.)
Slide37( Stochastic Shortest Paths (SSPs) )
Minimize the
expected
cost
of getting to the target
Markov Decision Processes (MDPs)
Slide38Random-Facet on the -cube[Kalai (1992)]
[Matoušek-Sharir-Welzl (1992)][Gärtner (2002)] Let
be the current vertex.
Choose a
random
index
.
Recursively
find the sink
of the
subcube
.
If
, then
is the sink of the whole cube.
Otherwise, move from
to
and
run the algorithm recursively from
.
Analysis of Random-Facet on the -cube
Let be an upper bound on the expected number of steps performed on any -AUSO from any starting point.
Analysis of Random-Facet on the -cube
Let such that
implies
Let
be the current vertex.
Let
be the optimal vertex
in the
subcube
.
Renumber the indices such that
.
The algorithm chooses a random index
and performs a first recursive call that computes
.
Let
.
Analysis of Random-Facet on the -cube
Let such that
implies
Let
be the current vertex.
Let
be the optimal vertex
in the
subcube
.
Renumber the indices such that
.
The algorithm chooses a random index
and performs a first recursive call that computes
.
Let
.
If
is not the global sink, the algorithm moves to
.
.