optimization formulations for inverse problems with
TRANSCRIPT
![Page 1: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/1.jpg)
Optimization formulations for inverse problems with applications to Mobile
Sensing
Christian ClaudelUC Berkeley
Department of Electrical Engineering and Computer S ciences
Los Alamos, 03/23//2010
![Page 2: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/2.jpg)
Mobile sensing for solving societal challenges
Fusion of mobile pollution data with models, for air pollution monitoring
source: DHS/NASA
source: Seto UC Berkeley
![Page 3: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/3.jpg)
Mobile sensing for solving societal challenges
Participatory sensing for traffic flow monitoring: fusion of fixed and mobile data into traffic models
![Page 4: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/4.jpg)
Mobile sensing for solving societal challenges
Impact of desalination on water ecosystems: how can we assess it?
![Page 5: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/5.jpg)
Overview
Model
Mathematical abstraction of how the system evolves
Best knowledge given the data
Data
Estimate
Measurements with information about model parameters and state of system
![Page 6: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/6.jpg)
Overview
Model
Mathematical abstraction of how the system evolves
Best knowledge given the data
Data
Estimate
Measurements with information about model parameters and state of system
Inverse modeling
![Page 7: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/7.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 8: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/8.jpg)
Assume that any evolution trajectory of a system can berepresented as a point in a (high dimensional) vector space V
Measurement of the state at time t=3
Inverse modeling problems
VMeasurement of the state at time t=1
Measurement of the state at time t=2
State trajectory
![Page 9: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/9.jpg)
The measurement data corresponds to a set of “possible statetrajectories” D of the system (i.e. trajectories that are compatiblewith the measurement data)
D is not a single point because of:
- Measurement uncertainty- Missing measurements
Inverse modeling problems
V
D(data)
![Page 10: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/10.jpg)
The model yields a set of “possible model trajectories” M (i. e.trajectories that are compatible with the model)
M is not a single point since:
- The evolution trajectory depends upon the initial condition- The model parameters may be uncertain
Inverse modeling problems
V
M(model)
![Page 11: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/11.jpg)
Case 1: compatibility between data and model
If both sets intersect, there exist state trajectories whic h are bothcompatible with the model and the data
V
M(model)
D(data)
![Page 12: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/12.jpg)
If both sets do not intersect, there are no state trajectorie s whichare both compatible with the model and the data. Two problemsmight still be of interest
Case 2: incompatibility between data and model
[Evensen 07][Wu, Litrico, Bayen 08]
V
M(model)
D(data)
![Page 13: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/13.jpg)
Data reconciliation problem : finding the trajectory satisfying themodel closest to the data
Case 2.a: incompatibility between data and model
Solution to the “data reconciliation” problem
[Evensen 07][Wu, Litrico, Bayen 08]
V
M(model)
D(data)
![Page 14: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/14.jpg)
Data assimilation problem : finding the trajectory satisfying thedata closest to be a solution of the model
Case 2.b: incompatibility between data and model
Solution to the “data assimilation” problem
[Evensen 07][Wu, Litrico, Bayen 08]
M(model)
D(data)
V
![Page 15: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/15.jpg)
Introductory example
Consider the measurement of a battery’s voltage x over (disc rete)time. If n discrete measurements are considered, the evolut iontrajectory space V is R n
For n=3, the state trajectory is defined by (x 1,x2,x3)
x1 (voltage at time t=1)
x2 (voltage at time t=2)
x3 (voltage at time t=3)
![Page 16: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/16.jpg)
Consider the simple discrete time linear model: x k+1= xk
Introductory example
M (model)
x1 (voltage at time t=1)
x2 (voltage at time t=2)
x3 (voltage at time t=3)
![Page 17: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/17.jpg)
Volta
ge (
xi)
Case 1: compatible data
There can exist more than one solution to theproblem
Volta
ge (
x
discrete time (i)1 2 3
![Page 18: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/18.jpg)
Case 2: incompatible dataVo
ltage
(x
i)There exist no solution to the problem
1 2 3
Volta
ge (
x
discrete time (i)
![Page 19: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/19.jpg)
Volta
ge (
xi)
Case 2.a: data reconciliation
1 2 3
Volta
ge (
x
discrete time (i)
![Page 20: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/20.jpg)
Volta
ge (
xi)
Case 2.b: data assimilation
1 2 3
Volta
ge (
x
discrete time (i)
![Page 21: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/21.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 22: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/22.jpg)
Context of this work: Mobile Millennium project
Mobile Millennium project: estimating traffic conditions on highways and secondary roads using multiple sources of data:
Dedicated infrastructure:
- Transponders (FasTrak)- Speed radars- Magnetometers- Loop detectors- Traffic cameras
http://traffic.berkeley.edu
![Page 23: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/23.jpg)
Context of this work: Mobile Millennium project
Mobile Millennium project: estimating traffic conditions on highways and secondary roads using multiple sources of data:
Participatory sensing
- Phones- Taxi location information- Aftermarket devices- Fleet location information
http://traffic.berkeley.edu
![Page 24: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/24.jpg)
Context of this work: Mobile Millennium project
![Page 25: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/25.jpg)
Context of this work: Mobile Millennium project
![Page 26: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/26.jpg)
Context of this work: Mobile Millennium project
![Page 27: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/27.jpg)
Architecture of the Mobile Millennium system
Proxy ID
Server
VTL
Server
VTLs PeMS
feedArterial
Models
Highway
Data
Server
NAVSTREETS
GPS
SatellitesVTL
feed
PeMS
Radar
feed
Radar
Cell Towers
Traffic Estimation
Server
Highway
Model
PublicNetwork
ProviderNokia / NAVTEQ UC Berkeley
Smartphones
Traffic
Report
Server
NAVTEQ
Traffic
PeMS
filter
Radar
filter
Real time traffic information system: - more than 2 million data points a day (from the fi xed infrastructure)- more than 500 000 data points a day (from mobile s ensing)- more than 200 GB of data (cumulated)
today’s talk
![Page 28: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/28.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 29: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/29.jpg)
System: highway section
The state of the system is described by the Moskowi tz function M(t,x).
- We attribute consecutive labels n to the vehicles entering a stretch of highway
- The Moskowitz function is a continuous function sat isfying M(t,x)=n
Problem description: state definition
[Newell 93], [Daganzo 03,06]
2 13456
vehicle label
position
6
5 43 2 1
![Page 30: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/30.jpg)
The Moskowitz function satisfies the following Hamilton-J acobi(HJ) partial differential equation (PDE), which can be deri vedfrom the Lighthill-Whitham -Richards (LWR) PDE
Model: Hamilton-Jacobi PDE
The function ψ is a parameter of the HJ PDE known as “fluxfunction”
[Lighthill, Whitham 55], [Richards 56], [Newell 93], [Daganzo 03,06]
Flo
w ψ
(ρ) (
veh/
h)
Density ρ (veh/mile)0 10050
1000
2000
Source: PeMS
![Page 31: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/31.jpg)
Data (boundary conditions)
To solve the PDE, one usually needs
• Initial condition (satellite, camera)• Upstream boundary condition (loop detector, radar)• Downstream boundary condition (loop detector, radar )
They are not always known, but we can use in additi on:
• Internal condition 1 (cellphone, taxi, fleet, after market device)• Internal condition 2 (cellphone, taxi, fleet, after market device)
![Page 32: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/32.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 33: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/33.jpg)
Model solution procedure
Input: data (with error bounds), model parameter
STEP 1: Define the proper mathematical solution to the PDE using control theory
STEP 2: Express the boundary conditions by piecewise affine (PWA) functions
[Claudel, Bayen, IEEE TAC 09 a,b] [Claudel, Bayen, SIAM SICON 10 (in review)]
STEP 3: Express the solutions to the HJ PDE as semi-analytic functions
STEP 4: Derive the model compatibility conditions as a set of inequality constraints
STEP 5: Solve the problem numerically (LP, QP, GP)
Output: possible traffic states compatible with the model a nd the data
![Page 34: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/34.jpg)
Controlled dynamical system:
STEP 1: Mathematical solution procedure
Control theoretic definition of the solution to the HJ PDE
[Aubin 91], [Cardaliaguet, Quincampoix, Saint Pierre 99], [Aubin, Bayen, Saint-Pierre 08]
![Page 35: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/35.jpg)
Capture basin
Capture basin of a target:
Set of initial points for which there exists at least one solution of the dynamical system
STEP 1: Mathematical solution procedure
Control theoretic definition of the solution to the HJ PDE
[Aubin 91], [Cardaliaguet, Quincampoix, Saint Pierre 99], [Aubin, Bayen, Saint-Pierre 08]
Target (representing a boundary condition)
dynamical system reaching the target in finite time
![Page 36: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/36.jpg)
Tangential property:
The lower envelope of the capture basin solves the Hamilton-Jacobi PDE in the Barron -Jensen/ Frankowska
Capture basin
STEP 1: Mathematical solution procedure
Control theoretic definition of the solution to the HJ PDE
[Crandall, Evans, Lions 84], [Aubin 91], [Aubin, Bay en Saint-Pierre 08]
Barron -Jensen/ Frankowskasense
Solution to the HJ PDE
Target
![Page 37: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/37.jpg)
Data (from sensors)
Boundary conditions
Capture basin
STEP 1: Mathematical solution procedure
Control theoretic definition of the solution to the HJ PDE
Solve the HJ PDE with these boundary conditions
Solution to the HJ PDE may (or may not) satisfy all the
prescribed boundary conditions
[Crandall, Evans, Lions 84], [Aubin 91], [Aubin, Bay en Saint-Pierre 08]Solution to the HJ PDE
Target
![Page 38: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/38.jpg)
A set of boundary conditions is compatible
with the model
STEP 1: Mathematical solution procedure
Capture basin
Control theoretic definition of the solution to the HJ PDE
[Aubin 91], [Aubin, Bayen, Saint-Pierre 08], [Claude l, Bayen, IEEE TAC 09 a]
There exists a solution to the HJ PDE satisfying all the boundary conditions
Target
![Page 39: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/39.jpg)
The data generated by mobile phones or fixed detectors corre spondsto a set of piecewise affine initial, boundary or internal co nditions
posi
tion
STEP 2: express the data as PWA functions
Loop detectors
[Claudel, Bayen, ACC 10 (accepted)], [Claudel, Baye n, SIAM SICON 10 (in review)]
time
posi
tion
Cellphone
Aftermarket device
Video or satellite
![Page 40: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/40.jpg)
• Standard schemes:- Lax-Friedrichs- Dynamic programming- Level set methods
• For piecewise affine initial,boundary and internal
Example: solution the HJ PDE associated with an affine internal condition
STEP 3: develop a fast semi-analytical scheme
boundary and internalboundary conditions, thesolution to the HJ PDE isthe minimum of closed-form expression functionsAdvantages:
- Exact- Faster
[Claudel, Bayen, IEEE TAC 09 b]
![Page 41: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/41.jpg)
Example: solution the HJ PDE associated with an affine internal condition
STEP 3: develop a fast semi-analytical scheme
Solution
Model parameter
Intermediate computation
[Claudel, Bayen, IEEE TAC 09 b]
Coefficients of the affine internal condition
Coordinates
![Page 42: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/42.jpg)
Problem : how can we check quickly if a set of coefficients (ofPWA boundary conditions) satisfies the model?
STEP 4: compatibility conditions as convex inequalities
M
[Claudel Bayen, SIAM SICON 10 (in review)]
Coefficients of the PWA boundary conditions
M
![Page 43: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/43.jpg)
Checking that there exists a solution to the Hamilton-Jacobi PDE amounts to checking that a
Depends on the boundary condition coefficients
STEP 4: compatibility conditions as convex inequalities
PDE amounts to checking that a set of convex inequalities is satisfied
[Claudel, Bayen, IEEE TAC 09 a,b]
Explicit solutions to the HJ PDE (depend on boundary condition
coefficients and model parameter)
![Page 44: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/44.jpg)
The inequalities defining the compatibility with the model areconvex in terms of the parameters of the PWA conditions
M
STEP 4: compatibility conditions as convex inequalities
[Claudel Bayen, SIAM SICON 10 (in review)]
Coefficients of the PWA boundary conditions
M
![Page 45: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/45.jpg)
The set of parameters of the PWA conditions that are compatib lewith the data is also convex (with usual error models)
M D
STEP 5: implementation using convex programming
[Claudel Bayen, SIAM SICON 10 (in review)]
Coefficients of the PWA boundary conditions
![Page 46: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/46.jpg)
Data assimilation and reconciliation problems: finding th eminimum distance between M and D
STEP 5: implementation using convex programming
y
M D
Minimize ||x-y||
s.t.x satisfies the convex inequality constraints(M)y satisfies the convex inequality constraints (D)
[Claudel Bayen, SIAM SICON 10 (in review)]
x y
Coefficients of the PWA boundary conditions
![Page 47: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/47.jpg)
Bounds on PWA condition parameters (traffic parameters) x i
STEP 5: implementation using convex programming
M D
[Claudel Bayen, SIAM SICON 10 (in review)]
Minimize (or maximize) x i
s.t.x satisifies the convex inequality constraints (M) and (D)
x
Coefficients of the PWA boundary conditions
![Page 48: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/48.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 49: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/49.jpg)
The model parameter ψ is usually uncertain. Nothing guarantees that real data solves the model exactly.
Model parameter uncertainty
M(ψ)
Real value of the parameters of the PWA conditions
[Claudel, Bayen, Allerton CCC 2009]
Flo
w ψ
(ρ) (
veh/
h)
Density ρ (veh/mile)0 10050
1000
2000
Coefficients of the PWA boundary conditions
of the PWA conditions
ψ
![Page 50: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/50.jpg)
Model parameter uncertainty
M(ψ1)
Real value of the parameters of the PWA conditions
The real value of the parameters is guaranteed to satisfy theHJ PDE, for a class of parameters ψ
[Claudel, Bayen, Allerton CCC 2009]
Flo
w ψ
(ρ) (
veh/
h)
Density ρ (veh/mile)0 10050
1000
2000
Coefficients of the PWA boundary conditions
of the PWA conditions
ψcertificate
![Page 51: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/51.jpg)
Model parameter uncertainty
M(ψ1)
The real value of the parameters is guaranteed to satisfy theHJ PDE, for a class of parameters ψ
D
[Claudel, Bayen, Allerton CCC 2009]
Coefficients of the PWA boundary conditions
D
Guaranteed bounds on traffic parameters
![Page 52: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/52.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 53: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/53.jpg)
Fault detection in sensor infrastructure
PeMS loop detector network:
- 1200 sensors in the San Francisco Bay Area- poor reliability (70% availability in average)- detecting sensor failures is a big problem
[Claudel, Bayen, Allerton CCC 2009]
spee
d (m
ph)
time
550 mph!
0 mph!
![Page 54: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/54.jpg)
Fault detection in sensor infrastructure
PeMS loop detector network:
- 1200 sensors in the San Francisco Bay Area- poor reliability (70% availability in average)- detecting sensor failures is a big problem
Status: good!
spee
d (m
ph)
time[Claudel, Bayen, Allerton CCC 2009]
Status: good!
Heavy congestion reported at 2:00AM!
![Page 55: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/55.jpg)
Fault detection in sensor infrastructure
Model-based fault detection
Checking (on multiple sensors) that model , dataand sensor specifications (maximal error level)are consistent
M
[Claudel, Bayen, Allerton CCC 2009]
Coefficients of the PWA boundary conditions
![Page 56: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/56.jpg)
Minimal error certificates:
What is the minimal possible relative error of both sensors so that the data is consistent with the model
Fault detection in sensor infrastructure
Minconsistent
D (20% error)
Coefficients of the PWA boundary conditions
[Claudel, Bayen, Allerton CCC 2009]
D (0% error)
D (40% error) D (Spec error limit)
![Page 57: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/57.jpg)
Minimal error certificates:
What is the minimal possible relative error of both sensors so that the data is consistent with the model
Fault detection in sensor infrastructure
Mconsistent
D (20% error)
Coefficients of the PWA boundary conditions
[Claudel, Bayen, Allerton CCC 2009]
D (0% error)
D (40% error) D (Spec error limit)
![Page 58: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/58.jpg)
Example of sensor fault detection (actually sensor misplacement)
Fault detection in sensor infrastructure
consistent consistentinconsistent inconsistent
sensor 1 sensor 2 sensor 3 sensor 4 sensor 5
![Page 59: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/59.jpg)
Example of sensor fault detection (actually sensor misplacement)
Fault detection in sensor infrastructure
![Page 60: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/60.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 61: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/61.jpg)
Google Maps vs. HJ PDE
Setup:
February March 20 th, 20091:30PM (Friday afternoon congestion)
Google Maps Mobile MillenniumGoogle Maps Mobile Millennium
![Page 62: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/62.jpg)
Google Maps vs. HJ PDE
![Page 63: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/63.jpg)
Outline
Inverse modeling problems- Problem description- Introductory example
Inverse modeling problems involving Hamilton-Jacobi equa tions- Background- Problem definition- Optimization formulations for inverse modeling- Optimization formulations for inverse modeling- Parameter uncertainty
Sensing applications in traffic flow- Fault detection in the PeMS sensor network- Other applications: racing with Google Maps
Conclusion
![Page 64: Optimization formulations for inverse problems with](https://reader034.vdocuments.mx/reader034/viewer/2022042922/626abc7b3c827e45e64ec705/html5/thumbnails/64.jpg)
Transformation of an estimation problem involving a nonsmooth, nonlinear PDE into a convex optimization problem: e fficient way of integrating the model constraints into the estimati on problem
Can be used for different applications (not showed here), including:
Conclusion
- data assimilation- data reconciliation- security analysis- user privacy analysis