scalable control of positive systems - iaaciaac.technion.ac.il/swisr/talks/rantzer.pdf · stability...
TRANSCRIPT
![Page 1: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/1.jpg)
Scalable Control of Positive Systems
Anders Rantzer
LCCC Linnaeus Center
Lund University
Sweden
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 2: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/2.jpg)
Wind Farms Need Control
Picture from http://www.hochtief.com/hochtief_en/9164.jhtml
Most wind farms today are paid to maximize power production.Future farms will have to curtial power at contracted levels.
New control objective:Minimize fatigue loads subject to fixed total production.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 3: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/3.jpg)
Minimizing Fatigue Loads
Single turbine control:Minimize tower pressure variance subject to linearizeddynamics with measurements of pitch angle and rotor speed.
Optimal controller: uloci (t)
Wind farm control:Minimize sum of all tower pressure variances subject tofixed total production of the farm:
!mi=1 ui = 0
Optimal controller: ui(t) = uloci (t)!1m
!mj=1 u
locj (t).
[PhD thesis by Daria Madjidian, Lund University, June 2014]
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 4: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/4.jpg)
Controller Structure
C
Linear quadratic control of m identical systems and a constraint!mi=1 ui = 0 gives an optimal feedback matrix with two parts:
One is localized (diagonal).
The other has rank one (control of the average state).
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 5: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/5.jpg)
Server Farms Need Control
Picture from http://www.dawn.com/news/1017980
Single server control:Assign resources (processor speed, memory, etc.) to minimizevariance in completion time.
Server farm control:Minimize sum of all time variances with fixed total resources.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 6: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/6.jpg)
Towards a Scalable Control Theory
Process
!Controller
"
!!
""
"P C
P
PC
P
P
PC
" ""
!
!!
##$ "!!
!
!
Linear quadratic control uses O(n3) flops, O(n2) memory
Model Predictive Control requires even more
Today: Exploiting monotone/positive systems
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 7: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/7.jpg)
Towards a Scalable Control Theory
Process
!Controller
"
!!
""
"P C
P
PC
P
P
PC
" ""
!
!!
##$ "!!
!
!
Linear quadratic control uses O(n3) flops, O(n2) memory
Model Predictive Control requires even more
Today: Exploiting monotone/positive systems
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 8: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/8.jpg)
Towards a Scalable Control Theory
Process
!Controller
"
!!
""
"P C
P
PC
P
P
PC
" ""
!
!!
##$ "!!
!
!
Linear quadratic control uses O(n3) flops, O(n2) memory
Model Predictive Control requires even more
Today: Exploiting monotone/positive systems
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 9: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/9.jpg)
Outline
• Positive and Monotone Systems
# Scalable Stability Analysis
# Input-Output Performance
# Trajectory Optimization
# Combination Therapy for HIV and Cancer
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 10: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/10.jpg)
Positive systems
A linear system is called positive if the state and output remainnonnegative as long as the initial state and the inputs arenonnegative:
dx
dt= Ax + Bu y= Cx
Equivalently, A, B and C have nonnegative coefficients exceptfor the diagonal of A.
Examples:
Probabilistic models.
Economic systems.
Chemical reactions.
Ecological systems.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 11: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/11.jpg)
Positive Systems and Nonnegative Matrices
Classics:
Mathematics: Perron (1907) and Frobenius (1912)
Economics: Leontief (1936)
Books:
Nonnegative matrices: Berman and Plemmons (1979)
Dynamical Systems: Luenberger (1979)
Recent control related work:
Biology inspired theory: Angeli and Sontag (2003)
Synthesis by linear programming: Rami and Tadeo (2007)
Switched systems: Liu (2009), Fornasini and Valcher (2010)
Distributed control: Tanaka and Langbort (2010)
Robust control: Briat (2013)
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 12: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/12.jpg)
Example 1: Transportation Networks
Cloud computing / server farms
Heating and ventilation in buildings
Traffic flow dynamics
Production planning and logistics
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 13: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/13.jpg)
A Transportation Network is a Positive System
2
1
4 3
"
##$
x1x2x3x4
%
&&'=
"
##$
!1! $31 $12 0 00 !$12 ! $32 $23 0$31 $32 !$23 ! $43 $340 0 $43 !4! $34
%
&&'
"
##$
x1x2x3x4
%
&&'+
"
##$
w1w2w3w4
%
&&'
How do we select $i j to minimize the gain from w to x?
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 14: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/14.jpg)
Example 2: A vehicle formation
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 15: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/15.jpg)
Example 2: Vehicle Formations
2
1
4 3
(
))))*
))))+
x1 = !x1 + $13(x3 ! x1) +w1x2 = $21(x1 ! x2) + $23(x3 ! x2) +w2x3 = $32(x2 ! x3) + $34(x4 ! x3) +w3x4 = !4x4 + $43(x3 ! x4) +w4
How do we select $i j to minimize the gain from w to x?
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 16: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/16.jpg)
Nonlinear Monotone Systems
The system
x(t) = f (x(t),u(t)), x(0) = a
is a monotone system if its linearization is a positive system.
x(0)
x(1)
x(2)
x(3)
y(0)
y(1)
y(2)
y(3)
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 17: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/17.jpg)
Macroscopic Models of Traffic Flow
1 Partial differential equation by Lighthill/Whitham (1955),Richards (1956) based on mass-conservation:
0 =%!
%t+%
%xf (!)
where !(x, t) is traffic density in position x at time t andf (!) expresses flow as function of density.
2 Spatial discretization by Daganzo (1994).
Both models are monotone systems!
Exploited for lines: [Gomes/Horowitz/Kurzhanskiy/Varaiya/Kwon, 2008].Exploited for networks: [Lovisari/Como/Rantzer/Savla, MTNS-14].
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 18: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/18.jpg)
Outline
# Positive and Monotone Systems
• Scalable Stability Analysis
# Input-Output Performance
# Trajectory Optimization
# Combination Therapy for HIV and Cancer
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 19: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/19.jpg)
Stability of Positive systems
Suppose the matrix A has nonnegative off-diagonal elements.Then the following conditions are equivalent:
(i) The system dxdt = Ax is exponentially stable.
(ii) There is a diagonal matrix P & 0 such thatATP+ PA ' 0
(iii) There exists a vector " > 0 such that A" < 0.(The vector inequalities are elementwise.)
(iv) There exits a vector z > 0 such that AT z < 0.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 20: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/20.jpg)
Lyapunov Functions of Positive systems
Solving the three alternative inequalities gives three differentLyapunov functions:
ATP+ PA ' 0 A" < 0 AT z < 0
V (x) = xT Px V (x) = maxk(xk/"k) V (x) = zT x
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 21: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/21.jpg)
A Scalable Stability Test
x1 x2 x3 x4
Stability of x = Ax follows from existence of "k > 0 such that"
##$
a11 a12 0 a14a21 a22 a23 0
0 a32 a33 a32a41 0 a43 a44
%
&&'
, -. /
A
"
##$
"1"2"3"4
%
&&'<
"
##$
0
0
0
0
%
&&'
The first node verifies the inequality of the first row.
The second node verifies the inequality of the second row.
. . .
Verification is scalable!Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 22: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/22.jpg)
A Distributed Search for Stabilizing Gains
Suppose
"
##$
a11 ! $1 a12 0 a14a21 + $1 a22 ! $2 a23 0
0 a32 + $2 a33 a32a41 0 a43 a44
%
&&'( 0 for $1, $2 ) [0, 1].
For stabilizing gains $1, $2, find 0 < µk < "k such that
"
##$
a11 a12 0 a14a21 a22 a23 0
0 a32 a33 a32a41 0 a43 a44
%
&&'
"
##$
"1"2"3"4
%
&&'+
"
##$
!1 0
1 !10 1
0 0
%
&&'
0
µ1µ2
1
<
"
##$
0
0
0
0
%
&&'
and set $1 = µ1/"1 and $2 = µ2/"2. Every row gives a local test.
Distributed synthesis by linear programming (gradient search).
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 23: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/23.jpg)
Max-separable Lyapunov Functions
Let x = f (x) be a monotone system such that the originglobally asymptotically stable and the compact set X * Rn+ isinvariant. Then there exist strictly increasing functionsVk : R+ + R+ for k = 1, . . . ,n, such thatV (x) = max{V1(x1), . . . ,Vn(xn)} satisfies
d
dtV (x(t)) = !V (x(t))
along all trajectories in X .
[Rantzer, Rüffer, Dirr, CDC-13]
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 24: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/24.jpg)
Proof idea
t = 0
t = 1
t = 2
t = 3
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 25: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/25.jpg)
Outline
# Positive and Monotone Systems
# Scalable Stability Analysis
• Input-Output Performance
# Trajectory Optimization
# Combination Therapy for HIV and Cancer
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 26: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/26.jpg)
Performance of Positive systems
Suppose that G(s) = C(sI ! A)!1B + D where A ) Rn,n isMetzler, while B ) Rn,1+ , C ) R1,n+ and D ) R+. Define-G-. = sup# /G(i# )/. Then the following are equivalent:
(i) The matrix A is Hurwitz and -G-. < $ .
(ii) The matrix
0
A B
C D ! $
1
is Hurwitz.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 27: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/27.jpg)
Example 1: Transportation Networks
2
1
4 3
"
##$
x1x2x3x4
%
&&'=
"
##$
!1! $31 $12 0 00 !$12 ! $32 $23 0$31 $32 !$23 ! $43 $340 0 $43 !4! $34
%
&&'
"
##$
x1x2x3x4
%
&&'+
"
##$
wwww
%
&&'
How do we select $i j ) [0, 1] to minimize the gain from w to!
i xi?
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 28: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/28.jpg)
Example 1: Transportation Networks
2
1
4 3
Minimize!
i " i subject to
0 ( !"1 ! µ31 + µ12 + 1
0 ( !µ12 ! µ32 + µ23 + 1
0 ( µ31 + µ32 ! µ23 ! µ43 + µ34 + 1
0 ( !4"4 + µ43 ! µ34 + 1
and 0 0 µ i j 0 " j . Then define $i j = µ i j/" j .
Optimal solution $12 = $32 = $43 = 1 and $31 = $23 = $34 = 0.Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 29: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/29.jpg)
Example 2: Vehicle Formations
"
##$
x1x2x3x4
%
&&'=
"
##$
!1! $13 0 $13 0
$21 !$21 ! $23 $23 0
0 $32 !$32 ! $34 $340 0 $43 !4! $43
%
&&'
"
##$
x1x2x3x4
%
&&'+ Bw
2
1
4 3 2
1
4 3 2
1
4 3
B =
!
"
"
#
101011
$
%
%
&
B =
!
"
"
#
1111
$
%
%
&
B =
!
"
"
#
111010
$
%
%
&
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 30: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/30.jpg)
Outline
# Positive and Monotone Systems
# Scalable Stability Analysis
# Input-Output Performance
• Trajectory Optimization
# Combination Therapy for HIV and Cancer
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 31: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/31.jpg)
Convex-Monotone Systems
The system
x(t) = f (x(t),u(t)), x(0) = a
is a monotone system if its linearization is a positive system. Itis a convex monotone system if every row of f is also convex.
Theorem. [Rantzer/ Bernhardsson (2014)]
For a convex monotone system x = f (x,u), each component ofthe trajectory % t(a,u) is a convex function of (a,u).
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 32: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/32.jpg)
Outline
# Positive and Monotone Systems
# Scalable Stability Analysis
# Input-Output Performance
# Trajectory Optimization
• Combination Therapy for HIV and Cancer
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 33: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/33.jpg)
Combination Therapy is a Control Problem
Evolutionary dynamics:
x =
2
A!3
i
uiDi
4
x
Each state xk is the concentration of a mutant. (There can behundreds!) Each input ui is a drug dosage.
A describes the mutation dynamics without drugs, whileD1, . . . ,Dm are diagonal matrices modeling drug effects.
Determine u1, . . . ,um ( 0 with u1 + & & &+ um 0 1 such that xdecays as fast as possible!
[Hernandez-Vargas, Colaneri and Blanchini, JRNC 2011][Jonsson, Rantzer,Murray, ACC 2014]
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 34: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/34.jpg)
Optimizing Decay Rate
Stability of the matrix A!!
i uiDi + $ I is equivalent to
existence of " > 0 with
(A!3
i
uiDi + $ I)" < 0
For row k, this means
Ak" !3
i
uiDik"k + $ "k < 0
or equivalently
Ak"
"k!3
i
uiDik + $ < 0
Maximizing $ is convex optimization in (log" i,ui,$ ) !
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 35: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/35.jpg)
Using Measurements of Virus Concentrations
Evolutionary dynamics:
x(t) =
2
A!3
i
ui(t)Di
4
x(t)
Can we get faster decay using time-varying u(t) based onmeasurements of x(t) ?
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 36: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/36.jpg)
Using Measurements of Virus Concentrations
The evolutionary dynamics can be written as a convexmonotone system:
d
dtlog xk(t) =
Akx(t)
xk(t)!3
i
ui(t)Dik
Hence the decay of log xk is a convex function of the input andoptimal trajectories can be found even for large systems.
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 37: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/37.jpg)
Example
A =
"
##$
!' µ µ 0
µ !' 0 µµ 0 !' µ0 µ µ !'
%
&&'
clearance rate ' = 0.24 day!1, mutation rate µ = 10!4 day!1
and replication rates for viral variants and therapies as follows
Variant Therapy 1 Therapy 2 Therapy 3
Wild type (x1) D11 = 0.05 D21 = 0.10 D31 = 0.30Genotype 1 (x2) D12 = 0.25 D22 = 0.05 D32 = 0.30Genotype 2 (x3) D13 = 0.10 D23 = 0.30 D33 = 0.30HR type (x4) D14 = 0.30 D24 = 0.30 D34 = 0.15
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 38: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/38.jpg)
Example
Optimized drug doses:
0 20 40 60 80 100 120 140 160 180 2000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
time [days]
u1u2u3
Total virus population:
0 20 40 60 80 100 120 140 160 180 20010-2
10-1
100
101
102
103
time [days]
time-varyingconstant
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 39: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/39.jpg)
For Scalable Control — Use Positive Systems!
Verification and synthesis scale linearly
Distributed controllers by linear programming
No need for global information
Optimal trajectiories by convex optimization
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 40: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/40.jpg)
Many Research Challenges Remain
Optimal Dynamic Controllers in Positive Systems
Analyze Trade-off Between Performance and Scalability
Distributed Controllers for Nonlinear Monotone Systems
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems
![Page 41: Scalable Control of Positive Systems - IAACiaac.technion.ac.il/SwIsr/talks/Rantzer.pdf · Stability of Positive systems Suppose the matrix A has nonnegative off-diagonal elements](https://reader033.vdocuments.site/reader033/viewer/2022060521/604fab62fd8f093acf6fc531/html5/thumbnails/41.jpg)
Thanks!
Enrico Vanessa Daria MartinaLovisari Jonsson Madjidian Maggio
Alessandro Bo FredrikPapadopoulos Bernhardsson Magnusson
Anders Rantzer, LCCC Linnaeus center Scalable Control of Positive Systems