↳estrang it

o-
se↳and wit it
3#
in DP solves the problem at one

- +Fi To dat Faith i "- in more

cals
03
6
MPC Solves short receding horizon MPC

"-
.-
2

o DP delivers law
F -Est
↳ control
· fT 8

·
T
i
a I MPC delivers control State

or -
for
=
-

find
'
O O
,2 oc

n
·

u
·

e function of X(i)
- 7
i li =
.T
Guarantes Resumine F
O
M 0 =

u(i) which is a * T
4 5
b 1
get
= =

by sal this we /
Ju
-
- ,
⑤

-I
& N
= 0
M > 0 7
-
COSTENC * =
4 additional constrainte
-
M
=
Mi -
> n(x)
- ·

CogEelYeach, 0"O
I
-

STABILITY Xn (I Step
EX0)tuld
X(I)
l
to E

White
X
* pred
-

* Xo u(0)
+ 0
2x(0)
=
2x + U -x()
=

L
X X

.ExVE EUI-5[U15
= +
,

O
= =
&
=
Herminal constraint
Ex-10XX 103
-
N+
e [x ,, v(k)
(x(k) x (0) =
-
U(0)
Xeq = O is table for all E> 0 Locally ·
toT
)
, = 5* EU(0)(U(0) O
2x

Backward Reachability 53 U N) o
= -

N*
=

< - *
if the exist 8(E) such that
54x12
a
X 2 / m o re
.

3
-
.

-t o
1X(d)(1 < S(E) then
Finite
horizon N >x -

5 Linear
②
OC -
lIXIkll - E KEN
④
in
0 1234
(unconstrained enc:
+=
Ax
control X + Bu up tra

cost:Jorg*XR +UCHIP)noterminalcost
+
Ax + BU

1
x =

is
cally asymptotic stable to
JN =
Joan (X u) = ,
HlIX/kIIR+ LUCKIPR) + ElXN I
iklin1x(kll
= 0 FX(0)
CC
O

is asymplated
S stobd him 1x(k)11 = 0 Ux(0) **
&~ 01) for Pr KRK8 (A + BKal]
(ep
T
= Q + +
/A + BKo) Pa kex
Prereq ,
issitie
-

fors tability The x is RoA

Po : Pos def + symmetric
Feedeach >N (X(i) -
Find opt control u(il
.

.
=
Mi-
trol law then time invariant
>
that: Minimize JieN = K(i)x(i) -

(i) =
KaX(i) Ka -IBTP B R) BT Do A
weight matrixof u
= +
u(i) K(i)x(i) ,
↳
=

K(i) - (BTP(i 1) B R) ' BT

= + P(i 1) A
Jo-a (40) Ex
+ +
opt costtogo : =
Po Xo

-
20
Ricattimatrix emmatrices - -

-
& P(i) A TPlit)A At Pli + 1) B(BP (i +1) B R) BP (i

↳
= + Q -
+ + 1) A
d
Of frominal weight
PIN) =

othe
o
~
initialization aur system
is is stable under U(i) = KxX(i)
find from
behind bility
S Theorem :

Jen (x) EXT P(0) Xo = If (A B) stabilizable

, ,
the
Xeq0 is a symptoticall stable
um
opt. Cost End in last step of iteration under LQ Control
to go
-
-

Preparation for MPC

Lyapunace O
4

I
Excursio e Stability of Liner system via Ear
.
Now the real deal

-
theory -
O 3 NEARMPC (we have constraints)
if equilibrium is stable => a Lypnorfuc . exists
and quadratic Ax Bu
if equi of liver system in stabl I Lypa fac exists system : x* = +
If

NEexc
.

- -
How to find Lyapunov
-

Linear system
-

+
with Stable
cost :
JN =
IUCKIII +
<IX (N
x =
x x=0 angap
QRPf > O

&
Lyon candidate : VIX) =
+
x PX with P > 0
constraints X(K) EX U(k) EN X(N) EXf
o such that v har devent
:
Find , ,

Descent Condition
by
:
Goal :
stabilize O
MPL
Xeq
- - - =

V(x+ )
+↑
Px x Px x Pax
V(X) =
=
-
x -

x
+
Px = x /P -
P) x < O
such that
stability
-

=> TP-P choose p

of MPC :

ETP-P
N
set & pos def and solve
.
- :
= -

Q
-
Terminal cost
If needs to be CLF
OnXf
regative defiits
>
-
Terminal constraint Xf needs to be control invariant

I
for
if P that solves LE some >O is pos .
Off
Discretization
=>
Xeq
= 0 is a
sympt Stable &
("recursive ")
a)Sequentia X(1) from cost with system dyni
elim
.

finally apply to control system

* a
JN /Xo 1) = (X(k) u(k)) J f(x)
x += Ax + BU u= KX
, +
, ,

-
AX + BKX
e
=
Ul

lescontrantproblem largeF(z) -z (U(0) , U(1), U(2) ..

=

#B()x copy from LQ

⑦ seritive dependance for
N

* PO =
Q + KRKx + (A + BKa)
"
Pa (A +
BKa)
b) parallel ("full discretization")

imma al- losfi

= R(X(k)
"
LE : (A + BK) P(A + BK) P =-
JN (Xo , , 1) j(X(M)
-

,
u(k)) +
c) Compromise Multiple shooting
:

KRK P P shooting times -

Q -
↓
Ricca
eg. +
=

- simple
->

zi
= >
-

erskatinga or onee
=
④ ,

⑦ easier for state

constraints z =
1
⑦larger NLP problem
① many
constraints
 with pe k salute
opt Algorithm
standinitialize
.

warm
.

3 - Lo
LQ --

=- See
B
-

Recursive Feasability

·
MPC

i↑
(U(0, Ull) )
*
- - [u(k)[l find
- N = 2 ,

f=&
*

N2-0,
Xo 10 <X (N) < 10 -10 -> X2/N) < 103
Vets
= -

Vis
=&
, ,

e
Pf S
/: ie s
+ =>
+ 110)
A = -1 , B = Q= 1 , R
+U
1 = X1 x2(0) y
-
=
2x
r
x
=
hard constraint
-

= =
x

- dining
,
, no optimization
-
=
rand X(o) given MnEU((23

I
=
s+ X(2) = X(1) + U(l) = D
, ,

K(k) =
-
(BTP(k 1) + B + ...

u= K(N)X(y) F 2]x
u(k) = k(i)x(i) ,
=

> U
*
D- X(1) Viz (D- 2
x (11)

is
-
= =

P(K) =
ATP(K + JA ...
-

N 1
-i
-

j =
2(x(k) Q x (k) + u(k) Ru(k)) +x (N) PX(N) +
X1 E ro-2 EV * 2 + e not had constraint
-102 X 1 < 10 ; < 2x X2 [
113 =min
Us
=
/
PIN) Pf
-

= = S K= 0

/UEN
Find cost expression 5 3
x = x > exist such that f(x, n) = X:X+ -
after substitution
-

mi E /D-X11(12+ U(0123 S t
.
X (1) = X (0) + U(0)
u(0)
K(k) =
PIK =
t
- J(X(0) U (0) ,U (D) xTol X 10) UCOITR U10) (1) TX(K) UCITR(l) P &(D X(0) U(0) )
2

>3
&
= + + U10
+ X
ut
- -

+
,

101

plemen
X, XI/OR

-Err=/P-0-(d)
=
+ .. - .

For 2 B+2

=
X 10 to a
PII) find
+ 2U(0)
4)
= -

st(x)k
6x1101 *
+2
x2 (0)" + 1 / U 012+ 104(1) 2 + = x(k) + u(k)k = 0 -
1

KIO) P(0) = B
25
....

=
ED
>
-
. .. .

U(0) LO
4x , (d) 2X 2(0) 0
) U* (0)
-

=
(N) D

②> &
224(0) + x
-

-
=
= =

added constraint minimizer

-
convert to
Step 1: f(2, φ) = c21 = 15, f(3, φ) = c31 = 8, f(4, φ) = c41 = 16, - - contraint

I
Step 2: D
Take bared Mo X+
Design rotut
= 0 5X + H
f(2, {3}) = c23 + f(3, φ) = 14 + 8 = 22 L (D
.

DP d (P-K(0)2 + U(ok U(0))

u =
+ M, -

f(2, {4}) = c24 + f(4, φ) = 11 + 16 = 27 A=0 5 B . = 1

f(3, {2}) = c32 + f(2, φ) = 15 + 15 = 30 z(k 1) + = 0 5Z(k) +

.
V(k)
f(3, {4}) = c34 + f(4, φ) = 18 + 16 = 34
b No pre stabilizatio needed x = 0 5<
f(4, {2}) = c42 + f(2, φ) = 10 + 15 = 25
.

case 1 :
f(4, {3}) = c43 + f(3, φ) = 17 + 8 = 25 Tube S(2) :
-

Step 3: e(2) = ((l) + 0 5W(K) .

Casez :

f(2, {3, 4}) = min{c23 + f(3, {4}), c24 + f(4, {3})} = min{14 + 34, 11 + 25} = 36 D-UID = 0 M , 70
f(3, {2, 4}) = min{c32 + f(2, {4}), c34 + f(4, {2})} = min{15 + 27, 18 + 25} = 42
S(2) =
[e(2) (w(d) , WIllEWB W E 11]
= ,

ethadoneigthedelier ED30 feasibl

f(4, {2, 3}) = min{c42 + f(2, {3}), c43 + f(3, {2})} = min{10 + 22, 17 + 30} = 32 S(2)) = + 5 AD U(0) = D
Mi
e
=
.

Step 4: It
implemente
=> n
↳ car bes *
D- x()
-

f(1, {2, 3, 4}) = min{c12 + f(2, {3, 4}), c13 + f(3, {2, 4}), c14 + f(4, {2, 3})} =min{48, 53, 41} 52) = 15
. = D (1) = =
D

L( Xo)=o computethe
Recima for worst can tube
- = ((k
1) + 0 3 W(k 2) -
. + 0 .
52w(k 3) -
+0 . 37 w(k - 2 ) + ...

(1 + Y2
optimal input Sequence [u(0) , l e
114 + 1/8 + ) 2

I (
=> + =
...

A X += ax + bu = = ((k) + U(k)2) + X /N) 2 =

K= 0 =>
-
2 and +2
,
④ N 2 => Find x(I) and
= (2) by system dyn With brown
X+=2x +U NEWTON
.

1
J 6U10)2 + 4U(0) UI) + 24/12 20U (0)
and (1) [Her seg X 10) , U(l) U10) inside placed OUll) + 21
-
=
X 10) U (0)
- -

NE(X(k)
, , X(0) = v
1
* b = -1 N = 2
find 5(X(0), U(0) U())
5= (X (N) -
x, 12 + -

x + u(k)2
,
X
Set z = (U(d) , U(l) ,:. )
, ,
mid
see N 2 , 5 (U (0) , U(l) F(z) zT Hz <Tz + d HYzzF
(1) bul (x10)
=
+
* =

It At heropla =>
*
U 10) u
WitXIOE) H [b =] YzF = 2z'H cl
+
us , 2
5 (alo) u/l) 6u10) + hu(0) Ull) + 2 V (1)2 + 20 x (0)0(0) + OX10)U(1) + 2(X (02
= (= [ 208] d 2)
[5]X(0)
= =
,

=> n =
20010) 8U(11 42x (0) + 21

[Nos] [12U(0) - 4U
- -

J(K) UT Hu + <(x(0)) +u + 0(X(0)

-

--
=
-T = =

b
I Kostrad law
-1
*a
o
2 =
1 = (U / 0) , u / /]T L /I/U(0)
=
, ,

u=nvo
Find * Ud
=> >
-
2 u lar ver saders P2C [Y4] H = [i] substituted
227 s
=

0 2x-U ↳
>
-
>
-
H
=

U = 2 x
,
022X < 1 => OEXE S
>
T
xoskvar 10
d(0) =
40 =
[i]
-
BuJ(y(*) /
K =1

-J(k() [6] exit

8) (x(0) 1)
H =

((x10)) [ 20
-

-
=
d() [8]
2])0[UN
, +
J(Ull)
52[i" "2] [ 28]
H
) [1 , 03
=

X (N)
x1 EX(N 1)
+ =
>
-

=> = -
-

[0, = =

6(X 10) 4(2) u+ d

[]
*
-

2) (0)2 42 X (6) 2) 4 =

-- = x -
+
= =

2 inputs /Newton w
with

I
KKT
opt (10) , U(l) For X(0) = 0
,
U1o) <I
example)
- 0 =] - same - e
Litera
enough

51
64(0)2 + 4U(0) U(l) +24(1)2-20U(0) 8Ul1) + 21 U(d - L
-

,
,

L =
64(02 + 44(a) U(l) + 24(1)2-204(0) 84(1) -
+ 21 + M(U (0) -

11)
↳ = 12U(0) + ... = 0
8 (10)

2 = 44( +... - = 0

24 y

care 1: M = 0 4(0) - 10

UID)***
,

u(o =
=2 ,
Ull) =
5 >
- netfearled
*

Case 2 : >0 UI0) -

1 = 0
-Ull)
-
-

U 10) =
1 U(1) = 1 N =
4 >
- femileh
equations.
Value iteration: slower convergence, correct solution “propagates” gradually from end
to beginning. Value function of current iteration is obtained from value function and
optimal policy of previous iteration, nodes can be solved in random order (even in
parallel).
① Q

③ ④