Upload
others
View
10
Download
0
Embed Size (px)
Citation preview
Optimization (Introduction)
Goal: Find the minimizer !∗that minimizes the objective (cost) function " ! :ℛ" → ℛ
Optimization
Unconstrained Optimization
ID f(x) : IR → IR
NE FCI) : IR"→ 112
-
fatal*
/H¥nfGY or x: arg yi±yf-
Goal: Find the minimizer !∗that minimizes the objective (cost) function " ! :ℛ" → ℛ
Optimization
Constrained Optimizationf- (x*) = min fCx)
X
s .t . hi G) ⇐ o →equalitygig:# so → inequality(in
Unconstrained Optimization• What if we are looking for a maximizer !∗?
" !∗ = max# " ! t÷f(x*) = main (-fan )
Calculus problem: maximize the rectangle area subject to perimeter constraint
max! ∈ ℛ!
✓area
-Max Area
① - -
perimeter
→perimeterconstraint
②ditto③dzflo }
di.dz#Lwithoutpericoust)-s/di--dz=l0-sA=IooT-
'$
'$
'%
'%
()*+ = '$'%
-*)./*0*) = 2('$ + '%)
Area [mata -100
perimeter t.a.pe (violates"
-f#¥Lr.D D
sizes A
¥"
pig ¥.dz):÷÷÷ILi€¥⇐:*:i÷÷÷.* i;
(d, .dz) d' dies
What is the optimal solution? (1D)
(First-order) Necessary condition
(Second-order) Sufficient condition
! "∗ = min"! "IT ,max-
•
gifted, pointsf"txt) > o → x
* is minimum
f-"(x*) co - X
* is maximum
Types of optimization problems
Gradient-free methods
Gradient (first-derivative) methods
Evaluate ! " , !′ " , !′′ "
Second-derivative methods
! "∗ = min" ! "
Evaluate ! "
Evaluate ! " , !′ "
5: nonlinear, continuous and smooth
If
•
off
Does the solution exists? Local or global solution? "°*^#
a IEEE
Example (1D)
-6 -4 -2 2 4 6
-200
-100
100
Consider the function " ! = 7!8 −
7"9 − 11 -
: + 40-. Find the stationary point and check the sufficient condition
miufcx)X
* lstarder necessary conditionmax: :
'
If-'(x) .-4¥ - 3¥ - 22×+40 ; min
f)G) = o ⇒ x3 - E - 22×+40=0 !
solutions ⇒ x =/-25 !* min
4
* 2nd order condition :
'''
as .ae#-aftIfa7::ii:i:.::.oIEhin ,f-" f-5) =3 (25) +10-2220
(Min)
What is the optimal solution? (ND)
(First-order) Necessary condition
(Second-order) Sufficient condition
1D: !## " > 0
1D: !′ " = 0
! *∗ = min$! * f- (x)
HI)= A
NI : If(E) = Q - gives stationary solution×*
NI i t ) is positive definite → x*is
minimizer
Taking derivatives…f : IR"→ R
f-(Xn) = f- (Xi , Xz , - - -,Xn)
¥.÷ . - - - :# ⇒ I:÷÷÷lgradient of
f CnxD
¥. If z
µ ,±,Ein Eas - - - - Tian Hi .-¥,
:*. :*. :#ox.- - -
- o¥oxn a ¥÷⇒KYE' o¥a.. ¥¥.
--
- . g¥)nth
From linear algebra:
A symmetric ; ×; matrix = is positive definite if >&= > > @ for any > ≠ @
A symmetric ; ×; matrix = is positive semi-definite if >&= > ≥ @ for any > ≠ @
A symmetric ; ×; matrix = is negative definite if >&= > < @ for any > ≠ @
A symmetric ; ×; matrix = is negative semi-definite if >&= > ≤ @ for any > ≠ @
A symmetric ; ×; matrix = that is not negative semi-definite and not positive semi-definite is called indefinite
I.i ta- O
y . HyGEo II's. s
e::÷.I :.
! *∗ = min$! *
First order necessary condition: +! * = ,Second order sufficient condition: - * is positive definiteHow can we find out if the Hessian is positive definite?
la. eight)
THW→ ( X, y ) → are eigenpairs of H
jyxyty -- x " yn: → "5fa¥:*.
* ki so ti →
/)yTHy >
o tty ftp.defsxt isminimizer
* Xi Lo fi ⇒ y'
Hy < o Hy → His neg def ⇒ x* is
maximizer
* Tgi,Yo }→ H is indefinite → x*is psaddetdeoint
Types of optimization problems
Gradient-free methods
Gradient (first-derivative) methods
Evaluate ! * , +! * , +%! *
Second-derivative methods
! *∗ = min$ ! *
Evaluate ! '
Evaluate ! ' , !" #
5: nonlinear, continuous and smooth
+HH
Consider the function " -E, -: = 2-E9 + 4-:: + 2-: − 24-EFind the stationary point and check the sufficient condition
Example (ND)me
E- if" -244¥:("o" : )8×2+2
1)of -- e → (GI!:{4]=fg]⇒
6×2--24 → xf=4→x,=±z8×2=-2 → Xz= - O -25
stationary points : x# III.as] x'*-- fo?⇐]
''et: Hina.fi#:.u.fttfo.-t-l::Ih::::i.n.
Optimization (1D Methods)
Optimization in 1D: Golden Section Search • Similar idea of bisection method for root finding• Needs to bracket the minimum inside an interval• Required the function to be unimodal
A function ":ℛ → ℛ is unimodal on an interval [4, 5]
ü There is a unique !∗ ∈ [4, 5] such that "(!∗) is the minimum in [4, 5]
ü For any -E, -: ∈ [4, 5] with -E < -:
§ -: < !∗⟹"(-E) > "(-:)§ -E > !∗⟹"(-E) < "(-:)
HuhfK¥,if'mI::¥:*"i÷/ / 4 I
fix.✓
*
OI- -
r
- -
r
+F
GH$ H% H'H(
5$
5%
5'
5(
+F
GH$ H% H'H(
5$
5%
5'
5(
¥7 (FITta fa
Fb Fb
• B.+**22
B A *a *2×* B
[email protected][ 4 , b ]
K£7 Propose the point asks- t.IT x
, e-atCl-E) hk
/XI Xz
HI,
Xz --at Chic
Ill- C) hk :
The 1 at the start h$⇐ (b -a)i I
1←Ehk'#£E)hI f , > f,or f, sfz
!¥%,9k Ex, , b ]
[a. xD
T.ee/-fEoeryiteration
aFI¥EIb * hkti-E.hn interval gets'
Em!i hatch. .ch#..EfET;db"I
'(I - E) hrs -
- Ehr
¥t%h* ,Cte) -- E - 12=0.6-187
I t
KI interval Caio) 12-0.618-1h. .- Cb -a)-
! I I → x, = at( t - E) ho
I,hofz-t-GD-qh.pe#qs.--cyh-/if f , sfz
: -→x*E[a,Xz]
" I 9k b=X2
µh-# xz=x, →fz - fi
a:④TTbtf+,
III.Ethan..
t¥h¥¥/( f , - FCK)
if f , > fz : →x*E Ex, , b]
¥+,¥¥¥hµ,That Ej÷¥÷nn-f=fz-
Xz= At Chu fz=f(Xz)
Golden Section Search
Golden Section Search What happens with the length of the interval after one iteration?
ℎ! = ( ℎ"
Or in general: ℎ#$! = ( ℎ#
Hence the interval gets reduced by )(for bisection method to solve nonlinear equations, (=0.5)
For recursion: ( ℎ! = (1 − () ℎ"( ( ℎ" = (1 − () ℎ"
(% = (1 − ()) = .. 012
• Derivative free method!
• Slow convergence:
limI→K|@ILE|@I
= 0.618 D = 1 (EFG@4D HIGJ@DK@GH@)
• Only one function evaluation per iteration
Golden Section Search II → the < tot
① IevaluateFG)
HE ha
-¥ -- hee ef÷=Y÷ -- Chien-
reI → T
-
-Xi cheap ,
Example
A = - to → ho = 20TT b -- to h
,= ?
he,=I ko → 0.618×20 = 12.36
Newton’s MethodUsing Taylor Expansion, we can approximate the function " with a quadratic function about -M" - ≈ " -M + "N -M (- − -M) + E
: "N′ -M (- − -M):
And we want to find the minimum of the quadratic function using the first-order necessary condition
Yeti = Xk t h
n¥near" =I
* = O →point5-
'Go) tyzf
''
Go) Cx - to)¢= ofstationary
sin:÷±÷÷÷µ÷¥¥÷7¥-
Newton’s Method• Algorithm:"3 = starting guess"456 = "4 − !′ "4 /!′′ "4
• Convergence: • Typical quadratic convergence• Local convergence (start guess close to solution)• May fail to converge, or converge to a maximum or
point of inflection
i
- ---
Newton’s Method (Graphical Representation)get A
tho)HA
IED'
•Iµ.¥i
l l l
#i→X3 Xz Xl Xo X
sequence of opt.
using quad. approx I
ExampleConsider the function " - = 4 -9 + 2 -: + 5 - + 40
If we use the initial guess -M = 2, what would be the value of - after one iteration of the Newton’s method?
-
-
x,= ?
f-'
G) = 12×2+4×+5
f-"
G) = 24 X t 4
h =-fifty -- - CMg¥£¥t# =
- ft
X , = Xoth → X
,= 2 - ¥2 -1×1--0.82697
Optimization (ND Methods)
Optimization in ND: Steepest Descent Method Given a function ! * :ℛ7 → ℛ at a point *, the function will decrease its value in the direction of steepest descent: −+! *
" -E, -: = (-E − 1)O+(-: − 1)O
What is the steepest descent direction?
min FG)x
EI-
FIFA
£Xz - - - - - - - - • ×
⇒
if"
(Xo)iX1
Steepest Descent Method
" -E, -: = (-E − 1)O+(-: − 1)OStart with initial guess:
!M = 33
Check the update:
¥2 = Ii - Of
XoKD
es I 0ItEtan .- IT, ] missed± . ft;]
-
i
Steepest Descent Method
" -E, -: = (-E − 1)O+(-: − 1)OUpdate the variable with: !ILE = !I − PIQ" !I
How far along the gradient should we go? What is the “best size” for PI?
=
I , = Io - 0=5 OfGo).-
-µ12=0.57How can we get a ?
I
i:"
÷÷÷÷÷:÷÷÷..
05¥
Steepest Descent Method Algorithm:
Initial guess: *3
Evaluate: ;4= −+! *4
Perform a line search to obtain <4 (for example, Golden Section Search)
<4 = argmin8
! *4 + < ;4
Update: *456 = *4 + <4 ;4
#of
'"
=
① ÷÷.
D Oret ask-
Line Search Xktt = Xk - Ak Tf(Xn)we want to find de St .
f- (Hett ) min f(Xn -④DfC)x he
1st order condition ¥70 → gives a)DI = ofArt, ) . Tf(Xr) = Oda OXKH
pg(XnH)•Tf¥f(Xkti ) is orthogonal tozigpoEEfrwnergefa-a.GR#
ExampleConsider minimizing the function
! "!, "" = 10("!)# − "" " + "! − 1
Given the initial guess"! = 2, ""= 2
what is the direction of the first step of gradient descent?
Miu far, , Xz)X, gXz
=@± .- I:]
a- =p if Iska -- III ]
4un±÷÷t→fi;
Newton’s Method Using Taylor Expansion, we build the approximation:
f- (Ets ) -- f t ELITE tf It④ = ICI-- tnonlinear quadratic+St order condition : II - o approx off
I⇒t⇒ → His :p.my#etric--/tfH.)s---TfGTf→ solve liusys to find
-Newton slip s
Newton’s MethodAlgorithm:Initial guess: *3
Solve:-9 *4 ;4 = −+! *4Update: *456 = *4 + ;4
Note that the Hessian is related to the curvature and therefore contains the information about how large the step should be.
,#ii=f¥any
-- -
→ solve 0¥⑤-
T G
Try this out!" -, R = 0.5-: + 2.5R:
When using the Newton’s Method to find the minimizer of this function, estimate the number of iterations it would take for convergence?
A) 1 B) 2-5 C) 5-10 D) More than 10 E) Depends on the initial guess
Newton’s Method SummaryAlgorithm:Initial guess: *3Solve:-9 *4 ;4 = −+! *4Update: *456 = *4 + ;4
About the method…• Typical quadratic convergence J• Need second derivatives L• Local convergence (start guess close to solution)• Works poorly when Hessian is nearly indefinite• Cost per iteration: >(@:)
=
-