Distributed Parametric-Nonparametric Estimation in...

Distributed Parametric-Nonparametric

Estimation in Networked Control Systems

Damiano Varagnolo

Department of Information Engineering - University of Padova

April, 18th 2011

entirely

writtenin

Beamer and Tik Z

Damiano Varagnolo (DEI) varagnolo@dei.unipd.it April, 18th 2011 1 / 48

introduction

Research topics

topics

multi-agent

systems

smart grids

parametric

regression

nonparametric

regression

estimation of

# of agents

application

of consensus

distributed

optimization

introduction

Research topics

topics

multi-agent

systems

smart grids

parametric

regression

nonparametric

regression

estimation of

# of agents

application

of consensus

distributed

optimization

introduction

Research topics

topics

multi-agent

systems

smart grids

parametric

regression

nonparametric

regression

estimation of

# of agents

application

of consensus

distributed

optimization

nonparametric

regression

estimation of

# of agents

introduction

Multi-agent systems: examples of applications

Multi-Robot

Coordination

Underwater

Exploration

Smart Grids

Deployment

Wind Power

Management

Wild Life

Monitoring

Highways

Deployment

Surveillance

Implementation

introduction

First problem considered in this speech

Assumption

noisy measurements of

f (x , t) : R3 × R 7→ R

that are

non uniformly sampled in space x

non uniformly sampled in time t

taken by di�erent agents

Objective

smoothing in space (x) and forecast in time (t) the quantity f (x , t)

introduction

Example 1 - channel gains in geographical areas

source: Dall'Anese et al., 2011

x ∈ R2 : position

t : time

f (x , t) : channel gain

introduction

Example 2 - waves power extraction

source: www.graysharboroceanenergy.com

x ∈ R2 : position

t : time

f (x , t) : sea level

introduction

Example 3 - multi robot exploration

source: http://www-robotics.jpl.nasa.gov

x ∈ R2 : position

f (x) : ground level

introduction

Di�culties related to this problem

Information-related di�culties

non-uniform samplings both in time and in space

unknown dynamics of f

unknown or extremely complex correlations in time and space

Hardware-related di�culties

energy & computational & memory & bandwidth limitations

Framework-related di�culties

mobile and time varying network

introduction

State of the art

proposed

distributed

solutions

Least Squares

Maximum Likelihoods

Kalman Filtering

Kriging

Other Learning Techniques

Choi et al. 2009, Predd et al. 2006, Boyd et al. 2005

Schizas et al. 2008, Barbarossa

and Scutari 2007, Boyd et al.

Cressie and Wikle 2002,

Olfati-Saber 2007, Carli

et al. 2008

Dall'Anese et al. 2011, Cortés 2010

Nguyen et al. 2005, Bazerque et al. 2010

nonparametric

introduction

State of the art

proposed

distributed

solutions

Least Squares

Maximum Likelihoods

Kalman Filtering

Kriging

et al. 2008

dynamic scenarios

introduction

State of the art

proposed

distributed

solutions

Least Squares

Maximum Likelihoods

Kalman Filtering

Kriging

et al. 2008

static scenarios

introduction

State of the art - Vision

obtain

f (x , t) = Ψ (past measurements)

distributed

capable of both smoothing and prediction

Our approach

nonparametric: Ψ (·) lives in an in�nite dimensional space

introduction

Why should we use a nonparametric approach?

Motivations

it could be di�cult or even impossible to de�ne a parametric

model (e.g. when only regularity assumptions are available)

parametric models could involve a large number of parameters

(could require nonlinear optimization techniques)

lead to convex optimization problems

consistent, i.e. f → f when # measurements →∞ (De Nicolao,

Ferrari-Trecate, 1999)

introduction

State of the art - where we actually contributed

our small

puzzle-piece

agents estimate

the same f

static scenario

(f indepen-

dent of t)

Regularization

Network

approach

Poggio and Girosi 1990

e.g. exploration

no needs to discard

old measurements

introduction

Our goal

obtain a simple, self-evaluating and auto-

tuning multi-agent regression strategy

framework & assumptions

Framework

Agents:

noisily sample the same f

limited computational &

communication capabilities

1 measurement × agent (ease

of notation)

M measurements in total

framework & assumptions

Measurement model

ym = f (xm) + νm (1)

f : X ⊂ Rd → R unknown (X compact)

νm ⊥ xm, zero mean and variance σ2

xm ∼ µ i.i.d. (agents know µ!!)

examples of µ:

uniform jitter generic

distributed estimators

Considered cost function

Q (f ) =M∑

(ym − f (xm))2 + γ ‖f ‖2K

lives in an in�nite

dimensional space

regularization factor,

K : X × X → RMercer kernel

Centralized optimal solution as a Regularization Network

fc =M∑

cmK (xm, ·)[

([K (x1, x1) · · · K (x1, xM )

K (xM , x1) · · · K (xM , xM )

)−1[y1

Q (f ) =M∑

(ym − f (xm))2 + γ ‖f ‖2K

lives in an in�nite

dimensional space

regularization factor,

K : X × X → RMercer kernel

Centralized optimal solution as a Regularization Network

fc =M∑

cmK (xm, ·)[

([K (x1, x1) · · · K (x1, xM )

K (xM , x1) · · · K (xM , xM )

)−1[y1

Drawbacks

fc =M∑

cmK (xm, ·)[

([K (x1, x1) · · · K (x1, xM )

K (xM , x1) · · · K (xM , xM )

)−1 [y1

computational cost: O (M3) (inversion of M ×M matrix)

transmission cost: O (M) (knowledge of whole {xm, ym}Mm=1)

need to �nd alternative solutions

Alternative centralized optimal solution (1st on 2)

Structure of K implies

• K (x1, x2) =+∞∑

λeφe(x1)φe(x2)λe = eigenvalue

φe = eigenfunction

• f (x) =+∞∑

beφe(x)

⇒ measurement model can be rewritten as

Cm:=︷︸︸︷[φ1 (xm) , φ2 (xm) , . . .]

b:=︷︸︸︷

b1b2...

+νm (2)

Alternative centralized optimal solution (2nd on 2)

)−1(1

CTm ym

involves in�nite dimensional objects:

• · · · · · ·...

. . ....

•......

⇒ cannot be computed exactly

Suboptimal �nite dimensional solution

New estimator

)−1(1

computable (involves E × E matrices and E -dimensional vectors)

minimizes QE (b) :=∑M

(ym − CE

+ γ∑E

e=1b2eλe

Drawbacks

1 O (E 3) computational e�ort

2 O (E 2) transmission e�ort

3 must know M

Suboptimal �nite dimensional solution

New estimator

)−1(1

computable (involves E × E matrices and E -dimensional vectors)

minimizes QE (b) :=∑M

(ym − CE

+ γ∑E

e=1b2eλe

Drawbacks

1 O (E 3) computational e�ort

2 O (E 2) transmission e�ort

3 must know M

Derivation of the distributed estimator

)−1(1

Consider the approximations

M → Mg (guess)

)TCEm → Eµ

Derivation of the distributed estimator

obtain: bd =

)−1(1

Advantages

1 O (E ) computational e�ort

2 O (E ) transmission e�ort

Summary of proposed estimation schemes

br (E )

br : O(E3

)comput., O

)transm.

bd (E ,Mg )

bd : O (E) comput., O (E) transm.

original hyp. space

bcbc : O

)comput., O (M) transm.

reduced hyp. s

br (E )

br : O(E3

)comput., O

)transm.

bd (E ,Mg )

original hyp. space

bcbc : O

reduced hyp. s

br (E )

br : O(E3

)comput., O

)transm.

bd (E ,Mg )

original hyp. space

bcbc : O

reduced hyp. s

Quanti�cation of performancesAssumption: E ,Mg already chosen, bd already computed

computable through distributed MC

local residuals

∝ I − 1

‖bc − bd‖2 ≤1

|rm| + ‖UMbd‖2 + ‖UCbd‖2

local residuals

∝ I − 1

‖bc − bd‖2 ≤1

local residuals

∝ I − 1

‖bc − bd‖2 ≤1

local residuals

∝ I − 1

‖bc − bd‖2 ≤1

local residuals

∝ I − 1

‖bc − bd‖2 ≤1

Tuning of the parameters - key ideasAssumption: have some information on the energy of f

parameters to

be estimated

number of

eigenfunctions E

number of mea-

surements Mg

assure ‖bc − br (E )‖to be su�ciently small

minimize the bound

on ‖bc − bd (E ,Mg )‖

parameters to

be estimated

number of

eigenfunctions E

number of mea-

surements Mg

minimize the bound

parameters to

be estimated

number of

eigenfunctions E

number of mea-

surements Mg

minimize the bound

Regression strategy e�ectiveness exampleM = 100, E = 20, Mmin = 90, Mmax = 110, SNR ≈ 2.5

0 0.2 0.4 0.6 0.8 1

true f

Regression strategy e�ectiveness exampleM = 100, E = 20, Mmin = 90, Mmax = 110, SNR ≈ 2.5

0 0.2 0.4 0.6 0.8 1

Accuracy of the computed boundM = 100, E = 20, Mmin = 90, Mmax = 110

0 0.25 0.5 0.75

true distance ‖bd − bc‖2

Comparison with oracleM = 100, E = 20, Mmin = 90, Mmax = 110

0 0.2 0.4 0.6

∥∥boracled − bc∥∥2

∥ ∥ bours

∥ ∥ 2

conclusions

Conclusions and future works for this part

Conclusions

Strategy:

is e�ective and easy to be implemented

has self-evaluation capabilities

has self-tuning capabilities

Future works

exploit statistical knowledge about M

incorporate e�ects of �nite number of steps in consensus

algorithms

extend to dynamic scenarios (long term objective)

Part Two

Introduction

Privacy-aware number of agents estimation

Estimation of the number of agents (A) can be important in:

distributed estimation

analysis of connectivity

We assume privacy concerns → do not use IDs!

Our goal: obtain an easily implementable

distributed estimator satisfying the constraints

A simple example

The basic idea

Algorithm:

generation

average

consensus

Maximum

LikelihoodA−1 = y 2

avedoes not require

sending IDs

A simple example

The basic idea

y1 ∼ N (0, 1)

y2 ∼ N (0, 1)

y3 ∼ N (0, 1)

y4 ∼ N (0, 1)

y5 ∼ N (0, 1)Algorithm:

generation

average

consensus

Maximum

avedoes not require

sending IDs

A simple example

The basic idea

y1 →1

y2 →1

y3 →1

y4 →1

y5 →1

yaAlgorithm:

generation

average

consensus

Maximum

avedoes not require

sending IDs

A simple example

The basic idea

yave ∼ N(0, 1

)yave ∼ N

yave ∼ N(0, 1

)Algorithm:

generation

average

consensus

Maximum

avedoes not require

sending IDs

A simple example

The basic idea

yave ∼ N(0, 1

)yave ∼ N

yave ∼ N(0, 1

)Algorithm:

generation

average

consensus

Maximum

avedoes not require

sending IDs

A simple example

Reformulating the idea as a block scheme

local distributed local

N (0, 1)

aver. cons. ML A

A simple example

Plausible ways to generalize the idea

local distributed local

p (·)

y 11 , . . . , yr1

y 12 , . . . , yr2

y 1A, . . . , yrA

F Ψ A

• N (µ, σ2)

• U [α, β]

• ??

• average• max• ??

• ML

• MMSE

• MAP

• ??Damiano Varagnolo (DEI) varagnolo@dei.unipd.it April, 18th 2011 35 / 48

Theoretical results

Which cost function we consider

notice: we want to estimate A−1 instead of A

↓estimator =: A−1

A−1 − A−1)2] (

≡ variance if A−1 unbiased)

convenient in order to obtain mathematical results

in our cases, asymptotically in r :

limr→+∞

(A−1 − A−1

)2 = E

(A− A

Theoretical results

Theoretical results: average-consensus + ML

p (·)

aver. cons. ML S

Assumptions

y ·a generated through Gaussian distributions N (µ, σ2)

fusion of y ·a is through average-consensus

Results: ML estimators:

writable in closed form are MVUE (Minimum Variance and Unbiased)

performances: var

(A−1 − A−1

r(independent of µ and σ2)

(conjecture: Law of Large Numbers) if r → +∞ then performances are

independent of p (·)Damiano Varagnolo (DEI) varagnolo@dei.unipd.it April, 18th 2011 37 / 48

Theoretical results

p (·)

aver. cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

aver. cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

aver. cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

aver. cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

Theoretical results: max-consensus + ML

p (·)

max cons. ML S

Assumptions

cumulative distribution P (·) of y ·a is strictly monotonic andcontinuous

fusion of y ·a is through max-consensus

performances: var

(A−1 − A−1

rindependent of P (·)

performances are twice as good as average-consensus

Theoretical results

p (·)

max cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

max cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

max cons. ML S

Assumptions

performances: var

(A−1 − A−1

Theoretical results

p (·)

max cons. ML S

Assumptions

performances: var

(A−1 − A−1

Simulative results

Results of various simulated systems (1)

N (0, 1) average consensus ML A = 10

5 10 15

p( A∣ ∣ ∣A)

r = 10

r = 40

r = 70

r = 100

Simulative results

U [0, 1] max consensus ML A = 10

5 10 15

p( A∣ ∣ ∣A)

r = 10

r = 40

r = 70

r = 100

Simulative results

U [0, 1] max consensus ML A = 100

50 100 150

p( A∣ ∣ ∣A)

r = 10

r = 40

r = 70

r = 100

Conclusions. . . and future extensions

Conclusions

e�ective and robust algorithm

quanti�able performances

rely on statistical concepts → preserves privacy

inherits good qualities of consensus strategies

Future extensions

analyze optimal quantization strategies

�nd optimal distributions for average consensus

use the strategy for topological change detection purposes

Distributed Parametric-Nonparametric

Estimation in Networked Control Systems

Damiano Varagnolo

Department of Information Engineering - University of Padova

April, 18th 2011

damiano.varagnolo@dei.unipd.it

www.dei.unipd.it/∼varagnolo/google: damiano varagnolo

licensed under the Creative Commons BY-NC-SA 2.5 Italy License:

entirely

writtenin

Beamer and Tik Z

Publications

Distributed Optimization

F. Zanella, D. Varagnolo, A. Cenedese, G. Pillonetto, L. Schenato.

Newton-Raphson consensus for distributed convex optimization.

IEEE Conference on Decision and Control (CDC 2011) (submitted)

Applications of Consensus

S. Bolognani, S. Del Favero, L. Schenato, D. Varagnolo.

Consensus-based distributed sensor calibration and least-square

parameter identi�cation in WSNs. International Journal of Robust

and Nonlinear Control, 2010

S. Bolognani, S. Del Favero, L. Schenato, D. Varagnolo.

Distributed sensor calibration and least-square parameter

identi�cation in WSNs using consensus algorithms. Proceedings of

Allerton Conference on Communication Control and Computing

(Allerton 2008)

Publications

Parametric Regression

D. Varagnolo, G. Pillonetto, L. Schenato. Distributed

consensus-based Bayesian estimation: su�cient conditions for

performance characterization. 2010 American Control Conference,

D. Varagnolo, P. Chen, L. Schenato, S. Sastry. Performance

analysis of di�erent routing protocols in wireless sensor networks

for real-time estimation. IEEE Conference on Decision and Control

(CDC 2008)

Publications

Nonparametric Regression / Classi�cation

D. Varagnolo, G. Pillonetto, L. Schenato. Distributed parametric

and nonparametric regression with on-line performance bounds

computation. Automatica (submitted)

D. Varagnolo, G. Pillonetto, L. Schenato. Auto-tuning procedures

for distributed nonparametric regression algorithms. IEEE

Conference on Decision and Control (CDC 2011) (submitted)

S. Del Favero, D. Varagnolo, F. Dinuzzo, L. Schenato, G.

Pillonetto. On the discardability of data in Support Vector

Classi�cation problems. IEEE Conference on Decision and Control

(CDC 2011) (submitted)

D. Varagnolo, G. Pillonetto, L. Schenato. Distributed Function

and Time Delay Estimation using Nonparametric Techniques. IEEE

Conference on Decision and Control (CDC 2009)

Publications

Number of Sensors Estimation

D. Varagnolo, G. Pillonetto, L. Schenato. Distributed statistical

estimation of the number of nodes in Sensor Networks. IEEE

Conference on Decision and Control (CDC 2010)

Smart Grids

E. Bitar, A. Giani, R. Rajagopal, D. Varagnolo, P. Khargonekar,

K. Poolla, V. Pravin. Optimal Contracts for Wind Power Producers

in Electricity Markets. Conference on Decision and Control

CDC10, 2010

left mouse / page down / space next slide

right mouse / page up / backspace previous slide

q / esc exit

tab overview toggling

t clock toggling

home go to �rst slide

end go to last slide

l go to the previously last seen slide

b fade the screen to black

w fade the screen to white

f full screen toggling

enter spotlight toggling

+ / - adjust the spotlight size

mouse wheel adjust the spotlight size

left mouse (dragging a box) highglight a box

right mouse (on a highlighted box) remove the highlight of that box

z zoom toggling

right mouse (in zoom modality) move on the image

Distributed Parametric-Nonparametric Estimation in...

Documents

Transcript of Distributed Parametric-Nonparametric Estimation in...

NONPARAMETRIC STATISTICS Nonparametric statistics or distribution free statistics is use when the population from which the samples are selected is not.

Generalized Linear Models - Yanfei Kang, Ph.D. · 2018. 6. 7. · Generalized Linear Models Lecture 10: Nonparametric regression. Nonparametric regression y ... Local linear and local

Collaborare per Crescere - Padova - 27_09_2013

Semi and Nonparametric Models in Econometrics - Part I - Crest

[Συλλογικό] - The Greek Question Speeches of Soviet Delegation at UN General Assembly 1947 [Soviet News 1947]

A Convex Hull Peeling Depth Approach to Nonparametric ...

Semi and Nonparametric Models in Econometrics

Nonparametric Estimation via Convex Programming · Nonparametric Estimation via Convex Programming Anatoli Juditsky⁄ Arkadi Nemirovskiy Abstract In the paper, we focus primarily

Universit a degli Studi di Padova Facolt a di Ingegneriatesi.cab.unipd.it/40204/1/TesiAlessia.pdf · Universit a degli Studi di Padova ... correct function of other metabolic pathways

International e-Congress on Physical Education & Sport Science...6 28th International e-Congress on Physical Education & Sport Science 12-14 June 2020 ORAL PRESENTATIONS Keynote Speeches

ORAL REPORT PRESENTED TO THE EUROPEAN COUNCIL IN ...european-convention.europa.eu/docs/speeches/9604.pdf · rule of adopting "European laws" and "European framework laws" by joint

Università degli Studi di Padova Facoltà di Ingegneria

Neutrinos from nuclear reactors - neutrino.pd.infn.itneutrino.pd.infn.it/lezioni/Corso2009/Reattori.pdf · Mauro Mezzetto, INFN Padova Neutrini da Reattori Corso di Dottorato, Maggio

Discussion of ABC talk by Stefano Cabras, Padova, March 21, 2013

Discussion of ABC talk by Francesco Pauli, Padova, March 21, 2013

Nonparametric Bayesian Methods 1 What is …larry/=sml/nonparbayes.pdfNonparametric Bayesian Methods 1 What is Nonparametric Bayes? In parametric Bayesian inference we have a model

PHYSICAL REVIEW D 112011 (2020) · 18Dipartimento di Fisica, INFN Sezione di Padova and Universit`a di Padova, I-35131 Padova, Italy 19INFN Sezione di Roma and Universit`a di Roma

Testing Parametric Regression Speciﬁcations with ...cshalizi/uADA/12/lectures/ch10.pdfChapter 10 Testing Parametric Regression Speciﬁcations with Nonparametric Regression 10.1

On the Testability of Identiﬁcation in Some Nonparametric ...

1 PARAMETRIC VERSUS NONPARAMETRIC STATISTICS Heibatollah Baghi, and Mastee Badii.