Predicting Emergent Behavior in Cardiac Tissue: A Grand Challenge - - PowerPoint PPT Presentation

Radu Grosu SUNY at Stony Brook

Predicting Emergent Behavior in Cardiac Tissue: A Grand Challenge

Joint work with

Ezio Bartocci, Gregory Batt, Flavio H. Fenton, James Glimm, Colas Le Guernic, and Scott A. Smolka

Excitable Cells

• Generate action potentials (elec. pulses)

in response to electrical stimulation

– Examples: neurons, cardiac cells, etc.

• Local regeneration allows electric signal

propagation without damping

• Building block for electrical signaling in

brain, heart, and muscles

Excitable Cells

– Examples: neurons, cardiac cells, etc.

Neurons of a squirrel

University College London

Artificial cardiac tissue

University of Washington

Excitable Cells

• Local regeneration allows electric signal

propagation without damping

Neurons of a squirrel

University College London

Artificial cardiac tissue

University of Washington

Excitable Cells

• Building block for electrical signaling in

brain, heart, and muscles

Neurons of a squirrel

University College London

Artificial cardiac tissue

University of Washington

Single Cell Reaction: Action Potential

Membrane’s AP depends on:

• Stimulus (voltage or current):

– External / Neighboring cells

• Cell itself (excitable or not):

– State / Parameters value

time voltage

Threshold Resting potential

Schematic Action Potential

Single Cell Reaction: Action Potential

• Cell itself (excitable or not):

– State / Parameters value

time voltage

Threshold Resting potential

Schematic Action Potential

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

nonlinear

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

Tissue: Reaction / diffusion

u t  R(u)  (Du)

nonlinear

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

Tissue: Reaction / diffusion

u t  R(u)  (Du)

Behavior In time

nonlinear

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

Tissue: Reaction / diffusion

u t  R(u)  (Du)

Reaction

nonlinear

Single Cell Reaction: Action Potential

time voltage

failed initiation Threshold Resting potential

Schematic Action Potential

Tissue: Reaction / diffusion

u t  R(u)  (Du)

Diffusion

nonlinear

Emergent Behavior in Cardiac Cells

Arrhythmia afflicts more than 3 million Americans alone

EKG Surface

Simulation

Ventricular Tachycardia Normal Heart Rhythm Ventricular Fibrillation

Optical Mapping

The Grand SB-Challenge

Identify (Learn) Model Parameters Personal Optical Map Faithful Simulation Personal Optical Map Control / Change Model Parameters Control / Cure Disorder Faithful Simulation

Human Cardiac-Cell Models

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Most Detailed Ionic Model

• Latest experimental data
• Multi-affine ODE (MA law)

Human Cardiac-Cell Models

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44

• Recent experimental data
• Sigmoidal ODE (Luo-Rudi)

Less Detailed Ionic Model

Human Cardiac-Cell Models

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

• Maesoscopic behavior
• Sigmoidal ODE (Luo-Rudi)

Minimal Ionic Model

Human Cardiac-Cell Models

reaction abstraction reaction abstraction

Reaction Abstraction (SPT)

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44 reaction abstraction reaction abstraction

Enzymatic Reactions (QSSA)

E  S k1Ä k1 C k2 E  P

Reaction Abstraction (SPT)

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44 reaction abstraction reaction abstraction

Enzymatic Reactions (QSSA)

E  S k1Ä k1 C k2 E  P

Using the Law of Mass Action (MA):

d[C] / dt  k1[E][S]  (k1  k2)[C] d[P] / dt  k2[C]

Reaction Abstraction (SPT)

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44 reaction abstraction reaction abstraction

Enzymatic Reactions (QSSA)

E  S k1Ä k1 C k2 E  P

Assume E = S and express E :

[C]  k1 / (k1  k2)[E][S]

[E]  [Et]  [C]

Using the Law of Mass Action (MA):

d[C] / dt  k1[E][S]  (k1  k2)[C] d[P] / dt  k2[C]

Reaction Abstraction (SPT)

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44 reaction abstraction reaction abstraction

Enzymatic Reactions (QSSA)

E  S k1Ä k1 C k2 E  P

Assume E = S and express E :

[C]  k1 / (k1  k2)[E][S]

[E]  [Et]  [C]

Michaelis-Menten Relation (Sigmoidal):

d[P] / dt  a / (1 e(u))  a S(u,,1)

u  ln[S], a  k2Et,   ln(k1  k2) / k1

Using the Law of Mass Action (MA):

d[C] / dt  k1[E][S]  (k1  k2)[C] d[P] / dt  k2[C]

Simulation: Hardware and Dimension

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

Multi-Core NVIDIA GPU 1s – 1s 1s – ?s

Tusscher-Noble2-Panfilov-03

Variables: 17 Parameters: 44

Simulation: Hardware and Dimension

NVIDIA GPU 1s – 20s

Iyer-Mazhari-Winslow-04

Variables: 67 Parameters: 94

Simulation: Hardware and Dimension

NVIDIA GPU 1s – 40s Wrong Behavior!

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

• Param. Estim: Hardware and Dimension
• Implem. SpaceEx in CUDA
• Extend SpaceEx to MA ODE

Hardware matters:

• Work on OCF model (MRM)
• MRM still intractable
• Will talk about our approach

Dimension matters:

Orovio-Cherry-Fenton-08

Variables: 4 Parameters: 27

• Param. Estim: Hardware and Dimension
• Implem. SpaceEx in CUDA
• Extend SpaceEx to MA ODE

Hardware matters:

• Work on OCF model (MRM)
• MRM still intractable
• Will talk about our approach

Dimension matters:

? ?

Lack of Excitability: Implications

Stimulus: bottom row, every 300ms No Obstacle Obstacle of UT

Problem to Solve

• What circumstances lead to a loss of excitability?
• What parameter ranges reproduce loss of excitability?

Problem to Solve

• What parameter ranges reproduce loss of excitability?

Problem to Solve

• What parameter ranges reproduce loss of excitability?

Experimental Data Minimal Resistor Model Minimal Conductor Model Minimal Multi-Affine Model RoverGene Analysis Tool

Biological Switching

  0.5

k  16

S(u,,k)  1 1 e2k(u) H (u,)  0 u   1 u       R(u,1,2)  u  1 u  1

2  1

else 1 u  2       

Minimal Resistor Model: Voltage ODE

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

Minimal Resistor Model: Voltage ODE

Voltage Rate

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

Minimal Resistor Model: Voltage ODE

Diffusion Laplacian

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

Minimal Resistor Model: Voltage ODE

Fast input current

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

Minimal Resistor Model: Voltage ODE

Slow input current

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

Minimal Resistor Model: Voltage ODE

Slow output current

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Heaviside (step)

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Constant Resistanc e

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Piecewise Nonlinear

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Piecewise Bilinear

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Piecewise Resistanc e Sigmoidal Resistanc e

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u))

MRM: Currents Equations

J fi(u,v)  H (u,v) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w) ws / si Jso(u)  H (u,w) u / o(u)  H (u,w) / so(u)

Piecewise Nonlinear

MRM: Gates ODEs

v(u,v)  H (u,v) (v  v) / v

(u)  H (u,v)v / v 

& w(u,w)  H (u,w)(w  w) / w

(u)  H (u,w)w / w 

& s(u,s)  (S(u,us,ks)  s) / s(u) J fi  H(u v)(u v)(uu  u)v / fi

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u)) J fi(u,v)  H (u,v,0,1) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w,0,1) ws / si Jso(u)  H (u,w,0,1) u / o(u)  H (u,w,0,1) / so(u)

MRM: Gates ODEs

v(u,v)  H (u,v) (v  v) / v

(u)  H (u,v)v / v 

& w(u,w)  H (u,w)(w  w) / w

(u)  H (u,w)w / w 

& s(u,s)  (S(u,us,ks)  s) / s(u) J fi  H(u v)(u v)(uu  u)v / fi

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u)) J fi(u,v)  H (u,v,0,1) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w,0,1) ws / si Jso(u)  H (u,w,0,1) u / o(u)  H (u,w,0,1) / so(u)

Piecewise Resistance Piecewise Resistance

MRM: Gates ODEs

v(u,v)  H (u,v) (v  v) / v

(u)  H (u,v)v / v 

& w(u,w)  H (u,w)(w  w) / w

(u)  H (u,w)w / w 

& s(u,s)  (S(u,us,ks)  s) / s(u) J fi  H(u v)(u v)(uu  u)v / fi

Sigmoidal Resistance

u(u,v,w,s)  (Du)  (J fi(u,v)  Jsi(u,w,s)  Jso(u)) J fi(u,v)  H (u,v,0,1) (u v)(uu  u)v / fi Jsi(u,w,s)  H (u,w,0,1) ws / si Jso(u)  H (u,w,0,1) u / o(u)  H (u,w,0,1) / so(u)

Sigmoid

 v

(u)  H (u,o)  v1   H (u,o)  v2 

 s (u)  H (u,w)  s1  H (u,w)  s2  o(u)  H (u,o)  o1  H (u,o)  o2 w(u)

MRM: Voltage-Controlled Resistances/SSV

Piecewis e Constant

 v

(u)  H (u,o)  v1   H (u,o)  v2 

 s (u)  H (u,w)  s1  H (u,w)  s2  o(u)  H (u,o)  o1  H (u,o)  o2 w(u)  w

(u)   w1 

 ( w2

   w1  ) S(u,us,kw )

 so(u)   so1  ( so2   so1 ) S(u,us,kso) w(u)

MRM: Voltage-Controlled Resistances/SSV

Sigmoidal

 v

(u)  H (u,o)  v1   H (u,o)  v2 

 s (u)  H (u,w)  s1  H (u,w)  s2  o(u)  H (u,o)  o1  H (u,o)  o2 w(u)  w

(u)   w1 

 ( w2

   w1  ) S(u,us,kw )

 so(u)   so1  ( so2   so1 ) S(u,us,kso) w(u)

MRM: Voltage-Controlled Resistances/SSV

v(u)  H (u,o) w(u)  H (u,o) (1 u / w)  H (u,o) w

*

 so(u)  H (u,o)  o1  H (u,o)  o2

Piecewis e Constant Piecewise Linear

 v

(u)  H (u,o, v1  , v2  )

 s (u)  H (u,w, s1, s2 )  o(u)  H (u,o, o1, o2 ) w(u)

MRM: Scaled Steps and Sigmoids

Piecewis e Constant

 v

(u)  H (u,o, v1  , v2  )

 s (u)  H (u,w, s1, s2 )  o(u)  H (u,o, o1, o2 ) w(u)  w

(u)  S(u,us,kw , w1  , w2  )

 so(u)  S(u,us,kso, so1, so2 ) w(u)

MRM: Scaled Steps and Sigmoids

Sigmoidal

u  o u  v u  w u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Resistance Model (MRM)

u  o u  v u  w u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Resistance Model (MRM)

0  u  o & u  (Du)  u / o1 & v  (1 v) / v1



& w  (1 u / w  w) / w

(u)

& s  (S(u,us,ks)  s) / s1

u  o u  v u  w o  u  w & u  (Du)  u / o2 & v  v / v2



& w  (w

*  w) / w (u)

& s  (S(u,us,ks)  s) / s1 u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Resistance Model (MRM)

0  u  o & u  (Du)  u / o1 & v  (1 v) / v1



& w  (1 u / w  w) / w

(u)

& s  (S(u,us,ks)  s) / s1

u  o u  v u  w o  u  w & u  (Du)  u / o2 & v  v / v2



& w  (w

*  w) / w (u)

& s  (S(u,us,ks)  s) / s1 w  u  v & u  (Du)  ws / si 1/ so(u) & v  v / v2



& w  w / w



& s  (S(u,us,ks)  s) / s2 u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Resistance Model (MRM)

0  u  o & u  (Du)  u / o1 & v  (1 v) / v1



& w  (1 u / w  w) / w

(u)

& s  (S(u,us,ks)  s) / s1

u  o u  v u  w o  u  w & u  (Du)  u / o2 & v  v / v2



& w  (w

*  w) / w (u)

& s  (S(u,us,ks)  s) / s1 w  u  v & u  (Du)  ws / si 1/ so(u) & v  v / v2



& w  w / w



& s  (S(u,us,ks)  s) / s2 u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Resistance Model (MRM)

0  u  o & u  (Du)  u / o1 & v  (1 v) / v1



& w  (1 u / w  w) / w

(u)

& s  (S(u,us,ks)  s) / s1 v  u < us & u  (Du)  (u v )(uu  u)v / fi  ws / fi 1/ so(u) & v  v / v



& w  w / w



& s  (S(u,us,ks)  s) /,s2

Sigmoid Closure Property

Theorem: For ab > 0, scaled sigmoids are closed under the reciprocal operation:

S(u,k,,a,b)1  S(u,k,  ln(a b) 2k , 1 b , 1 a)

Sigmoid Closure Property

Proof: a b-a



S(u,k,,a,b)

b S(u,k,,a,b)1  (a  b  a 1 e2k u

 )1

S(u,k,,a,b)1  S(u,k,  ln(a b) 2k , 1 b , 1 a)

Sigmoid Reciprocal Closure

S(u,k,,a,b)1  1 a  1 a  1 b 1 e

2k(u(  ln alnb 2k ))

Proof:

S(u,k,,a,b)1

  ln(a / b) / 2k

1 a 1 b

1 a  1 b

S(u,k,,a,b)1  S(u,k,  ln(a b) 2k , 1 b , 1 a)

From Resistances to Conductances

v

0.3 u uu 1.55 us 0.9087

Removing Divisions using Sigmoid Reciprocal:

gw

  1/w   S(u,kw ,u'w ,w1 1,w2 1)

 w

  S(u,kw ,uw , w1  , w2  )

uw



0.03 u'w



0.04

From Resistances to Conductances

v

0.3 u uu 1.55 us 0.9087

Removing Divisions using Sigmoid Reciprocal:

gw

  1/w   S(u,kw ,u'w ,w1 1,w2 1)

 w

  S(u,kw ,uw , w1  , w2  )

uw



0.03 u'w



uso 0.65 u'so

 so  S(u,kso,uso, so1, so2 ) gso  1/ s0  S(u,kso,u'so, 1

so1, 1 so2 )

0.04

1.48

From Resistances to Conductances

v

0.3 u uu 1.55 us 0.9087 gw

  1/w   S(u,kw ,u'w ,w1 1,w2 1)

 w

  S(u,kw ,uw , w1  , w2  )

uw



0.03 u'w



Removing Divisions using Step Reciprocal:

uso 0.65 u'so

 so  S(u,kso,uso, so1, so2 ) gso  1/ s0  S(u,kso,u'so, 1

so1, 1 so2 )

v  H (u,o,0,1) w  H (u,o,0,1) (1 ugw) H (u,o,0,w

* )

o

0.006 gv

  1/ v   H (u,o,v1 1,v2 1)

 v

  H (u,o, v1  , v2  )

 o  H (u,o, o1, o2 ) go  1/ o  H (u,o, 1

• 1, 1
• 2 )

0.04

1.48

From Resistances to Conductances

v

0.3 u uu 1.55 us 0.9087 gw

  1/w   S(u,kw ,u'w ,w1 1,w2 1)

 w

  S(u,kw ,uw , w1  , w2  )

uw



0.03 u'w



Removing Divisions using Step Reciprocal:

uso 0.65 u'so

 so  S(u,kso,uso, so1, so2 ) gso  1/ s0  S(u,kso,u'so, 1

so1, 1 so2 )

v  H (u,o,0,1) w  H (u,o,0,1) (1 ugw) H (u,o,0,w

* )

o

0.006 gv

  1/ v   H (u,o,v1 1,v2 1)

 v

  H (u,o, v1  , v2  )

 o  H (u,o, o1, o2 ) go  1/ o  H (u,o, 1

• 1, 1
• 2 )

w 0.13  s  H (u,w, s1, s2 ) gs  1/ s  H (u,w, s1

1, s2 1)

0.04

1.48

u  o u  v u  w o  u  w & u  (Du)  u go2 & v  v gv2



& w  (w

*  w) gw  (u)

& s  (S(u,us,ks,0,1)  s) gs1 w  u  v & u  (Du)  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  (S(u,us,ks,0,1)  s) gs2 u  o  0.006 u  w  0. 13 u  v  0.3

Minimal Conductance Model (MCM)

v  u < us & u  (Du)  (u v )(uu  u)v g fi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  (S(u,us,ks,0,1)  s) gs2 0  u  o & u  (Du)  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

 (u)

& s  (S(u,us,ks,0,1)  s) gs1

Gene Regulatory Networks (GRN)

GRN canonical sigmoidal form:

ui  aij S(uk,kk

k1 nj



,k,ak,bk) 

j1 mi



biui

Gene Regulatory Networks (GRN)

GRN canonical sigmoidal form:

ui  aij S(uk,kk

k1 nj



,k,ak,bk) 

j1 mi



biui

where:

aij :

are activation / inhibition constants

bi :

are decay constants

S(..): are on / off sigmoidal functions

Gene Regulatory Networks (GRN)

GRN canonical sigmoidal form:

ui  aij S(uk,kk

k1 nj



,k,ak,bk) 

j1 mi



biui

where:

aij :

are activation / inhibition constants

bi :

are decay constants

S(..): are on / off sigmoidal functions

Note: steps and ramps are sigmoid approximations

Optimal Polygonal Approximation

Given: One nonlinear curve and desired # segments Find: Optimal polygonal approximation

Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the blue curve with 3 segments ?

Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the blue curve with 3 segments ?

Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the blue curve with 3 segments ?



Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the blue curve with 3 segments ?



Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the blue curve with 3 segments ?



Optimal Polygonal Approximation

Dynamic Programming Algorithm

• M. Salotti, Pattern Recognition Letters 22 (2001), Pag 215-221
• Complexity: O(P2)
• P: # points of the curve



Optimal Polygonal Approximation

Dynamic Programming Algorithm

• M. Salotti, Pattern Recognition Letters 22 (2001), Pag 215-221
• Complexity: O(P2)
• P: # points of the curve



Optimal Polygonal Approximation

Dynamic Programming Algorithm

• M. Salotti, Pattern Recognition Letters 22 (2001), Pag 215-221
• Complexity: O(P2)
• P: # points of the curve



Optimal Polygonal Approximation

Dynamic Programming Algorithm

• M. Salotti, Pattern Recognition Letters 22 (2001), Pag 215-221
• Complexity: O(P2)
• P: # points of the curve



Optimal Polygonal Approximation

Dynamic Programming Algorithm

• M. Salotti, Pattern Recognition Letters 22 (2001), Pag 215-221
• Complexity: O(P2)
• P: # points of the curve



Globally-Optimal Polygonal Approximation

Given: Set of nonlinear curves and desired # of segments Find: Globally optimal polygonal approximation

Globally-Optimal Polygonal Approximation

Example: What is the optimal polygonal approxima- tion of the curves below with 5 segments ?

Globally-Optimal Polygonal Approximation

Combining the two we obtain 8 segments and not 5 segments

Example: What is the optimal polygonal approxima- tion of the curves below with 5 segments ?

Globally-Optimal Polygonal Approximation

Solution: modify the OPAA to minimize the maximum error of a set of curves simultaneously. Example: What is the optimal polygonal approxima- tion of the curves below with 5 segments ?

Deriving the Piecewise Multi-Affine Model

(v  u < uu) & u  e  (u v )(uu  u)v gfi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 u v w  u  v & u  e  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 o  u  w & u  e  u go2 & v  v gv2



& w  (w

*  w) gw (u)

& s  S(u,ks,us) gs1  s gs1 0  u  o & u  e  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

(u)

& s  S(u,ks,us) gs1  s gs1 u v u w u w u o u o

Deriving the Piecewise Multi-Affine Model

(v  u < uu) & u  e  (u v )(uu  u)v gfi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 u v w  u  v & u  e  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 o  u  w & u  e  u go2 & v  v gv2



& w  (w

*  w) gw (u)

& s  S(u,ks,us) gs1  s gs1 0  u  o & u  e  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

(u)

& s  S(u,ks,us) gs1  s gs1 u v u w u w u o u o

Deriving the Piecewise Multi-Affine Model

(v  u < uu) & u  e  (u v )(uu  u)v gfi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 u v w  u  v & u  e  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  S(u,ks,us,0,1) gs2  s gs2 o  u  w & u  e  u go2 & v  v gv2



& w  (w

*  w) gw (u)

& s  S(u,ks,us) gs1  s gs1 0  u  o & u  e  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

(u)

& s  S(u,ks,us) gs1  s gs1 u v u w u w u o u o

Deriving the Piecewise Multi-Affine Model

(v  u < uu) & u  e  (u v )(uu  u)v gfi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  S(u,ks,us) gs2  s gs2 u v w  u  v & u  e  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  S(u,ks,us) gs2  s gs2 o  u  w & u  e  u go2 & v  v gv2



& w  (w

*  w) gw (u)

& s  S(u,ks,us) gs1  s gs1 0  u  o & u  e  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

(u)

& s  S(u,ks,us) gs1  s gs1 u v u w u w u o u o

Deriving the Piecewise Multi-Affine Model

(v  u < uu) & u  e  (u v )(uu  u)v gfi  ws gsi  gso(u) & v  v gv



& w  w gw



& s  S(u,ks,us) gs2  s gs2 u v w  u  v & u  e  ws gsi  gso(u) & v  v gv2



& w  w gw



& s  S(u,ks,us) gs2  s gs2 o  u  w & u  e  u go2 & v  v gv2



& w  (w

*  w) gw (u)

& s  S(u,ks,us) gs1  s gs1 0  u  o & u  e  u go1 & v  (1 v) gv1



& w  (1 u gw  w) gw

(u)

& s  S(u,ks,us) gs1  s gs1 u v u w u w u o u o

Deriving the Piecewise Multi-Affine Model

12  v < u  uu  26 & u  e  R(u,i,i1,u fii ,u fii1 )

i12 25



v g fi  ws gsi  R(u,i,i1,usoi ,usoi1 ) gso

i12 25



& v  v gv



& w  w gw



& s  ( R(u,i,i1,usi ,usi1 )

i12 25



 s) gs2 u v 8  w  u  v  12 & u  e  ws gsi  R(u,i,i1,usoi ,usoi1 ) gso

i8 11



& v  v gv2



& w  w gw



& s  ( R(u,i,i1,usi ,usi1 )

i8 11



 s) gs2 2  o  u  w  8 & u  e  u go2 & v  v gv2



& w  (w

*  w)

R(u,i,i1,uwi ,uwi1 )

i2 7



gwb & s  ( R(u,i,i1,usi ,usi1 )

i2 7



 s) gs1 0  0  u  o  2 & u  e  u go1 & v  (1 v) gv1



& w  ( (R(u,i,i1,uwi

 ,uwi1  ) i0 1



 wR(u,i,i1,uwi

 ,uwi1  )) gwa

& s  ( R(u,i,i1,usi ,usi1 )

i0 1



 s) gs1 u v u w u w u o u o

2D Comparison

Analysis Problem

• Find parameter ranges reproducing non-excitability:

– Restated as an LTL formula: G (u v)

Analysis Problem

G (u v)

• Initial region:

u [0,1] v [0.95,1] w [0.95,1] s [0,0.01]

Analysis Problem

G (u v) u [0,1] v [0.95,1] w [0.95,1] s [0,0.01]

• Uncertain parameter ranges:

go1 [1,180] go2 [0,10] gsi [0.1,100] gso [0.9,50]

Analysis Problem

G (u v) u [0,1] v [0.95,1] w [0.95,1] s [0,0.01] go1 [1,180] go2 [0,10] gsi [0.1,100] gso [0.9,50]

• Stimulus:

e  1

State Space Partition

• Hyperrectangles: 4 dimensional (uv-projection)

– Arrows: indicate the vector field

u v 1.00 0.95 0.00 0 1 2 3 7 8 12 13 25 26 11 9

Embedding Transition System TX(p)

u v 1.00 0.95 0.00 0 1 2 3 7 8 12 13 25 26 11 9

x

TX (p)

   x' iff there is a solution  and time  such that:

x' x

 (0)  x, ()  x'  t [0,]. (t) rect(x) rect(x')

 rect(x) is adjacent to rect(x')



The Discrete Abstraction TR(p)

u v 1.00 0.95 0.00 0 1 2 3 7 8 12 13 25 26 11 9

x : R( p) x' iff rect(x)  rect(x') TR(p) is the quotiont of TX(p) with respect to : R(p)

The Discrete Abstraction TR(p)

u v 1.00 0.95 0.00 0 1 2 3 7 8 12 13 25 26 11 9

x : R( p) x' iff rect(x)  rect(x') TR(p) is the quotiont of TX(p) with respect to : R(p) Theorem: p. TX(p)  TR(p)

Computing TR(p)

Theorem: If f is multi-affine then x R. f (x) cHull({f (v)| v VR})

f (v1) f (v4) f (v2) f (v3) v3 v4 v1 v2 x f (x) R

Computing TR(p)

x R. f (x) cHull({f (v)| v VR})

f (v1) f (v4) f (v2) f (v3) v3 v4 v1 v2 x f (x) R

Corollary:

1.00 0.95 0.00 0 1

2

1.00 0.95 0.00 0 1 2

Partitioning the Parameter Space

• In each vertex: affine equation in the parameters

1.00

0.95

0.00 1 2

u  1 u go1  0

go1  1/2 go1  1/1

u

Partitioning the Parameter Space

go2

go1

go1

go2

1

go2 go1

1

go1

2

go1

m1

go1

m

go2

n

go2

n1

go2

2

• Parameter space: 4 dimensional (go1/go2 projection)

– Each rectangle: a different transition system

1.00

0.95

0.00 1 2

u  1 u go1  0

go1  1/2 go1  1/1

u

Results

• Rovergene: intelligently explores the PS rectangles

go2 go1

1

166.94 180 10 7.69

gso gsi

0.1 0.9 90.18 100 50 26.95

independent linearly dependent simulation

 

Conclusions and Outlook

• First automatic parameter-range identification for CC
• Validated both in MCM and MRM
• Can be validated experimentally as for ischemia

Conclusions and Outlook

• Currently work on time-dependent properties of CC
• Extend SpaceEx with RoverGene MA-techniques

Conclusions and Outlook

• Moving towards 2D/3D parameter-range identification
• Use PS partitioning, simulation and curvature analysis

Conclusions and Outlook

• Moving towards 2D/3D parameter-range identification
• Use PS partitioning, simulation and curvature analysis

Conclusions and Outlook

• Moving towards 2D/3D parameter-range identification
• Use PS partitioning, simulation and curvature analysis

Conclusions and Outlook

• Moving towards 2D/3D parameter-range identification
• Use PS partitioning, simulation and curvature analysis

Conclusions and Outlook

• Moving towards 2D/3D parameter-range identification
• Use PS partitioning, simulation and curvature analysis

Conclusions and Outlook

• Derive the MRM from Iyer model through TS abstraction