SlideShare a Scribd company logo
Startup ADMM
Shin Matsushima
Department of Statistics
Purdue University
Lab Seminar
April 2, 2012
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 1 / 22
The Paper
”Distributed Optimization and Statistical Learning
via the Alternating Direction Method of Multipliers”
Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato and Jonathan Eckstein
Foundations and Trends in Machine Learning Vol. 3, No. 1 (2010) 1-122
URL: http://www.stanford.edu/˜boyd/papers/pdf/admm distr stats.pdf
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 2 / 22
Outline
1 Dual Ascent
2 Method of Multipliers
3 Alterative Direction Method of Multipliers (ADMM)
4 The following contents
Consider the equality-comnstrained convex optimization problem
minimize f (x)
subject to Ax = b
where x ∈ Rn
, A ∈ Rm×n
and f is convex.
Lagrangian:
L(x, y) = f (x) + y (Ax − b)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 4 / 22
Assume strong duality:
inf
x
sup
y
L(x, y) = sup
y
inf
x
L(x, y)
g(y):concave
Aiming to solve the dual problem:
maximize g(y) = inf
x
f (x) + y (Ax − b)
and recover a primal optimal point using y (optimal solution of above)
x = argmin
x
L(x, y )
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 5 / 22
Dual Ascent
Procedure(Dual Ascent):
xk+1
= argmin
x
L(x, yk
)
= argmin
x
f (x) + y Ax
yk+1
= yk
+ αk
(Axk+1
− b)
Axk+1 − b ∈ ∂g(yk) because
g(yk
) = min
x
f (x) + yk
(Ax − b) = f (xk+1
) + yk
(Axk+1
− b)
g(y) = min
x
f (x) + y (Ax − b) ≤ f (xk+1
) + y (Axk+1
− b)
⇒ g(y) − g(yk
) ≤ (y − yk
) (Axk+1
− b)
Necessary to chose appropriate stepsize αk
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 6 / 22
Dual Decomposition
In the problem
minimize f (x)
subject to Ax = b,
assume f (x) = N
i=1 fi (xi ) where xi ∈ Rni
, x = [x1 · · · xN ]
Let Ai ∈ Rm×ni
, A = [A1 · · · AN]
Then the algorithm becomes decentraized:
xk+1
= argmin
x
L(x, yk
)
= argmin
xi
f (xi ) + yk
Ai xi for i = 1, . . . , N
yk+1
broadcast
= yk
+ αk
(A xk+1
gather
−b)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 7 / 22
Outline
1 Dual Ascent
2 Method of Multipliers
3 Alterative Direction Method of Multipliers (ADMM)
4 The following contents
Original Problem:
minimize f (x)
subject to Ax = b
Original Lagrangian:
L(x, y) = f (x) + y (Ax − b)
Augmented Lagrangian:
Lρ(x, y) = f (x) + y (Ax − b) + ρ/2 Ax − b 2
is considered Lagrangian for the following equality-comnstrained convex
optimization problem wich is equivalent to the original problem
minimize f (x) + ρ/2 Ax − b 2
subject to Ax = b
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 9 / 22
Method of Multipliers
Procedure(Method of Multipliers):
xk+1
= argmin
x
Lρ(x, yk
)
= argmin
x
L(x, yk
) + ρ/2 Ax − b 2
yk+1
= yk
+ ρ(Axk+1
− b)
stepsize is now a fixed constant ρ.
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 10 / 22
An easy understanding for setting stepsize to ρ is the following:
We can see
xk+1
= argmin
x
Lρ(x, yk
)
⇒ 0 ∈ ∂xk+1 {f (xk+1
) + yk
Axk+1
+ ρ/2 Axk+1
− b 2
}
= ∂xk+1 f (xk+1
) + A yk
+ ρ(Axk+1
− b)
= ∂xk+1 f (xk+1
) + A yk+1
This implies that MM keeps
0 ∈ ∂f (xk+1
) + A yk+1
after every iteration. Note that
0 ∈ ∂f (x ) + A y (dual feasibility)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 11 / 22
Method of Multipliers has more improved convergence property.
But the augmented term disables to make it separate.
xk+1
= argmin
x
Lρ(x, yk
)
= argmin
x
L(x, yk
) + ρ/2 Ax − b 2
= argmin
xi
fi (xi ) + yk
Ai xi + · · ·
yk+1
= yk
+ ρ(Axk+1
− b)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 12 / 22
Outline
1 Dual Ascent
2 Method of Multipliers
3 Alterative Direction Method of Multipliers (ADMM)
4 The following contents
ADMM solves problems in the following form
minimize f (x) + g(z)
subject to Ax + Bz = c
Augmented Lagrangian:
Lρ(x, z, y) = f (x) + g(z) + y (Ax + Bz − c) + ρ/2 Ax − Bz − c 2
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 14 / 22
Alternative Direction Method of Mutipliers(ADMM)
Procedure(ADMM):
xk+1
= argmin
x
Lρ(x, zk
, yk
)
= argmin
x
f (x) + yk
(Ax + Bzk
− c) + ρ/2 Ax + Bzk
− c 2
zk+1
= argmin
z
Lρ(xk+1
, z, yk
)
yk+1
= yk
+ ρ(Axk+1
+ Bzk+1
− c)
This includes x-minimization step and z-minimization step.
c.f. Method of multipliers should be
xk+1
, zk+1
= argmin
x,z
Lρ(x, z, yk
)
yk+1
= yk
+ ρ(Axk+1
+ Bzk+1
− c)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 15 / 22
Scaled Form of Alternative Direction Method of
Mutipliers(ADMM)
Let rk = Axk + Bzk − c (residual) and uk = (1/ρ)yk(scaled dual variable),
The procedure of ADMM can be rewritten as follows:
Procedure(scaled form of ADMM):
xk+1
= argmin
x
Lρ(x, zk
, yk
)
= argmin
x
f (x) + yk
(Ax + Bzk
− c) + ρ/2 Ax + Bzk
− c 2
= argmin
x
f (x) + ρ/2 Ax + Bzk
− c + uk 2
zk+1
= argmin
z
Lρ(xk+1
, z, yk
)
uk+1
= uk
+ (Axk+1
+ Bzk+1
− c)
= uk
+ rk+1
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 16 / 22
Convergence of ADMM
Assumption :
1 f and g are closed, proper, and convex
2 L(x, z, y) = L0(x, z, y) has a saddle point
Result : As k → ∞,
1 Residual Convergence
rk
→ 0
2 Objective Convergence
f (xk
) + g(zk
) → p∗
3 Dual variable Convergence
yk
→ y∗
where y∗
is a dual optimal point
In practice:
a few tens of iteration will often produce modest accuracy solution
it can be very slow to get high accuracy solution
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 17 / 22
Outline
1 Dual Ascent
2 Method of Multipliers
3 Alterative Direction Method of Multipliers (ADMM)
4 The following contents
Other Characteristics of ADMM
optimality condition and stop conditioin
some variants
Varying ρk
→ 0
More general Augumenting terms
Inexact x/z-minimization step
...
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 19 / 22
Remarkable Application Discussed in the Following Chapter
Chap 5: Problem with more general constraint
minimize f (x)
subject to x ∈ C
is transfoemed as
minimize f (x) + IC(z)
subject to x − z = 0
Procedure:
xk+1
= argmin
x
f (x) + ρ/2 x − zk
+ uk 2
zk+1
= argmin
z∈C
ρ/2 xk+1
− z + uk 2
uk+1
= uk
+ (xk+1
− zk+1
)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 20 / 22
Chap 7: Distributed Version (Consensus) of
minimize f (x) =
N
i=1
fi (x)
is transfromed as
minimize
N
i=1
fi (xi )
subject to xi − z = 0 i = 1, . . . , N
Procedure:
xk+1
= argmin
x
Lρ(x, zk
, yk
)
= argmin
xi
fi (xi ) + yk
i (xi − z) + ρ/2 xi − z 2
zk+1
= 1/N
N
i=1
(xk+1
i + (1/ρ)yk
i )
yk+1
= yk
+ ρ(xk+1
− zk+1
)
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 21 / 22
Contents
4 General patterns
Tips and tools used after this chaper
5 Constrained Convex Optimization
How to incorpolate general constraints
6
1-Norm Problems
Discussion about problems involving 1-Norm
7 Consensus and Sharing
Framework for distributed optimization
8 Distributed Model Fitting
Examples for the distributed optimization
9 Nonconvex problems
10 Implementation
11 Numerical Examples
12 Conclusion
Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 22 / 22

More Related Content

What's hot

Multilinear Twisted Paraproducts
Multilinear Twisted ParaproductsMultilinear Twisted Paraproducts
Multilinear Twisted Paraproducts
VjekoslavKovac1
 
An application of the hyperfunction theory to numerical integration
An application of the hyperfunction theory to numerical integrationAn application of the hyperfunction theory to numerical integration
An application of the hyperfunction theory to numerical integration
HidenoriOgata
 
WITMSE 2013
WITMSE 2013WITMSE 2013
WITMSE 2013
Joe Suzuki
 
On Twisted Paraproducts and some other Multilinear Singular Integrals
On Twisted Paraproducts and some other Multilinear Singular IntegralsOn Twisted Paraproducts and some other Multilinear Singular Integrals
On Twisted Paraproducts and some other Multilinear Singular Integrals
VjekoslavKovac1
 
Bayesian Criteria based on Universal Measures
Bayesian Criteria based on Universal MeasuresBayesian Criteria based on Universal Measures
Bayesian Criteria based on Universal Measures
Joe Suzuki
 
Introductory maths analysis chapter 02 official
Introductory maths analysis   chapter 02 officialIntroductory maths analysis   chapter 02 official
Introductory maths analysis chapter 02 official
Evert Sandye Taasiringan
 
Boundedness of the Twisted Paraproduct
Boundedness of the Twisted ParaproductBoundedness of the Twisted Paraproduct
Boundedness of the Twisted Paraproduct
VjekoslavKovac1
 
Density theorems for anisotropic point configurations
Density theorems for anisotropic point configurationsDensity theorems for anisotropic point configurations
Density theorems for anisotropic point configurations
VjekoslavKovac1
 
Introductory Mathematical Analysis for Business Economics International 13th ...
Introductory Mathematical Analysis for Business Economics International 13th ...Introductory Mathematical Analysis for Business Economics International 13th ...
Introductory Mathematical Analysis for Business Economics International 13th ...
Leblancer
 
Tales on two commuting transformations or flows
Tales on two commuting transformations or flowsTales on two commuting transformations or flows
Tales on two commuting transformations or flows
VjekoslavKovac1
 
A Szemeredi-type theorem for subsets of the unit cube
A Szemeredi-type theorem for subsets of the unit cubeA Szemeredi-type theorem for subsets of the unit cube
A Szemeredi-type theorem for subsets of the unit cube
VjekoslavKovac1
 
Simplified Runtime Analysis of Estimation of Distribution Algorithms
Simplified Runtime Analysis of Estimation of Distribution AlgorithmsSimplified Runtime Analysis of Estimation of Distribution Algorithms
Simplified Runtime Analysis of Estimation of Distribution Algorithms
PK Lehre
 
A Szemerédi-type theorem for subsets of the unit cube
A Szemerédi-type theorem for subsets of the unit cubeA Szemerédi-type theorem for subsets of the unit cube
A Szemerédi-type theorem for subsets of the unit cube
VjekoslavKovac1
 
Some fixed point theorems in fuzzy mappings
Some fixed point theorems in fuzzy mappingsSome fixed point theorems in fuzzy mappings
Some fixed point theorems in fuzzy mappings
Alexander Decker
 
Scattering theory analogues of several classical estimates in Fourier analysis
Scattering theory analogues of several classical estimates in Fourier analysisScattering theory analogues of several classical estimates in Fourier analysis
Scattering theory analogues of several classical estimates in Fourier analysis
VjekoslavKovac1
 
A common fixed point theorem for two weakly compatible mappings satisfying a ...
A common fixed point theorem for two weakly compatible mappings satisfying a ...A common fixed point theorem for two weakly compatible mappings satisfying a ...
A common fixed point theorem for two weakly compatible mappings satisfying a ...
Alexander Decker
 
11.a common fixed point theorem for two weakly compatible mappings satisfying...
11.a common fixed point theorem for two weakly compatible mappings satisfying...11.a common fixed point theorem for two weakly compatible mappings satisfying...
11.a common fixed point theorem for two weakly compatible mappings satisfying...
Alexander Decker
 
Introductory maths analysis chapter 10 official
Introductory maths analysis   chapter 10 officialIntroductory maths analysis   chapter 10 official
Introductory maths analysis chapter 10 official
Evert Sandye Taasiringan
 
A Note on Latent LSTM Allocation
A Note on Latent LSTM AllocationA Note on Latent LSTM Allocation
A Note on Latent LSTM Allocation
Tomonari Masada
 
Mesh Processing Course : Active Contours
Mesh Processing Course : Active ContoursMesh Processing Course : Active Contours
Mesh Processing Course : Active Contours
Gabriel Peyré
 

What's hot (20)

Multilinear Twisted Paraproducts
Multilinear Twisted ParaproductsMultilinear Twisted Paraproducts
Multilinear Twisted Paraproducts
 
An application of the hyperfunction theory to numerical integration
An application of the hyperfunction theory to numerical integrationAn application of the hyperfunction theory to numerical integration
An application of the hyperfunction theory to numerical integration
 
WITMSE 2013
WITMSE 2013WITMSE 2013
WITMSE 2013
 
On Twisted Paraproducts and some other Multilinear Singular Integrals
On Twisted Paraproducts and some other Multilinear Singular IntegralsOn Twisted Paraproducts and some other Multilinear Singular Integrals
On Twisted Paraproducts and some other Multilinear Singular Integrals
 
Bayesian Criteria based on Universal Measures
Bayesian Criteria based on Universal MeasuresBayesian Criteria based on Universal Measures
Bayesian Criteria based on Universal Measures
 
Introductory maths analysis chapter 02 official
Introductory maths analysis   chapter 02 officialIntroductory maths analysis   chapter 02 official
Introductory maths analysis chapter 02 official
 
Boundedness of the Twisted Paraproduct
Boundedness of the Twisted ParaproductBoundedness of the Twisted Paraproduct
Boundedness of the Twisted Paraproduct
 
Density theorems for anisotropic point configurations
Density theorems for anisotropic point configurationsDensity theorems for anisotropic point configurations
Density theorems for anisotropic point configurations
 
Introductory Mathematical Analysis for Business Economics International 13th ...
Introductory Mathematical Analysis for Business Economics International 13th ...Introductory Mathematical Analysis for Business Economics International 13th ...
Introductory Mathematical Analysis for Business Economics International 13th ...
 
Tales on two commuting transformations or flows
Tales on two commuting transformations or flowsTales on two commuting transformations or flows
Tales on two commuting transformations or flows
 
A Szemeredi-type theorem for subsets of the unit cube
A Szemeredi-type theorem for subsets of the unit cubeA Szemeredi-type theorem for subsets of the unit cube
A Szemeredi-type theorem for subsets of the unit cube
 
Simplified Runtime Analysis of Estimation of Distribution Algorithms
Simplified Runtime Analysis of Estimation of Distribution AlgorithmsSimplified Runtime Analysis of Estimation of Distribution Algorithms
Simplified Runtime Analysis of Estimation of Distribution Algorithms
 
A Szemerédi-type theorem for subsets of the unit cube
A Szemerédi-type theorem for subsets of the unit cubeA Szemerédi-type theorem for subsets of the unit cube
A Szemerédi-type theorem for subsets of the unit cube
 
Some fixed point theorems in fuzzy mappings
Some fixed point theorems in fuzzy mappingsSome fixed point theorems in fuzzy mappings
Some fixed point theorems in fuzzy mappings
 
Scattering theory analogues of several classical estimates in Fourier analysis
Scattering theory analogues of several classical estimates in Fourier analysisScattering theory analogues of several classical estimates in Fourier analysis
Scattering theory analogues of several classical estimates in Fourier analysis
 
A common fixed point theorem for two weakly compatible mappings satisfying a ...
A common fixed point theorem for two weakly compatible mappings satisfying a ...A common fixed point theorem for two weakly compatible mappings satisfying a ...
A common fixed point theorem for two weakly compatible mappings satisfying a ...
 
11.a common fixed point theorem for two weakly compatible mappings satisfying...
11.a common fixed point theorem for two weakly compatible mappings satisfying...11.a common fixed point theorem for two weakly compatible mappings satisfying...
11.a common fixed point theorem for two weakly compatible mappings satisfying...
 
Introductory maths analysis chapter 10 official
Introductory maths analysis   chapter 10 officialIntroductory maths analysis   chapter 10 official
Introductory maths analysis chapter 10 official
 
A Note on Latent LSTM Allocation
A Note on Latent LSTM AllocationA Note on Latent LSTM Allocation
A Note on Latent LSTM Allocation
 
Mesh Processing Course : Active Contours
Mesh Processing Course : Active ContoursMesh Processing Course : Active Contours
Mesh Processing Course : Active Contours
 

Similar to A

Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
Tomoya Murata
 
02 basics i-handout
02 basics i-handout02 basics i-handout
02 basics i-handout
sheetslibrary
 
Universal Prediction without assuming either Discrete or Continuous
Universal Prediction without assuming either Discrete or ContinuousUniversal Prediction without assuming either Discrete or Continuous
Universal Prediction without assuming either Discrete or Continuous
Joe Suzuki
 
Normal density and discreminant analysis
Normal density and discreminant analysisNormal density and discreminant analysis
Normal density and discreminant analysis
VARUN KUMAR
 
Lec04 min cost linear problems
Lec04 min cost linear problemsLec04 min cost linear problems
Lec04 min cost linear problems
Danny Luk
 
NIPS2010: optimization algorithms in machine learning
NIPS2010: optimization algorithms in machine learningNIPS2010: optimization algorithms in machine learning
NIPS2010: optimization algorithms in machine learning
zukun
 
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
Geoffrey Négiar
 
Stochastic Alternating Direction Method of Multipliers
Stochastic Alternating Direction Method of MultipliersStochastic Alternating Direction Method of Multipliers
Stochastic Alternating Direction Method of Multipliers
Taiji Suzuki
 
Distributed Support Vector Machines
Distributed Support Vector MachinesDistributed Support Vector Machines
Distributed Support Vector Machines
Harsha Vardhan Tetali
 
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
The Statistical and Applied Mathematical Sciences Institute
 
Lecture note4coordinatedescent
Lecture note4coordinatedescentLecture note4coordinatedescent
Lecture note4coordinatedescent
Xudong Sun
 
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
UK Carbon Capture and Storage Research Centre
 
01. Differentiation-Theory & solved example Module-3.pdf
01. Differentiation-Theory & solved example Module-3.pdf01. Differentiation-Theory & solved example Module-3.pdf
01. Differentiation-Theory & solved example Module-3.pdf
RajuSingh806014
 
Accelerated reconstruction of a compressively sampled data stream
Accelerated reconstruction of a compressively sampled data streamAccelerated reconstruction of a compressively sampled data stream
Accelerated reconstruction of a compressively sampled data stream
Pantelis Sopasakis
 
Tensor Train data format for uncertainty quantification
Tensor Train data format for uncertainty quantificationTensor Train data format for uncertainty quantification
Tensor Train data format for uncertainty quantification
Alexander Litvinenko
 
Recursive Compressed Sensing
Recursive Compressed SensingRecursive Compressed Sensing
Recursive Compressed Sensing
Pantelis Sopasakis
 
Ece3075 a 8
Ece3075 a 8Ece3075 a 8
Ece3075 a 8
Aiman Malik
 
Adaptive Three Operator Splitting
Adaptive Three Operator SplittingAdaptive Three Operator Splitting
Adaptive Three Operator Splitting
Fabian Pedregosa
 
Kernels and Support Vector Machines
Kernels and Support Vector  MachinesKernels and Support Vector  Machines
Kernels and Support Vector Machines
Edgar Marca
 
2. Definite Int. Theory Module-5.pdf
2. Definite Int. Theory Module-5.pdf2. Definite Int. Theory Module-5.pdf
2. Definite Int. Theory Module-5.pdf
RajuSingh806014
 

Similar to A (20)

Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
Doubly Accelerated Stochastic Variance Reduced Gradient Methods for Regulariz...
 
02 basics i-handout
02 basics i-handout02 basics i-handout
02 basics i-handout
 
Universal Prediction without assuming either Discrete or Continuous
Universal Prediction without assuming either Discrete or ContinuousUniversal Prediction without assuming either Discrete or Continuous
Universal Prediction without assuming either Discrete or Continuous
 
Normal density and discreminant analysis
Normal density and discreminant analysisNormal density and discreminant analysis
Normal density and discreminant analysis
 
Lec04 min cost linear problems
Lec04 min cost linear problemsLec04 min cost linear problems
Lec04 min cost linear problems
 
NIPS2010: optimization algorithms in machine learning
NIPS2010: optimization algorithms in machine learningNIPS2010: optimization algorithms in machine learning
NIPS2010: optimization algorithms in machine learning
 
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
Stochastic Frank-Wolfe for Constrained Finite Sum Minimization @ Montreal Opt...
 
Stochastic Alternating Direction Method of Multipliers
Stochastic Alternating Direction Method of MultipliersStochastic Alternating Direction Method of Multipliers
Stochastic Alternating Direction Method of Multipliers
 
Distributed Support Vector Machines
Distributed Support Vector MachinesDistributed Support Vector Machines
Distributed Support Vector Machines
 
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
Deep Learning Opening Workshop - ProxSARAH Algorithms for Stochastic Composit...
 
Lecture note4coordinatedescent
Lecture note4coordinatedescentLecture note4coordinatedescent
Lecture note4coordinatedescent
 
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
Beck Workshop on Modelling and Simulation of Coal-fired Power Generation and ...
 
01. Differentiation-Theory & solved example Module-3.pdf
01. Differentiation-Theory & solved example Module-3.pdf01. Differentiation-Theory & solved example Module-3.pdf
01. Differentiation-Theory & solved example Module-3.pdf
 
Accelerated reconstruction of a compressively sampled data stream
Accelerated reconstruction of a compressively sampled data streamAccelerated reconstruction of a compressively sampled data stream
Accelerated reconstruction of a compressively sampled data stream
 
Tensor Train data format for uncertainty quantification
Tensor Train data format for uncertainty quantificationTensor Train data format for uncertainty quantification
Tensor Train data format for uncertainty quantification
 
Recursive Compressed Sensing
Recursive Compressed SensingRecursive Compressed Sensing
Recursive Compressed Sensing
 
Ece3075 a 8
Ece3075 a 8Ece3075 a 8
Ece3075 a 8
 
Adaptive Three Operator Splitting
Adaptive Three Operator SplittingAdaptive Three Operator Splitting
Adaptive Three Operator Splitting
 
Kernels and Support Vector Machines
Kernels and Support Vector  MachinesKernels and Support Vector  Machines
Kernels and Support Vector Machines
 
2. Definite Int. Theory Module-5.pdf
2. Definite Int. Theory Module-5.pdf2. Definite Int. Theory Module-5.pdf
2. Definite Int. Theory Module-5.pdf
 

Recently uploaded

Pitch Deck Teardown: Kinnect's $250k Angel deck
Pitch Deck Teardown: Kinnect's $250k Angel deckPitch Deck Teardown: Kinnect's $250k Angel deck
Pitch Deck Teardown: Kinnect's $250k Angel deck
HajeJanKamps
 
Discover the Beauty and Functionality of The Expert Remodeling Service
Discover the Beauty and Functionality of The Expert Remodeling ServiceDiscover the Beauty and Functionality of The Expert Remodeling Service
Discover the Beauty and Functionality of The Expert Remodeling Service
obriengroupinc04
 
Cover Story - China's Investment Leader - Dr. Alyce SU
Cover Story - China's Investment Leader - Dr. Alyce SUCover Story - China's Investment Leader - Dr. Alyce SU
Cover Story - China's Investment Leader - Dr. Alyce SU
msthrill
 
Best Competitive Marble Pricing in Dubai - ☎ 9928909666
Best Competitive Marble Pricing in Dubai - ☎ 9928909666Best Competitive Marble Pricing in Dubai - ☎ 9928909666
Best Competitive Marble Pricing in Dubai - ☎ 9928909666
Stone Art Hub
 
8328958814KALYAN MATKA | MATKA RESULT | KALYAN
8328958814KALYAN MATKA | MATKA RESULT | KALYAN8328958814KALYAN MATKA | MATKA RESULT | KALYAN
8328958814KALYAN MATKA | MATKA RESULT | KALYAN
➑➌➋➑➒➎➑➑➊➍
 
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
Cambridge Product Management Network
 
The latest Heat Pump Manual from Newentide
The latest Heat Pump Manual from NewentideThe latest Heat Pump Manual from Newentide
The latest Heat Pump Manual from Newentide
JoeYangGreatMachiner
 
GKohler - Retail Scavenger Hunt Presentation
GKohler - Retail Scavenger Hunt PresentationGKohler - Retail Scavenger Hunt Presentation
GKohler - Retail Scavenger Hunt Presentation
GraceKohler1
 
DearbornMusic-KatherineJasperFullSailUni
DearbornMusic-KatherineJasperFullSailUniDearbornMusic-KatherineJasperFullSailUni
DearbornMusic-KatherineJasperFullSailUni
katiejasper96
 
Prescriptive analytics BA4206 Anna University PPT
Prescriptive analytics BA4206 Anna University PPTPrescriptive analytics BA4206 Anna University PPT
Prescriptive analytics BA4206 Anna University PPT
Freelance
 
list of states and organizations .pdf
list of  states  and  organizations .pdflist of  states  and  organizations .pdf
list of states and organizations .pdf
Rbc Rbcua
 
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
irhcs
 
Satta Matka Dpboss Kalyan Matka Results Kalyan Chart
Satta Matka Dpboss Kalyan Matka Results Kalyan ChartSatta Matka Dpboss Kalyan Matka Results Kalyan Chart
Satta Matka Dpboss Kalyan Matka Results Kalyan Chart
Satta Matka Dpboss Kalyan Matka Results
 
Pro Tips for Effortless Contract Management
Pro Tips for Effortless Contract ManagementPro Tips for Effortless Contract Management
Pro Tips for Effortless Contract Management
Eternity Paralegal Services
 
Enhancing Adoption of AI in Agri-food: Introduction
Enhancing Adoption of AI in Agri-food: IntroductionEnhancing Adoption of AI in Agri-food: Introduction
Enhancing Adoption of AI in Agri-food: Introduction
Cor Verdouw
 
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
BBPMedia1
 
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
dpbossdpboss69
 
Efficient PHP Development Solutions for Dynamic Web Applications
Efficient PHP Development Solutions for Dynamic Web ApplicationsEfficient PHP Development Solutions for Dynamic Web Applications
Efficient PHP Development Solutions for Dynamic Web Applications
Harwinder Singh
 
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
concepsionchomo153
 
Kirill Klip GEM Royalty TNR Gold Copper Presentation
Kirill Klip GEM Royalty TNR Gold Copper PresentationKirill Klip GEM Royalty TNR Gold Copper Presentation
Kirill Klip GEM Royalty TNR Gold Copper Presentation
Kirill Klip
 

Recently uploaded (20)

Pitch Deck Teardown: Kinnect's $250k Angel deck
Pitch Deck Teardown: Kinnect's $250k Angel deckPitch Deck Teardown: Kinnect's $250k Angel deck
Pitch Deck Teardown: Kinnect's $250k Angel deck
 
Discover the Beauty and Functionality of The Expert Remodeling Service
Discover the Beauty and Functionality of The Expert Remodeling ServiceDiscover the Beauty and Functionality of The Expert Remodeling Service
Discover the Beauty and Functionality of The Expert Remodeling Service
 
Cover Story - China's Investment Leader - Dr. Alyce SU
Cover Story - China's Investment Leader - Dr. Alyce SUCover Story - China's Investment Leader - Dr. Alyce SU
Cover Story - China's Investment Leader - Dr. Alyce SU
 
Best Competitive Marble Pricing in Dubai - ☎ 9928909666
Best Competitive Marble Pricing in Dubai - ☎ 9928909666Best Competitive Marble Pricing in Dubai - ☎ 9928909666
Best Competitive Marble Pricing in Dubai - ☎ 9928909666
 
8328958814KALYAN MATKA | MATKA RESULT | KALYAN
8328958814KALYAN MATKA | MATKA RESULT | KALYAN8328958814KALYAN MATKA | MATKA RESULT | KALYAN
8328958814KALYAN MATKA | MATKA RESULT | KALYAN
 
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
2024.06 CPMN Cambridge - Beyond Now-Next-Later.pdf
 
The latest Heat Pump Manual from Newentide
The latest Heat Pump Manual from NewentideThe latest Heat Pump Manual from Newentide
The latest Heat Pump Manual from Newentide
 
GKohler - Retail Scavenger Hunt Presentation
GKohler - Retail Scavenger Hunt PresentationGKohler - Retail Scavenger Hunt Presentation
GKohler - Retail Scavenger Hunt Presentation
 
DearbornMusic-KatherineJasperFullSailUni
DearbornMusic-KatherineJasperFullSailUniDearbornMusic-KatherineJasperFullSailUni
DearbornMusic-KatherineJasperFullSailUni
 
Prescriptive analytics BA4206 Anna University PPT
Prescriptive analytics BA4206 Anna University PPTPrescriptive analytics BA4206 Anna University PPT
Prescriptive analytics BA4206 Anna University PPT
 
list of states and organizations .pdf
list of  states  and  organizations .pdflist of  states  and  organizations .pdf
list of states and organizations .pdf
 
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
1Q24_HYUNDAI CAPITAL SERVICES INC. AND SUBSIDIARIES
 
Satta Matka Dpboss Kalyan Matka Results Kalyan Chart
Satta Matka Dpboss Kalyan Matka Results Kalyan ChartSatta Matka Dpboss Kalyan Matka Results Kalyan Chart
Satta Matka Dpboss Kalyan Matka Results Kalyan Chart
 
Pro Tips for Effortless Contract Management
Pro Tips for Effortless Contract ManagementPro Tips for Effortless Contract Management
Pro Tips for Effortless Contract Management
 
Enhancing Adoption of AI in Agri-food: Introduction
Enhancing Adoption of AI in Agri-food: IntroductionEnhancing Adoption of AI in Agri-food: Introduction
Enhancing Adoption of AI in Agri-food: Introduction
 
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
NIMA2024 | De toegevoegde waarde van DEI en ESG in campagnes | Nathalie Lam |...
 
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
Call 8867766396 Dpboss Matka Guessing Satta Matta Matka Kalyan Chart Indian M...
 
Efficient PHP Development Solutions for Dynamic Web Applications
Efficient PHP Development Solutions for Dynamic Web ApplicationsEfficient PHP Development Solutions for Dynamic Web Applications
Efficient PHP Development Solutions for Dynamic Web Applications
 
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
欧洲杯投注-欧洲杯投注外围盘口-欧洲杯投注盘口app|【​网址​🎉ac22.net🎉​】
 
Kirill Klip GEM Royalty TNR Gold Copper Presentation
Kirill Klip GEM Royalty TNR Gold Copper PresentationKirill Klip GEM Royalty TNR Gold Copper Presentation
Kirill Klip GEM Royalty TNR Gold Copper Presentation
 

A

  • 1. Startup ADMM Shin Matsushima Department of Statistics Purdue University Lab Seminar April 2, 2012 Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 1 / 22
  • 2. The Paper ”Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers” Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato and Jonathan Eckstein Foundations and Trends in Machine Learning Vol. 3, No. 1 (2010) 1-122 URL: http://www.stanford.edu/˜boyd/papers/pdf/admm distr stats.pdf Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 2 / 22
  • 3. Outline 1 Dual Ascent 2 Method of Multipliers 3 Alterative Direction Method of Multipliers (ADMM) 4 The following contents
  • 4. Consider the equality-comnstrained convex optimization problem minimize f (x) subject to Ax = b where x ∈ Rn , A ∈ Rm×n and f is convex. Lagrangian: L(x, y) = f (x) + y (Ax − b) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 4 / 22
  • 5. Assume strong duality: inf x sup y L(x, y) = sup y inf x L(x, y) g(y):concave Aiming to solve the dual problem: maximize g(y) = inf x f (x) + y (Ax − b) and recover a primal optimal point using y (optimal solution of above) x = argmin x L(x, y ) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 5 / 22
  • 6. Dual Ascent Procedure(Dual Ascent): xk+1 = argmin x L(x, yk ) = argmin x f (x) + y Ax yk+1 = yk + αk (Axk+1 − b) Axk+1 − b ∈ ∂g(yk) because g(yk ) = min x f (x) + yk (Ax − b) = f (xk+1 ) + yk (Axk+1 − b) g(y) = min x f (x) + y (Ax − b) ≤ f (xk+1 ) + y (Axk+1 − b) ⇒ g(y) − g(yk ) ≤ (y − yk ) (Axk+1 − b) Necessary to chose appropriate stepsize αk Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 6 / 22
  • 7. Dual Decomposition In the problem minimize f (x) subject to Ax = b, assume f (x) = N i=1 fi (xi ) where xi ∈ Rni , x = [x1 · · · xN ] Let Ai ∈ Rm×ni , A = [A1 · · · AN] Then the algorithm becomes decentraized: xk+1 = argmin x L(x, yk ) = argmin xi f (xi ) + yk Ai xi for i = 1, . . . , N yk+1 broadcast = yk + αk (A xk+1 gather −b) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 7 / 22
  • 8. Outline 1 Dual Ascent 2 Method of Multipliers 3 Alterative Direction Method of Multipliers (ADMM) 4 The following contents
  • 9. Original Problem: minimize f (x) subject to Ax = b Original Lagrangian: L(x, y) = f (x) + y (Ax − b) Augmented Lagrangian: Lρ(x, y) = f (x) + y (Ax − b) + ρ/2 Ax − b 2 is considered Lagrangian for the following equality-comnstrained convex optimization problem wich is equivalent to the original problem minimize f (x) + ρ/2 Ax − b 2 subject to Ax = b Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 9 / 22
  • 10. Method of Multipliers Procedure(Method of Multipliers): xk+1 = argmin x Lρ(x, yk ) = argmin x L(x, yk ) + ρ/2 Ax − b 2 yk+1 = yk + ρ(Axk+1 − b) stepsize is now a fixed constant ρ. Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 10 / 22
  • 11. An easy understanding for setting stepsize to ρ is the following: We can see xk+1 = argmin x Lρ(x, yk ) ⇒ 0 ∈ ∂xk+1 {f (xk+1 ) + yk Axk+1 + ρ/2 Axk+1 − b 2 } = ∂xk+1 f (xk+1 ) + A yk + ρ(Axk+1 − b) = ∂xk+1 f (xk+1 ) + A yk+1 This implies that MM keeps 0 ∈ ∂f (xk+1 ) + A yk+1 after every iteration. Note that 0 ∈ ∂f (x ) + A y (dual feasibility) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 11 / 22
  • 12. Method of Multipliers has more improved convergence property. But the augmented term disables to make it separate. xk+1 = argmin x Lρ(x, yk ) = argmin x L(x, yk ) + ρ/2 Ax − b 2 = argmin xi fi (xi ) + yk Ai xi + · · · yk+1 = yk + ρ(Axk+1 − b) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 12 / 22
  • 13. Outline 1 Dual Ascent 2 Method of Multipliers 3 Alterative Direction Method of Multipliers (ADMM) 4 The following contents
  • 14. ADMM solves problems in the following form minimize f (x) + g(z) subject to Ax + Bz = c Augmented Lagrangian: Lρ(x, z, y) = f (x) + g(z) + y (Ax + Bz − c) + ρ/2 Ax − Bz − c 2 Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 14 / 22
  • 15. Alternative Direction Method of Mutipliers(ADMM) Procedure(ADMM): xk+1 = argmin x Lρ(x, zk , yk ) = argmin x f (x) + yk (Ax + Bzk − c) + ρ/2 Ax + Bzk − c 2 zk+1 = argmin z Lρ(xk+1 , z, yk ) yk+1 = yk + ρ(Axk+1 + Bzk+1 − c) This includes x-minimization step and z-minimization step. c.f. Method of multipliers should be xk+1 , zk+1 = argmin x,z Lρ(x, z, yk ) yk+1 = yk + ρ(Axk+1 + Bzk+1 − c) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 15 / 22
  • 16. Scaled Form of Alternative Direction Method of Mutipliers(ADMM) Let rk = Axk + Bzk − c (residual) and uk = (1/ρ)yk(scaled dual variable), The procedure of ADMM can be rewritten as follows: Procedure(scaled form of ADMM): xk+1 = argmin x Lρ(x, zk , yk ) = argmin x f (x) + yk (Ax + Bzk − c) + ρ/2 Ax + Bzk − c 2 = argmin x f (x) + ρ/2 Ax + Bzk − c + uk 2 zk+1 = argmin z Lρ(xk+1 , z, yk ) uk+1 = uk + (Axk+1 + Bzk+1 − c) = uk + rk+1 Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 16 / 22
  • 17. Convergence of ADMM Assumption : 1 f and g are closed, proper, and convex 2 L(x, z, y) = L0(x, z, y) has a saddle point Result : As k → ∞, 1 Residual Convergence rk → 0 2 Objective Convergence f (xk ) + g(zk ) → p∗ 3 Dual variable Convergence yk → y∗ where y∗ is a dual optimal point In practice: a few tens of iteration will often produce modest accuracy solution it can be very slow to get high accuracy solution Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 17 / 22
  • 18. Outline 1 Dual Ascent 2 Method of Multipliers 3 Alterative Direction Method of Multipliers (ADMM) 4 The following contents
  • 19. Other Characteristics of ADMM optimality condition and stop conditioin some variants Varying ρk → 0 More general Augumenting terms Inexact x/z-minimization step ... Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 19 / 22
  • 20. Remarkable Application Discussed in the Following Chapter Chap 5: Problem with more general constraint minimize f (x) subject to x ∈ C is transfoemed as minimize f (x) + IC(z) subject to x − z = 0 Procedure: xk+1 = argmin x f (x) + ρ/2 x − zk + uk 2 zk+1 = argmin z∈C ρ/2 xk+1 − z + uk 2 uk+1 = uk + (xk+1 − zk+1 ) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 20 / 22
  • 21. Chap 7: Distributed Version (Consensus) of minimize f (x) = N i=1 fi (x) is transfromed as minimize N i=1 fi (xi ) subject to xi − z = 0 i = 1, . . . , N Procedure: xk+1 = argmin x Lρ(x, zk , yk ) = argmin xi fi (xi ) + yk i (xi − z) + ρ/2 xi − z 2 zk+1 = 1/N N i=1 (xk+1 i + (1/ρ)yk i ) yk+1 = yk + ρ(xk+1 − zk+1 ) Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 21 / 22
  • 22. Contents 4 General patterns Tips and tools used after this chaper 5 Constrained Convex Optimization How to incorpolate general constraints 6 1-Norm Problems Discussion about problems involving 1-Norm 7 Consensus and Sharing Framework for distributed optimization 8 Distributed Model Fitting Examples for the distributed optimization 9 Nonconvex problems 10 Implementation 11 Numerical Examples 12 Conclusion Shin Matsushima (Purdue University) Startup ADMM April 2, 2012 22 / 22