tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
is to construct the feedback policy using a known control
Lyapunov function Vclfinstead of Vin 4 This leads to
for any xRn
gxarg min
uRℓx uVclfAxBu 5
In this work and as proposed in 6 we define Vclfas the
optimal value function associated with the unconstrained LQR
problem in the sense that the input can take any value in R ie
VclfxxPxfor any xRnwhere Pis the unique real
symmetric positive definite solution to the Riccati equation
APAPAPBSBPBR1APBSQ0
which exists as long as the pair A Bis controllable
In view of the expressions of ℓandVclf 5 can be re
written as
gxarg minuRRBPBuKx2
max0 Kx6
with KRBPB1BPASthe optimal gain
for the unconstrained LQR problem This equivalence follows
from 5 and the equalities KRBPBKAPAPQ
andRBPBKBPAS
B Optimal insulin infusion
We apply the approach of Section IIA to the problem of
optimal insulin infusion of 7 where the primary objective
is to minimize the peak of the blood glucose level BGL
resulting from a food impulse while ensuring that insulin
flow remains nonnegative We reformulate the problem as
a LQR problem with a control nonnegativity constraint and
apply the associated QCLP strategy 5 The continuoustime
model of 7 was obtained from clinical trials Here we use a
sampleddata version with a sampling period T5min see
Appendix A for more details As a result the discretetime
state model is given by 1 with n4 and
A083510115000521 0
00716 0 995400021 0
00014 0 0390 1 0000 0
0008203249164423 0 9277
B16702
01431
00029
001637We have selected the weighting matrices Q S andRas in
Appendix A which leads to
K04936699881047360 1 66268
A solution to the corresponding closedloop system with 6
is depicted in Figure 1 together with the solutions using
the optimal openloop strategy given in 7 and the zero
input strategy As it can be seen in Figure 1 the control
is effective in reducing the excursion due to an impulse of
food ingested 60g compared to the openloop behavior
with zero input in black dashdotted line Our result shows
a maximum BGL excursion of 2mmolL The best feedback
policy in 7 Section 7 for the corresponding continuous
time model see Appendix A has a maximum BGL excursion
of128mmolL However the associated controller assumes
the BGL can be exactly differentiated which is unfortunately
unrealistic To make the strategy realistic the bandwidth of
these differentiators has been restricted in 7 and the results
show larger peaks 571mmolL
Fig 1 Top glucose evolution openloop with zero input in
black dashdotted line closedloop with Kin red solid line
optimal openloop of 7 in blue dashed line Bottom Insulin
flow closedloop with Kin 8
QCLP 6 is thus promising for this application However
the stability of the corresponding closedloop system is an
open question as we now explain
C Limitations of existing Lyapunovbased techniques
System 1 in closedloop with 6 is a CLS given by
xt1A1xtxtC1
A2xtxtC29
withA1ABK andA2Aand
C1xRnKx0
C2xRn Kx0clRnC1104 GENERIC COLORIZED JOURNAL VOL XX NO XX XXXX 2021
where Kis the gain associated with the optimal unconstrained
LQR problem It appears that certifying the stability of the
origin for system 9 is challenging 4 even when A1andA2
are Schur as in Section IIB
Interestingly this stability problem can be formulated in
other frameworks such as optimizationbased systems due
to 5 piecewise linear systems to which system 9 belongs
to or linear complementarity systems1LCS 20 Never
theless none of the corresponding Lyapunovbased stability
tools of the literature certifies the stability of the system
in Section IIB To elaborate a bit more on this claim we
introduce below a novel comparison between three distinct
classes of Lyapunov function candidates commonly encoun
tered in the above mentioned fields The first one is found
in the optimizationbased framework 13 21 22 and
corresponds to the value function associated to 5 ie for
anyxRn
Vxmin
uRxQx2xSuuRu
AxBuPAxBu11
The second Lyapunov function candidate is the quadratic form
inherited from the LCS literature 23 ie for any xRnx
andugxas in 5
VquadxxuX1X2
X
2X3
rllrlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmrlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlr
Xx
u 12

is to construct the feedback policy using a known control

Lyapunov function Vclfinstead of Vin 4 This leads to

for any xRn

gxarg min

uRℓx uVclfAxBu 5

In this work and as proposed in 6 we define Vclfas the

optimal value function associated with the unconstrained LQR

problem in the sense that the input can take any value in R ie

VclfxxPxfor any xRnwhere Pis the unique real

symmetric positive definite solution to the Riccati equation

APAPAPBSBPBR1APBSQ0

which exists as long as the pair A Bis controllable

In view of the expressions of ℓandVclf 5 can be re

written as

gxarg minuRRBPBuKx2

max0 Kx6

with KRBPB1BPASthe optimal gain

for the unconstrained LQR problem This equivalence follows

from 5 and the equalities KRBPBKAPAPQ

andRBPBKBPAS

B Optimal insulin infusion

We apply the approach of Section IIA to the problem of

optimal insulin infusion of 7 where the primary objective

is to minimize the peak of the blood glucose level BGL

resulting from a food impulse while ensuring that insulin

flow remains nonnegative We reformulate the problem as

a LQR problem with a control nonnegativity constraint and

apply the associated QCLP strategy 5 The continuoustime

model of 7 was obtained from clinical trials Here we use a

sampleddata version with a sampling period T5min see

Appendix A for more details As a result the discretetime

state model is given by 1 with n4 and

A083510115000521 0

00716 0 995400021 0

00014 0 0390 1 0000 0

0008203249164423 0 9277

B16702

01431

00029

001637We have selected the weighting matrices Q S andRas in

Appendix A which leads to

K04936699881047360 1 66268

A solution to the corresponding closedloop system with 6

is depicted in Figure 1 together with the solutions using

the optimal openloop strategy given in 7 and the zero

input strategy As it can be seen in Figure 1 the control

is effective in reducing the excursion due to an impulse of

food ingested 60g compared to the openloop behavior

with zero input in black dashdotted line Our result shows

a maximum BGL excursion of 2mmolL The best feedback

policy in 7 Section 7 for the corresponding continuous

time model see Appendix A has a maximum BGL excursion

of128mmolL However the associated controller assumes

the BGL can be exactly differentiated which is unfortunately

unrealistic To make the strategy realistic the bandwidth of

these differentiators has been restricted in 7 and the results

show larger peaks 571mmolL

Fig 1 Top glucose evolution openloop with zero input in

black dashdotted line closedloop with Kin red solid line

optimal openloop of 7 in blue dashed line Bottom Insulin

flow closedloop with Kin 8

QCLP 6 is thus promising for this application However

the stability of the corresponding closedloop system is an

open question as we now explain

C Limitations of existing Lyapunovbased techniques

System 1 in closedloop with 6 is a CLS given by

xt1A1xtxtC1

A2xtxtC29

withA1ABK andA2Aand

C1xRnKx0

C2xRn Kx0clRnC1104 GENERIC COLORIZED JOURNAL VOL XX NO XX XXXX 2021

where Kis the gain associated with the optimal unconstrained

LQR problem It appears that certifying the stability of the

origin for system 9 is challenging 4 even when A1andA2

are Schur as in Section IIB

Interestingly this stability problem can be formulated in

other frameworks such as optimizationbased systems due

to 5 piecewise linear systems to which system 9 belongs

to or linear complementarity systems1LCS 20 Never

theless none of the corresponding Lyapunovbased stability

tools of the literature certifies the stability of the system

in Section IIB To elaborate a bit more on this claim we

introduce below a novel comparison between three distinct

classes of Lyapunov function candidates commonly encoun

tered in the above mentioned fields The first one is found

in the optimizationbased framework 13 21 22 and

corresponds to the value function associated to 5 ie for

anyxRn

Vxmin

uRxQx2xSuuRu

AxBuPAxBu11

The second Lyapunov function candidate is the quadratic form

inherited from the LCS literature 23 ie for any xRnx

andugxas in 5

VquadxxuX1X2

X

2X3

rllrlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmrlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlmorlr

Xx

u 12