Optimization, Learnability, and Games: From the Lens of Smoothed Analysis Shang-Hua Teng Computer Science@Viterbi School of Engineering@USC TexPoint fonts.

Optimization, Learnability, and Games:From the Lens of Smoothed Analysis

Shang-Hua TengComputer Science@Viterbi School of Engineering@USC

Joint work with Daniel Spielman (Yale), Heiko Röglin (Maastricht University), Adam Kalai (Microsoft New England Lab), Alex Samorodnitsky (Hebrew University), Xi Chen (USC), Xiaotie Deng (City University of Hong Kong)

This Talk

• Part I: Overview of Smoothed Analysis

• Part II: Multiobjective Optimization

• Part III: Machine Learning

• Part VI: Games, Markets and Equilibrium

• Part V: Discussions

Practical Performance of Algorithms “While theoretical work on models of computation and

methods for analyzing algorithms has had enormous payoff, we are not done. In many situations, simple algorithms do well. Take for example the Simplex algorithm for linear programming, or the success of simulated annealing of contain supposedly intractable problems. We don't understand why! It is apparent that worst-case analysis does not provide useful insights on the performance of algorithms and heuristics and our models of computation need to be further developed and refined. Theoreticians are investing increasingly in careful experimental work leading to identification of important new questions in algorithms area. Developing means for predicting the performance of algorithms and heuristics on real data and on real computers is a grand challenge in algorithms.”

-- Challenges for Theory of Computing: Report for an NSF-Sponsored Workshop on Research in Theoretical Computer Science (Condon, Edelsbrunner, Emerson, Fortnow, Haber, Karp, Leivant, Lipton, Lynch, Parberry, Papadimitriou, Rabin, Rosenberg, Royer, Savage,Selman, Smith, Tardos, and Vitter), 1999

Linear Programming & Simplex Method

max s.t.

Worst-Case: exponentialAverage-Case: polynomialWidely used in practice

Smoothed Analysis of Simplex Method(Spielman + Teng, 2001)

Theorem: For all A, b, c, simplex method takes expected time polynomial in

max s.t.

maxs.t.

G is Gaussian

Smoothed Complexity

Interpolates between worst and average case

Considers neighborhood of every input

If low, all bad inputs are unstable

Data in practice are not arbitrary but could be generated with noises and imprecision

Optimization: Single Criterion & Multiobjective

min f(x) subject to x ∈ S.

Examples:

• Linear Programming

• Shortest path

• Minimum spanning tree

• TSP

• Set cover

Optimization: Single Criterion & Multiobjective

Real-life logistical problems often involve multiple

objectives

• Travel time, fare, departure time

• Delay, cost, reliability

• Profit and risk

Optimization: Single Criterion & Multiobjective

min f1(x), ..., min fd(x) subject to x ∈ S

There may not be a solution that is simultaneously optimal for all fi

Question: What can we do algorithmically to support a decision maker?

Pareto-Optimal Solutions

x S ∈ dominates y S ∈

iff

∀i : fi(x) ≤ fi(y) and ∃i : fi(x) < fi(y)

Pareto-Optimal Solutions

Pareto-Optimal Solutions

Pareto Curve

Pareto Surface

Decision Makers only Choose Pareto-Optimal Solutions

Fact: Every monotone function, e.g., 1 f1(x)+ ... +d fd(x)is optimized by a Pareto-optimal solution.

Computational Problem:Return the Pareto curve (surface, set)

Decision Makers only Choose Pareto-Optimal Solutions

Return the Pareto curve (surface, set)

Central Question: How large is the Pareto set?

A Concrete Model

S : can encode arbitrary combinatorial structure.

Examples: all paths from s to t, all Hamiltonian cycles, all spanning trees, . . .

How Large can a Pareto Set be?

• Worst Case: Exponential

• In Practice: Usually smaller

– Müller-Hannemann, Weihe (2001)

Train Connection

travel time, fare, number of train changes

Smoothed Models

Pareto Set is Usually Small(Röglin-Teng)

d = 2 [Beier-Röglin-Vöcking, 2007]: O(n2φ)

How Many Pareto Points in an -interval

The Winner

The Losers and their Gaps

A Non-Concentration Lemma

Putting Together

Nearly Tight Smoothed Bounds for 2D: Many Moments

Three or More Objectives

Not So Tight Yet: But Polynomial Smoothed Bound for Fixed Dimensions

This Talk

• Part I: Overview of Smoothed Analysis

• Part II: Multi-objective Optimization

• Part III: Machine Learning

• Part VI: Games, Markets and Equilibrium

• Part V: Discussions

P.A.C. Learning!?

X = {0,1}ⁿ f: X → {–1,+1}

PAC assumption: target is from a particular concept class

(for example, an AND, e.g. f(x) = “Bank” & “Adam” & “Free”)

Input: training data (xj from D, f(xj)) j≤mNoiseless

NIGERIA BANK VIAGRA ADAM LASER SALE FREE IN f(x)

x1 YES YES YES NO NO YES NO YES SPAM

x2 YES NO NO YES YES YES YES YES LEGIT

x3 NO YES YES YES YES YES YES YES LEGIT

x4 YES YES YES NO NO NO NO YES SPAM

x5 YES YES YES YES YES NO YES YES SPAM

[Valiant84]

P.A.C Learning

Poly-time learning algorithm– Succeed with prob. ≥ 1- (e.g. 0.99)

– m = # examples = poly(n/ε)

Output: h: X → {–1,+1} with

err(h) = Prx←D[ h(x)≠f(x) ] ≤

OPTIONAL: “Proper” learning: the class from which h is.

Agnostic P.A.C. Learning!?

X = {0,1}ⁿ f: X → {–1,+1}

Without PAC assumption: target is from a particular concept class

Input: training data (xj from D, f(xj)) j≤m

Poly-time learning algorithm

– Succeed with prob. ≥ 1- (e.g. 0.99)

– m = # examples = poly(n/ε)

Output: h: X → {–1,+1} with

err(h) = Prx←D[ h(x) ≠ f(x) ] ≤ + ming from the class err(g)

[Kearns Schapire Sellie 92]

• Computation is limiting resource– “Easy” ignoring computation

– YET: Children learn many things computers can’t

– Worst-case poly-time algorithms?• PAC-learn DNF, Decision trees, juntas

• Learning parity with noise

Computational Learning Theory

Some Smoothed Results in Learning(Kalai-Samorodnitsky-Teng)

• PAC learn decision trees over smoothed (constant-bounded) product distributions

• PAC learn DNFs over smoothed (constant-bounded) product distribution

• Agnostically learn decision trees over smoothed (constant-bounded) product distributions

A Formal Statement of the First Result

For μ ϵ [0,1]ⁿ, let πμ be the product distribution where entries of μ define the mean of Boolean variables

Theorem 1: Concept Function: decision tree f: {0,1}ⁿ → {–1,+1} of size s Distribution: πμ defined by μ ϵ ν+[–.01,.01]ⁿ where ν ϵ [.02,.98]ⁿ Data: m=poly(ns/ε) training examples (xj, f(xj)) j≤m: xj iid from πμ, Learning Algorithm: a polynomial-time algorithm Output: a function h Quality: Prx←πμ

[ sgn(h(x))≠f(x) ] ≤ ε.

Fourier over Product Distributions

• x ϵ {0,1}ⁿ, μ ϵ [0,1]ⁿ,

1

22

2

(1 )

[ ] (also called ( , ))

ˆ ˆ( ) ( ) ( ) E [ ( ) ]

ˆ ˆ ( ) (Pars

for any

, where

1 eval)

ˆ ˆ ( )

i ii

i i

S i Si S

S SS

S

S

x

n x

f x f S f

x

x x S

x S f x

f f

x

S

f f S

E [ ]i ix

Non-Concentration Bound on Fourier Structures

For any f:{0,1}ⁿ→{–1,1}, α,β > 0, and d ≥ 1,

Continuous generalization of Schwartz-Zippel theoremLet p:Rⁿ→R be a degree-d multi-linear polynomial with leading

coefficient of 1. Then, for any ό>0,

25[.49,.51]

ˆ ˆPr s.t. ( ) ( ) 200n

dS T f S f T T d

[ 1,1]Pr ( ) 2

n

d

xp x

ò ò

e.g., p(x)=x1x2x9+.3x7–0.2

Some Related Work

• Decision Trees:

• P.A.C. Membership Queries:

• Uniform Distributions

[Kushilevitz-Mansour’91; Goldreich-Levin’89]

• [Bshouty’94]

• Agnostic Membership Queries: Uniform D

[Gopalan-Kalai-Klivans’08]• DNF: P.A.C. Membership Queries + Uniform D [Jackson’94]

Some Smoothed Results in Learning(Kalai-Samorodnitsky-Teng)

• PAC learn decision trees over smoothed (constant-bounded) product distributions

• PAC learn DNFs over smoothed (constant-bounded) product distribution

• Agnostically learn decision trees over smoothed (constant-bounded) product distributions

Games and Optimization

Optimization

President UUSA(xUSA,xCA,xMA,…)

Global optimum Local optimumApproximation

Multi-Objective Optimization

President UUSA(xUSA,xCA,xMA,…)

Pareto optimum [Approximation]

UCA(xUSA,xCA,xMA …) UMA(xUSA,xCA,xMA,…)

Multi-Player Games

President UUSA(xUSA,xCA,xMA,…)

Best response Nash equilibrium

Governor of CAUCA(xUSA,xCA,xMA,…)

Governor of MAUMA(xUSA,xCA,xMA,…)

“Is the smoothed complexity of (another classic algorithm,)Lemke-Howson (algorithm) for two-player games,

polynomial?”

0

0

1

-1

-1

1

-1

1

0

0

1

-1

1

-1

-1

1

0

0

BIMATRIX Games

Mixed Strategies

Mixed equilibrium always exists:

Search Problem: Find an equilibrium

Nash Equilibria in Two-Player Games

Exchange Economies

• Traders

• Goods

• Initial Endowments:

• Utilities:

Arrow-Debreu Equilibrium Price

A price vector

Distributed Exchange

• Every Trader:– Sells the initial endowment to “market”: (to get a budget)

– Buys from the “market” to optimize her individual utilities

• Market Clearing Price

Smoothed Model

Complexity of Nash Equilibria

[Daskalakis-Goldberg-Papadimitriou, 2005]• For any constant k ≥ 3, NASH is PPAD-hard.

[Chen-Deng, 2005]• 2-player NASH is PPAD-complete.

[Chen-Deng-Teng, 2006]• If PPAD is not in P, then 2-player NASH does not have a fully

polynomial-time approximation scheme

Smoothed Complexity of Equilibria

[Chen-Deng-Teng, 2006]

• NO Smoothed Polynomial-Time Complexity for Lemke-Howson or any BIMATRIX algorithm, unless computation of game and market equilibria and Brouwer fixed points is in randomized P!

[Huang-Teng, 2006]

• Computation of Arrow-Debreu equilibria in Leontief Exchange Economies is not in Smoothed P, unless …

PSPACE

NP

PLS PPAD

Complexity Classes and Complete Problems

P

Tale of Two Types of Equilibria

Local Search

(Potential Games)• Linear Programming

– P

• Simplex Method– Smoothed P

• PLS– FPTAS

• Intuitive

Fixed-Point Computation

(Matrix Games)• 2-Player Nash equilibrium

– Unknown

• Lemke-Howson Algorithm– If in P, then NASH in RP

• PPAD– FPTAS, then NASH in RP

• Intuitive to some

A Basic Question

Is fixed point computation fundamentally harder than local search?

Random Separation of Local Search and Fixed Point Computation

Aldous (1983): • Randomization helps local search

Chen & Teng (2007):• Randomization doesn’t help Fixed-Point-

Computation!!!

… in the black-box query model

Open Questions

• How hard is PPAD?

• Non-concentration of multi-linear polynomials

• Optimal smoothed bound for Pareto Sets

Non-Concentration of Multi-linear Polynomials

Continuous Schwartz-Zippel Conjecture:Let p:Rⁿ→R be a degree-d multi-linear polynomial with constant

coefficient of 1. Then, for any ό>0,

[ 1,1]Pr ( ) 2

n

d

xp x

ò ò