algorithms for submodular objectives: continuous extensions & dependent randomized rounding

65
Algorithms for submodular objectives: continuous extensions & dependent randomized rounding Chandra Chekuri Univ. of Illinois, Urbana-Champaign

Upload: cleta

Post on 23-Feb-2016

34 views

Category:

Documents


0 download

DESCRIPTION

Algorithms for submodular objectives: continuous extensions & dependent randomized rounding. Chandra Chekuri Univ. of Illinois, Urbana-Champaign. Combinatorial Optimization. N a finite ground set w : N ! R weights on N. max/min w(S) s.t S µ N satisfies constraints. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Algorithms for submodular objectives: continuous extensions &

dependent randomized rounding

Chandra ChekuriUniv. of Illinois, Urbana-Champaign

Page 2: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding
Page 3: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Combinatorial Optimization

• N a finite ground set• w : N ! R weights on N

max/min w(S)

s.t S µ N satisfies constraints

Page 4: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Combinatorial Optimization

• N a finite ground set• w : N ! R weights on N• S µ 2N feasible solutions to problem

max/min w(S)

s.t S 2 S

Page 5: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Examples: poly-time solvable

• max weight matching • s-t shortest path in a graph• s-t minimum cut in a graph• max weight independent set in a matroid

and intersection of two matroids• ...

Page 6: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Examples: NP-Hard

• max cut• min-cost multiway/multiterminal cut• min-cost (metric) labeling• max weight independent set in a graph• ...

Page 7: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Approximation Algorithms

A is an approx. alg. for a problem:• A runs in polynomial time• maximization problem: for all instances I of

the problem A(I) ¸ ® OPT(I)• minimization problem: for all instances I of

the problem A(I) · ® OPT(I)• ® is the worst-case approximation ratio of A

Page 8: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

This talk

f is a non-negative submodular set function on NMotivation:• several applications• mathematical interest• modeling power and new results

min/max f(S)

s.t. S 2 S

Page 9: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular Set Functions

A function f : 2N ! R+ is submodular if

A Bj

f(A+j) – f(A) ¸ f(B+j) – f(B) for all A ½ B, j 2 N\B

f(A+j) – f(A) ≥ f(A+i+j) – f(A+i) for all A N , i, j N\A

Page 10: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular Set Functions

A function f : 2N ! R+ is submodular if

A Bj

f(A+j) – f(A) ¸ f(B+j) – f(B) for all A ½ B, i 2 N\B

f(A+j) – f(A) ≥ f(A+i+j) – f(A+i) for all A N , i, j N\A

Equivalently: f(A) + f(B) ≥ f(AB) + f(AB) 8 A,B N

Page 11: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

• G=(V,E) undirected graph• f : 2V ! R+ where f(S) = |δ(S)|

Cut functions in graphs

S

Page 12: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Coverage in Set Systems

• X1, X2, ..., Xn subsets of set U

• f : 2{1,2, ..., n} ! R+ where f(A) = |[ i in A Xi |

X1

X2 X3

X4X5

X1

X2 X3

X4X5

Page 13: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular Set Functions

• Non-negative submodular set functionsf(A) ≥ 0 8 A ) f(A) + f(B) ¸ f(A[ B) (sub-additive)

• Monotone submodular set functionsf(ϕ) = 0 and f(A) ≤ f(B) for all A B

• Symmetric submodular set functionsf(A) = f(N\A) for all A

Page 14: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Other examples• Cut functions in hypergraphs (symmetric non-

negative)• Cut functions in directed graphs (non-negative)• Rank functions of matroids (monotone)• Generalizations of coverage in set systems

(monotone)• Entropy/mutual information of a set of random

variables• ...

Page 15: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Max-Cut

• f is cut function of a given graph G=(V,E)• S = 2V : unconstrained • NP-Hard!

max f(S)

s.t S 2 S

Page 16: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Unconstrained problem

• minimization poly-time solvable assuming value oracle for f• Ellipsoid method [GLS’79]• Strongly-polynomial time combinatorial

algorithms [Schrijver, Iwata-Fleischer-Fujishige ’00]

• maximization NP-Hard even for explicit cut-function

min/max f(S)

Page 17: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Techniques

f is a non-negative submodular set function on N• Greedy • Local Search• Mathematical Programming Relaxation +

Rounding

min/max f(S)

s.t. S 2 S

Page 18: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Math. Programming approach

min/max w(S)

s.t S 2 S

min/max w¢x

s.t x 2 P(S)Exact algorithm: P(S) = convexhull( {1S : S 2

S})

xi 2 [0,1] indicator variable for i

Page 19: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Math. Programming approach

min/max w(S)

s.t S 2 S

min/max w¢x

s.t x 2 P(S)

Round x* 2 P(S) to S* 2 S

Exact algorithm: P(S) = convexhull( {1S : S 2 S})

Approx. algorithm: P(S) ¾ convexhull( {1S : S 2 S})

P(S) solvable: can do linear optimization over it

Page 20: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Math. Programming approach

P(S) ¶ convexhull( {1S : S 2 S}) and solvable

min/max f(S)

s.t S 2 S

min/max g(x)

s.t x 2 P(S)

Round x* 2 P(S) to S* 2 S

Page 21: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Math. Programming approach

• What is the continuous extension g ?• How to optimize with objective g ?• How do we round ?

min/max f(S)

s.t S 2 S

min/max g(x)

s.t x 2 P(S)

Round x* 2 P(S) to S* 2 S

Page 22: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Continuous extensions of f

For f : 2N ! R+ define g : [0,1]N ! R+ s.t • for any S µ N want f(S) = g(1S)

• given x = (x1, x2, ..., xn) [0,1]N want polynomial time algorithm to evaluate g(x)

• for minimization want g to be convex and for maximization want g to be concave

Page 23: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Canonical extensions: convex and concave closure

x = (x1, x2, ..., xn) [0,1]N

min/max S ®S f(S)

S ®S = 1

S ®S = xi for all i

®S ¸ 0 for all S

f-(x) for minimization and f+(x) for maximization: convex and concave respectively for any f

Page 24: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular f

• For minimization f-(x) can be evaluated in poly-time via submodular function minimization• Equivalent to the Lovasz-extension

• For maximization f+(x) is NP-Hard to evaluate even when f is monotone submodular • Rely on the multi-linear-extension

Page 25: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Lovasz-extension of f

f»(x) = Eµ 2 [0,1][ f(xµ) ] where xµ = { i | xi ¸ µ }

Example: x = (0.3, 0, 0.7, 0.1) xµ = {1,3} for µ = 0.2 and xµ = {3} for µ = 0.6f»(x) = (1-0.7) f(;) + (0.7-0.3)f({3}) + (0.3-0.1) f({1,3}) + (0.1-0) f({1,3,4}) + (0-0) f({1,2,3,4})

Page 26: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Properties of f»

• f» is convex iff f is submodular

• f»(x) = f-(x) for all x when f is submodular• Easy to evaluate f»

• For submod f : solve relax. via convex optimization min f»(x)

s.t x 2 P(S)

Page 27: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Multilinear extension of f

[Calinescu-C-Pal-Vondrak’07] inspired by [Ageev-Svir.]

For f : 2N ! R+ define F : [0,1]N ! R+ as x = (x1, x2, ..., xn) [0,1]N

R: random set, include i independently with prob. xi

F(x) = E[ f(R) ] = S N f(S) i S xi i N\S (1-xi)

Page 28: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Properties of F• F(x) can be evaluated by random sampling • F is a smooth submodular function• 2F/xixj ≤ 0 for all i,j. Recall f(A+j) – f(A) ≥ f(A+i+j) – f(A+i) for all A, i, j

• F is concave along any non-negative direction vector

• F/xi ≥ 0 for all i if f is monotone

Page 29: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Maximizing F

max { F(x) | xi 2 [0,1] for all i} is NP-Hard

equivalent to unconstrained maximization of f

When f is monotonemax { F(x) | i xi · k, xi 2 [0,1] for all i} is NP-

Hard

Page 30: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Approximately maximizing F

[Vondrak’08]Theorem: For any monotone f, there is a (1-

1/e) approximation for the problem max { F(x) | x P } where P [0,1]N is any solvable polytope.

Algorithm: Continuous-Greedy

Page 31: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Approximately maximizing F

[C-Vondrak-Zenklusen’11]Theorem: For any non-negative f, there is a ¼

approximation for the problem max { F(x) | x P } where P [0,1]n is any down-closed solvable polytope.

Remark: 0.325-approximation can be obtainedRemark: Current best 1/e ' 0.3678 [Feldman-Naor-

Schwartz’11]Algorithms: variants of local-search and continuous-

greedy

Page 32: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Math. Programming approach

• What is the continuous extension g ?• Lovasz-extension for min and multilinear ext. for max

• How to optimize with objective g ?• Convex optimization for min and O(1)-approx. alg for max

• How do we round ?

min/max f(S)

s.t S 2 S

min/max g(x)

s.t x 2 P(S)

Round x* 2 P(S) to S* 2 S

Page 33: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

Rounding and approximation depend on S and P(S)Two competing issues:• Obtain feasible solution S* from fractional

x*• Want f(S*) to be close to g(x*)

Page 34: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding approach

Viewpoint: objective function is complex• round x* to S* to approximately preserve

objective• fix/alter S* to satisfy constraints• analyze loss in fixing/altering

Page 35: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding to preserve objective

x* : fractional solution to relaxation

Minimization: f»(x) = Eµ 2 [0,1][ f(xµ) ]Pick µ uniformly at random in [0,1] (or [a, b])S* = { i | x*i ¸ µ }

Maximization: F(x) = E[f(R)] S* = pick each i 2 N independently with probability ® x*i

(® · 1)

Page 36: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Maximization

I µ 2N is a downward closed family A 2 I & B ½ A ) B 2 I

Captures “packing” problems

max f(S)

s.t S 2 I

Page 37: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Maximization

High-level results:• optimal rounding in matroid polytopes

[Calinescu-C-Vondrak-Pal’07,C-Vondrak-Zeklusen’09]]

• contention resolution scheme based rounding framework [C-Vondrak-Zenklusen’11]

Page 38: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Max k-Coverage

• X1,X2,...,Xn subsets of U and integer k

• N = {1,2,...,n}• f is the set coverage function (monotone)• I = { A µ N : |A| · k } (cardinality constraint)• NP-Hard

max f(S)

s.t S 2 I

Page 39: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Greedy

[Nemhauser-Wolsey-Fisher’78, FNW’78]• Greedy gives (1-1/e)-approximation for the

problem max { f(S) | |S| · k } when f is monotone Obtaining a (1-1/e + ²)-approximation requires exponentially many value queries to f

• Greedy give ½ for maximizing monotone f over a matroid constraint

• Unless P=NP no (1-1/e +²)-approximation for special case of Max k-Coverage [Feige’98]

Page 40: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Matroid Rounding[Calinescu-C-Pal-Vondrak’07]+[Vondrak’08]=[CCPV’09]

Theorem: There is a randomized (1-1/e) ' 0.632 approximation for maximizing a monotone f subject to any matroid constraint.

[C-Vondrak-Zenklusen’09]

Theorem: (1-1/e-²)-approximation for monotone f subject to a matroid and a constant number of packing/knapsack constraints.

Page 41: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding in Matroids[Calinescu-C-Pal-Vondrak’07]Theorem: Given any point x in P(M), there is a

randomized polynomial time algorithm to round x to a vertex X(hence an indep set of M) such that • E[X] = x• f(X) = F(X) ≥ F(x)

[C-Vondrak-Zenklusen’09] Different rounding with additional properties and apps.

Page 42: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Contention Resolution Schemes

• I an independence family on N• P(I) a relaxation for I and x 2 P(I)• R: random set from independent rounding

of xCR scheme for P(I): given x, R outputs R’ µ R s.t.

1. R’ 2 I 2. and for all i, Pr[i 2 R’ | i 2 R] ¸ c

Page 43: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding and CR schemes

Theorem: A monotone CR scheme for P(I) can be used to round s.t.

E[f(S*)] ¸ c F(x*)Via FKG inequality

max F(x)

s.t x 2 P(I)

Round x* 2 P(I)to S* 2 I

Page 44: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Summary for maximization

• Optimal results in some cases• Several new technical ideas and results• Questions led to results even for modular

case• Similar results for modular and submodular

(with in constant factors) for most known problems

Page 45: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Minimization

• Landscape is more complex• Many problems that are “easy” in modular

case are hard in submodular case: shortest paths, spanning trees, sparse cuts ...

• Some successes via Lovasz-extension• Future: need to understand special families

of submodular functions and applications

Page 46: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular-cost Vertex Cover

• Input: G=(V,E) and f : 2V ! R+

• Goal: min f(S) s.t S is a vertex cover in G• 2-approx for modular case well-known • 2-approx for sub-modular costs

[Koufogiannakis-Young’ 99, Iwata-Nagano’99, Goel etal’99]

Page 47: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular-cost Vertex Cover

• Input: G=(V,E) and f : 2V ! R+

• Goal: min f(S) s.t S is a vertex cover in G

min f»(x)xi + xj ¸ 1 for all ij 2 E xi ¸ 0 for all i 2 V

Page 48: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

Pick µ 2 [0, 1/2] uniformly at randomOutput S = { i | xi ¸ µ }

min f»(x)xi + xj ¸ 1 for all ij 2 E xi ¸ 0 for all i 2 V

Page 49: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding Analysis

Pick µ 2 [0, 1/2] uniformly at randomOutput S = { i | x*i ¸ µ }

Claim 1: S is a vertex cover with probability 1Claim 2: E[ f(S) ] · 2 f»(x*)Proof: 2f»(x) = 2s1

0 f(xµ) dµ ¸ 2s1/20 f(xµ) = E[ f(S) ]

min f»(x)xi + xj ¸ 1 for all ij 2 E xi ¸ 0 for all i 2 V

Page 50: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular-cost Set Cover

• Input: Subsets X1,...,Xn of U, f : 2N ! R+

• Goal: min f(S) s.t [i 2 S Xi = U

• Rounding according to objective gives only k-approx where k is max-element frequency. Also integrality gap of (k)

• [Iwata-Nagano’99] (k/log k)-hardness

Page 51: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

A Labeling Problem

Assign label to each vertex to minimize cost:• to label v with i cost c(v,i) • if edge uv get assigned different labels pay cost w(uv)

G=(V,E)

Labels

Page 52: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

A Labeling Problem

Assign label to each vertex to minimize cost:• to label v with i cost c(v,i) • if edge uv get assigned different labels pay cost w(uv)

G=(V,E)

Labels

Page 53: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

A Labeling Problem• Labeling problem has many applications in computer

vision [Boykov-Veksler-Zabih, ...]• Generalized to metric labeling by [Kleinberg-

Tardos’99]• 2-approx for uniform case by interesting relaxation and

rounding• O(log k)-approximation for general metric building on

uniform case• LP relaxation and O(log k) gap [C-Khanna-Naor-Zosin’01]

• Multiway-cut is special case

Page 54: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular Cost Allocation

[C-Ene’11] A new model • V a finite ground set• L = {1, 2, ..., k} set of labels• for each i a submodular cost fn fi : 2V ! R+

• Goal: assign labels to V to minimize i fi(Ai) where Ai µ V assigned label i

Page 55: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Labeling Problem as Submodular Cost Allocation

For each label i define function fi : 2V ! R

fi(S) = v 2 S c(v,i) + w(±(S))/2

G=(V,E)

Labels

Page 56: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Submodular Cost Allocation: Relaxation

x(u,i) 2 {0,1} whether u allocated label ixi = (x(u1,i), x(u2,i),....,x(un,i))

min i fi»(xi)

i x(u,i) = 1 for all u 2 V x(u,i) ¸ 0 for all u 2 V, i 2 L

Page 57: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Relaxation

• [Calinescu-Karloff-Rabani’98] relaxation for multiway cut is a special case

• [Kleinberg-Tardos’99] relaxation for uniform metric labeling is a special case

Page 58: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

Pick µ 2 [0,1] at randomfor each i let Ai = { u | x(u,i) ¸ µ }

Claim: E[ i f(Ai) ] = i fi»(xi)

min i fi»(xi)

i x(u,i) = 1 for all u 2 V x(u,i) ¸ 0 for all u 2 V, i 2 L

Page 59: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

Pick µ 2 [0,1] at randomfor each i let Ai = { u | x(u,i) ¸ µ }

Claim: E[ i f(Ai) ] = i fi»(xi)

The Ai overlap & not all vertices assigned labels!

min i fi»(xi)

i x(u,i) = 1 for all u 2 V x(u,i) ¸ 0 for all u 2 V, i 2 L

Page 60: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

(1,0,0)

(0,1,0)

(0,0,1)

Page 61: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Rounding

(1,0,0)

(0,1,0)

(0,0,1)

Page 62: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Fixing

Overlap• If fi are monotone overlap does not matter• If fi arise from single symmetric function

uncross using posi-modularity• Ensure no overlap by picking µ 2 (1/2, 1]

Not all vertices labeled• Repeat rounding until all vertices labeled • Assign all unlabeled vertices to some label

Page 63: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Results

[C-Ene’11a,C-Ene’11b]• Rounding “explains” [CKR’98,KT’99,...]• Submodular multiway partition: 1.5-

approximation for symmetric, 2-approximation for general. Improve 2 and (k-1) respectively

• Recover 2-approx for uniform metric labeling• Obtain O(log k) for generalization of labeling,

hub-location, non-metric facility location ...

Page 64: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Concluding Remarks

• Substantial progress on constrained submodular function optimization in the last few years

• New tools and connections including a general framework via continuous extensions and dependent randomized rounding

• Increased awareness and applications• Much work to be done, especially on

minimization problems

Page 65: Algorithms for  submodular objectives:  continuous extensions  & dependent  randomized  rounding

Thanks!