a machine learning approach for constrained …a machine learning approach for constrained sensor...
TRANSCRIPT
Problem formulationMotivation
ContributionFurther improvements
A machine learning approach for constrainedsensor placement
Kévin KASPER, Lionel MATHELIN, Hisham ABOU-KANDIL
ENS Cachan - CNRS
July 3, 2015
Kévin KASPER A machine learning approach for constrained sensor placement 1 / 27
Problem formulationMotivation
ContributionFurther improvements
GoalFind sensor locations that enable efficient field estimation
through the use of a linear estimator
Available DataSnapshot sequence Y(Machine Learning)Number of sensors : nSPossible sensor locations(constraint)
The sensor locations must :
Minimise εr = ||Y − Y ||F/||Y ||F“Work” with few sensorsNot overlap (constraint)
Popular in Monitoring and Control
Structure integrity monitoring (bridges, ...)Non-invasive defect detectionFluid flow state estimation and control
Kévin KASPER A machine learning approach for constrained sensor placement 2 / 27
Problem formulationMotivation
ContributionFurther improvements
GoalFind sensor locations that enable efficient field estimation
through the use of a linear estimator
Available DataSnapshot sequence Y(Machine Learning)Number of sensors : nSPossible sensor locations(constraint)
The sensor locations must :
Minimise εr = ||Y − Y ||F/||Y ||F“Work” with few sensorsNot overlap (constraint)
Popular in Monitoring and Control
Structure integrity monitoring (bridges, ...)Non-invasive defect detectionFluid flow state estimation and control
Kévin KASPER A machine learning approach for constrained sensor placement 2 / 27
Problem formulationMotivation
ContributionFurther improvements
GoalFind sensor locations that enable efficient field estimation
through the use of a linear estimator
Available DataSnapshot sequence Y(Machine Learning)Number of sensors : nSPossible sensor locations(constraint)
The sensor locations must :
Minimise εr = ||Y − Y ||F/||Y ||F“Work” with few sensorsNot overlap (constraint)
Popular in Monitoring and Control
Structure integrity monitoring (bridges, ...)Non-invasive defect detectionFluid flow state estimation and control
Kévin KASPER A machine learning approach for constrained sensor placement 2 / 27
Problem formulationMotivation
ContributionFurther improvements
Contents
1 Problem formulation
2 Motivation
3 Contribution
4 Further improvements
Kévin KASPER A machine learning approach for constrained sensor placement 3 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
Contents
1 Problem formulationLearning SequenceSensorsConstraints
2 Motivation
3 Contribution
4 Further improvements
Kévin KASPER A machine learning approach for constrained sensor placement 4 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
A snapshoty i ∈ Rnx
(nx = 3015)
A collection of nsnap ∈ N∗ pressurefield snapshots is formed(nsnap = 1875)
Each snapshot is spatiallydiscretized on the same grid
Learning Sequence
Y =[y iy2 . . . ynsnap
]∈ Rnx×nsnap
contains spatially and temporallydiscretized data
Kévin KASPER A machine learning approach for constrained sensor placement 5 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
A snapshoty i ∈ Rnx
(nx = 3015)
A collection of nsnap ∈ N∗ pressurefield snapshots is formed(nsnap = 1875)
Each snapshot is spatiallydiscretized on the same grid
Learning Sequence
Y =[y iy2 . . . ynsnap
]∈ Rnx×nsnap
contains spatially and temporallydiscretized data
Kévin KASPER A machine learning approach for constrained sensor placement 5 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
Choose wiselyThe learning sequence is critical tobe able to guarantee goodestimation performances→Requires expertise
x1
x2
−1 0 1
−1
0
1
Time
Drag
0 50 100 150 200 250 300 350 400 450 5001.2
1.3
1.4
1.5
1.6
1.7
1.8
Kévin KASPER A machine learning approach for constrained sensor placement 6 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
Sensor specificitiesThe nS ∈ N∗ sensors :
are linked to specific grid points.give s i , the exact value of y i onthose grid points
x1
x2
−1 0 1
−1
0
1
To manipulate the sensors, we use :the set containing the grid indexes of the sensors : posS orthe matrix C ∈ RnS×nx with s i = Cy i
C has a special structure → it contains only one 1 per line(indexed by posS)→ CY contains the posS row restriction of Y
Kévin KASPER A machine learning approach for constrained sensor placement 7 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
Sensor specificitiesThe nS ∈ N∗ sensors :
are linked to specific grid points.give s i , the exact value of y i onthose grid points
x1
x2
−1 0 1
−1
0
1
To manipulate the sensors, we use :the set containing the grid indexes of the sensors : posS orthe matrix C ∈ RnS×nx with s i = Cy i
C has a special structure → it contains only one 1 per line(indexed by posS)→ CY contains the posS row restriction of Y
Kévin KASPER A machine learning approach for constrained sensor placement 7 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
ConstraintsSensor locations have to follow certain rules
Environment constraintsSensors cannot be setanywhere on the grid→ the search domain isrestricted.
Sensor geometry constraintsSensors cannot overlap→ a new scheme is introduced
Kévin KASPER A machine learning approach for constrained sensor placement 8 / 27
Problem formulationMotivation
ContributionFurther improvements
Learning SequenceSensorsConstraints
ConstraintsSensor locations have to follow certain rules
Environment constraintsSensors cannot be setanywhere on the grid→ the search domain isrestricted.
Sensor geometry constraintsSensors cannot overlap→ a new scheme is introduced
Kévin KASPER A machine learning approach for constrained sensor placement 8 / 27
Problem formulationMotivation
ContributionFurther improvements
Contents
1 Problem formulation
2 Motivation
3 Contribution
4 Further improvements
Kévin KASPER A machine learning approach for constrained sensor placement 9 / 27
Problem formulationMotivation
ContributionFurther improvements
In order to guarantee good recovery performances with few sensorsand in constrained cases, we cannot use methods such as :Effective Independence (EI) orFrameSense (FS)
Why? Because they rely on a reduced order basis such as UnD
UnD is the POD “basis” of order nD ∈ N∗Y → UΣV T (SVD) → UnD ∈ Rnx×nD (restriction of U)
What will be the price to pay for not using a reduced orderbasis ?
Kévin KASPER A machine learning approach for constrained sensor placement 10 / 27
Problem formulationMotivation
ContributionFurther improvements
In order to guarantee good recovery performances with few sensorsand in constrained cases, we cannot use methods such as :Effective Independence (EI) orFrameSense (FS)
Why? Because they rely on a reduced order basis such as UnD
UnD is the POD “basis” of order nD ∈ N∗Y → UΣV T (SVD) → UnD ∈ Rnx×nD (restriction of U)
What will be the price to pay for not using a reduced orderbasis ?
Kévin KASPER A machine learning approach for constrained sensor placement 10 / 27
Problem formulationMotivation
ContributionFurther improvements
In order to guarantee good recovery performances with few sensorsand in constrained cases, we cannot use methods such as :Effective Independence (EI) orFrameSense (FS)
Why? Because they rely on a reduced order basis such as UnD
UnD is the POD “basis” of order nD ∈ N∗Y → UΣV T (SVD) → UnD ∈ Rnx×nD (restriction of U)
What will be the price to pay for not using a reduced orderbasis ?
Kévin KASPER A machine learning approach for constrained sensor placement 10 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Contents
1 Problem formulation
2 Motivation
3 ContributionOptimization ProblemReducing computational costGreedy minimisation schemeResults
4 Further improvements
Kévin KASPER A machine learning approach for constrained sensor placement 11 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
The goal is to minimize ||Y − Y ||F where Y = RCY :R ∈ Rnx×nS is a linear estimatorCY is the measurement sequence
For a given C , ||Y − RCY ||F exhibits its lowest value for
R = Y (CY )+
(with the use of the Moore-Penrose Pseudo-Inverse)
Criterion
Find C ∈ argminC∈MC
||Y − Y(CY
)+︸ ︷︷ ︸
linear estimator
measurements︷︸︸︷CY ||F = εr
(C)
whereMC indicates the special structure of C (one 1 per line)
Kévin KASPER A machine learning approach for constrained sensor placement 12 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
The goal is to minimize ||Y − Y ||F where Y = RCY :R ∈ Rnx×nS is a linear estimatorCY is the measurement sequence
For a given C , ||Y − RCY ||F exhibits its lowest value for
R = Y (CY )+
(with the use of the Moore-Penrose Pseudo-Inverse)
Criterion
Find C ∈ argminC∈MC
||Y − Y(CY
)+︸ ︷︷ ︸
linear estimator
measurements︷︸︸︷CY ||F = εr
(C)
whereMC indicates the special structure of C (one 1 per line)
Kévin KASPER A machine learning approach for constrained sensor placement 12 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
The goal is to minimize ||Y − Y ||F where Y = RCY :R ∈ Rnx×nS is a linear estimatorCY is the measurement sequence
For a given C , ||Y − RCY ||F exhibits its lowest value for
R = Y (CY )+
(with the use of the Moore-Penrose Pseudo-Inverse)
Criterion
Find C ∈ argminC∈MC
||Y − Y(CY
)+︸ ︷︷ ︸
linear estimator
measurements︷︸︸︷CY ||F = εr
(C)
whereMC indicates the special structure of C (one 1 per line)
Kévin KASPER A machine learning approach for constrained sensor placement 12 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
DrawbackThe criterion is computationally costly to evaluate
To overcome this, the invariance of the Frobenius norm underunitary transformation is used.A SVD decomposition of Y of order no ∈ N∗ is computed withno � nS → Y ≈ Uno ΣnoV T
no
εr (C ) ≈ ||UnoΣnoVTno −UnoΣnoV
Tno (CY )+ CUno ΣnoV
Tno ||F
= ||Σno − ΣnoVTno (CY )+ CUno Σno ||F
→ the error matrix Y − Y of size nx × nsnap is approximated by ano square matrix → considerable speed-up
Kévin KASPER A machine learning approach for constrained sensor placement 13 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
DrawbackThe criterion is computationally costly to evaluate
To overcome this, the invariance of the Frobenius norm underunitary transformation is used.A SVD decomposition of Y of order no ∈ N∗ is computed withno � nS → Y ≈ Uno ΣnoV T
no
εr (C ) ≈ ||UnoΣnoVTno −UnoΣnoV
Tno (CY )+ CUno ΣnoV
Tno ||F
= ||Σno − ΣnoVTno (CY )+ CUno Σno ||F
→ the error matrix Y − Y of size nx × nsnap is approximated by ano square matrix → considerable speed-up
Kévin KASPER A machine learning approach for constrained sensor placement 13 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Seeing as nS � nsnap, CY is highly likely to be full row rank
(CY )+ = (CY )T(CY (CY )T
)−1
By using the previous reduced SVD
(CY )+ = Vno ΣTnoU
TnoC
T(CYY TCT
)−1
The criterion becomes :
εr (C ) ≈ ||Σno − Σno ΣTnoU
TnoC
T(CYY TCT
)−1CUno Σno ||F
→ Once and for all pre-computation of the blue terms→ considerable speed-up (C acts as a restriction operator)
Kévin KASPER A machine learning approach for constrained sensor placement 14 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Seeing as nS � nsnap, CY is highly likely to be full row rank
(CY )+ = (CY )T(CY (CY )T
)−1
By using the previous reduced SVD
(CY )+ = Vno ΣTnoU
TnoC
T(CYY TCT
)−1
The criterion becomes :
εr (C ) ≈ ||Σno − Σno ΣTnoU
TnoC
T(CYY TCT
)−1CUno Σno ||F
→ Once and for all pre-computation of the blue terms→ considerable speed-up (C acts as a restriction operator)
Kévin KASPER A machine learning approach for constrained sensor placement 14 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
→ Greedy SchemeMade viable becausefew sensors are used
InitializationSelect nS admissiblegrid points at randomto initialize the sensorconfiguration
Kévin KASPER A machine learning approach for constrained sensor placement 15 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Main loop1→ select a randomsensor (red)
Kévin KASPER A machine learning approach for constrained sensor placement 16 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Main loop2→ remove theselected sensor anddefine the new possiblesearch domain (green)
Kévin KASPER A machine learning approach for constrained sensor placement 17 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Main loop3→ replace the sensorin the location thatminimizes the criterion
Kévin KASPER A machine learning approach for constrained sensor placement 18 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
Main loop4→ select a sensor atrandom among theremaining ones andrepeat the above steps
Exit the loop whennS − 1 consecutiveupdates do not reducethe criterion
Kévin KASPER A machine learning approach for constrained sensor placement 19 / 27
Problem formulationMotivation
ContributionFurther improvements
Optimization ProblemReducing computational costGreedy minimisation schemeResults
UnconstrainedCase
ConstrainedCase
Number of Sensors
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
70
80
FSEISS
Number of Sensors
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
70
80
EISS
Kévin KASPER A machine learning approach for constrained sensor placement 20 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Contents
1 Problem formulation
2 Motivation
3 Contribution
4 Further improvementsGoal-OrientedNoisy measurements
Kévin KASPER A machine learning approach for constrained sensor placement 21 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Goal-OrientedHow to place sensors that retrieve data from Y in order to recover
H ∈ Rnh×nsnap , nh ∈ N∗ ?CY → H
Example : Recover the velocity field from pressure measurementsonly
εr (C ) = ||H − H (CY )+ CY ||F
+ computational cost reduction through the reduced order SVD ofH
Kévin KASPER A machine learning approach for constrained sensor placement 22 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Goal-OrientedHow to place sensors that retrieve data from Y in order to recover
H ∈ Rnh×nsnap , nh ∈ N∗ ?CY → H
Example : Recover the velocity field from pressure measurementsonly
εr (C ) = ||H − H (CY )+ CY ||F
+ computational cost reduction through the reduced order SVD ofH
Kévin KASPER A machine learning approach for constrained sensor placement 22 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Noise RobustnessHow to place sensors that allow for noise robust estimation ?
The sensors deliver : s = Cy + ξ with ξ ∈ RnS
Each sensor is plagued with additive white noise assumedstationary, statistically independent and drawn from the sameprobability distribution.
εr (C ) = ||Y − Y (CYξ)+ CYξ||F
where Yξ = Y + Ξ and Ξ ∈ Rnx×nsnap contains elements drawnfrom above distribution (generating sensor noise).
Redundancy can be introduced to further increase robustness :s i + ξi ,1 → y i , s i + ξi ,2 → y i , s i + ξi ,3 → y iduplicate snapshots can be added in Y (Ξ is generated accordingly)
Kévin KASPER A machine learning approach for constrained sensor placement 23 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Noise RobustnessHow to place sensors that allow for noise robust estimation ?
The sensors deliver : s = Cy + ξ with ξ ∈ RnS
Each sensor is plagued with additive white noise assumedstationary, statistically independent and drawn from the sameprobability distribution.
εr (C ) = ||Y − Y (CYξ)+ CYξ||F
where Yξ = Y + Ξ and Ξ ∈ Rnx×nsnap contains elements drawnfrom above distribution (generating sensor noise).
Redundancy can be introduced to further increase robustness :s i + ξi ,1 → y i , s i + ξi ,2 → y i , s i + ξi ,3 → y iduplicate snapshots can be added in Y (Ξ is generated accordingly)
Kévin KASPER A machine learning approach for constrained sensor placement 23 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Two noise distributions are considered for the training and the testsequence →both zero-mean Gaussian but with different variance
Training NoiseUsed by thealgorithm→ Variance σ2
n
Testing NoiseUsed to testthe locations→ Variance σ2
b
Noise Variance σ2b
ErrorExpectationE[ǫ
r](%
)
0 0.02 0.05 0.1 0.220
40
60
80
100
120
140
160
180
EISS (σ2
n= 1)
SS (σ2n= 0.5)
SS (σ2n= 0.05)
SS (σ2n= 0.01)
Kévin KASPER A machine learning approach for constrained sensor placement 24 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Two noise distributions are considered for the training and the testsequence →both zero-mean Gaussian but with different variance
Training NoiseUsed by thealgorithm→ Variance σ2
n
Testing NoiseUsed to testthe locations→ Variance σ2
b Noise Variance σ2b
ErrorExpectationE[ǫ
r](%
)
0 0.02 0.05 0.1 0.220
40
60
80
100
120
140
160
180
EISS (σ2
n= 1)
SS (σ2n= 0.5)
SS (σ2n= 0.05)
SS (σ2n= 0.01)
Kévin KASPER A machine learning approach for constrained sensor placement 24 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Thank you for your [email protected]
Kévin KASPER A machine learning approach for constrained sensor placement 25 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited by using nD = nS modes.
This originates from the neccessity of having a well-posed problem.Let ynD
be the nD POD approximation of y , a snapshot.With snD = CynD
, we have s = snD + δs.
Sensor placement methodsrely on measuring snD
(implicitly though theFisher Information Matrix).For “low” values of nD ,||δs||2 can be problematicto correctly estimate y .
nD
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
Kévin KASPER A machine learning approach for constrained sensor placement 26 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited by using nD = nS modes.
This originates from the neccessity of having a well-posed problem.Let ynD
be the nD POD approximation of y , a snapshot.With snD = CynD
, we have s = snD + δs.
Sensor placement methodsrely on measuring snD
(implicitly though theFisher Information Matrix).For “low” values of nD ,||δs||2 can be problematicto correctly estimate y .
nD
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
Kévin KASPER A machine learning approach for constrained sensor placement 26 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited by using nD = nS modes.
This originates from the neccessity of having a well-posed problem.Let ynD
be the nD POD approximation of y , a snapshot.With snD = CynD
, we have s = snD + δs.
Sensor placement methodsrely on measuring snD
(implicitly though theFisher Information Matrix).For “low” values of nD ,||δs||2 can be problematicto correctly estimate y .
nD
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
Kévin KASPER A machine learning approach for constrained sensor placement 26 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited by using nD = nS modes.
This originates from the neccessity of having a well-posed problem.Let ynD
be the nD POD approximation of y , a snapshot.With snD = CynD
, we have s = snD + δs.
Sensor placement methodsrely on measuring snD
(implicitly though theFisher Information Matrix).For “low” values of nD ,||δs||2 can be problematicto correctly estimate y .
nD
RelativeErrorǫr(in%)
1 2 3 4 50
10
20
30
40
50
60
Kévin KASPER A machine learning approach for constrained sensor placement 26 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27
Problem formulationMotivation
ContributionFurther improvements
Goal-OrientedNoisy measurements
Because most methods rely on UnD , the sensor placement andsubsequent estimation are limited when sensors are constrained.
By relying on UnD that is optimal in the `2 sense, such methodsimplicitly rely on being able to set the sensors freely to achieve thelowest estimation error.
To deal with prohibited sensorlocations, most methodsrestrict UnD
→ UnD is no longer optimal
To prevent overlapping sensors,most methods require are-work of their minimizationprocedure. Such overlapping iscaused by UnD
→ UnD is no longer optimal
Kévin KASPER A machine learning approach for constrained sensor placement 27 / 27