• No results found

Support Vectors, Duals, and the Kernel Trick

N/A
N/A
Protected

Academic year: 2021

Share "Support Vectors, Duals, and the Kernel Trick"

Copied!
100
0
0

Loading.... (view fulltext now)

Full text

(1)

1

Support Vectors, Duals, and the Kernel Trick

Machine Learning Fall 2017

Supervised Learning: The Setup

1

Machine Learning Spring 2018

The slides are partly from Vivek Srikumar

(2)

Support vector machines

• Training by maximizing margin

• The SVM objective

• Solving the SVM optimization problem

• Support vectors, duals and kernels

2

(3)

This lecture

1. Dual forms, and support vectors 2. Kernels & kernel trick

3. Properties of kernels 4. Nonlinear SVM

3

(4)

This lecture

1. Dual forms, and support vectors

2. Kernels & kernel trick 3. Properties of kernels 4. Nonlinear SVM

4

(5)

So far we have seen

• Support vector machines

• Hinge loss and optimizing the regularized loss

More broadly, different algorithms for learning linear classifiers

5

(6)

So far we have seen

• Support vector machines

• Hinge loss and optimizing the regularized loss

More broadly, different algorithms for learning linear classifiers

What about non-linear models?

6

(7)

One way to learn non-linear models

Explicitly introduce non-linearity into the feature space

7

If the true separator is quadratic

(8)

One way to learn non-linear models

Explicitly introduce non-linearity into the feature space

8

If the true separator is quadratic Transform all input points as

(9)

One way to learn non-linear models

Explicitly introduce non-linearity into the feature space

9

If the true separator is quadratic Transform all input points as

Now, we can try to find a weight vector in this higher dimensional space That is, predict using wT!(x1, x2) + b≥ 0

(10)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

10

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5

(11)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

11

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5

Lagrange form

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w0>w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwxi ) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

5

(12)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

12

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5

Lagrange form

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w0>w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwxi ) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

5

Equivalent formulation, removing constraints on x

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

ix

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

min

x

max L(x, ) = f (x) +

X

m i=1

i

g

i

(x) s.t.

1

0, . . . ,

m

0

5

(13)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

13

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5

Lagrange form

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w0>w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwxi ) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

5

Equivalent formulation, removing constraints on x

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

ix

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

min

x

max L(x, ) = f (x) +

X

m i=1

i

g

i

(x) s.t.

1

0, . . . ,

m

0

min

x

max

0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

(2)

5

(14)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

14

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5 Why equivalent?

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w>0 w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwix) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

minx max L(x, ) = f (x) +

Xm i=1

igi(x) s.t. 1 0, . . . , m 0

minx max

0 L(x, ) = f (x) + Xm

i=1

igi(x)

(2)

max0 f (x) +

Xm i=1

igi(x) =

⇢f (x) if x satisfy all the constraints 1 otherwise

5

(15)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

15

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5 Why equivalent?

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w>0 w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwix) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

minx max L(x, ) = f (x) +

Xm i=1

igi(x) s.t. 1 0, . . . , m 0

minx max

0 L(x, ) = f (x) + Xm

i=1

igi(x)

(2)

max0 f (x) +

Xm i=1

igi(x) =

⇢f (x) if x satisfy all the constraints 1 otherwise

5

1

= . . . =

m

= 0

<latexit sha1_base64="KwPfTpSGDOCgPJNEw2zpRNa+MCg=">AAACC3icbZDLSsNAFIYnXmu9RV26GVoEVyURQTdC0Y3LCvYCbQiTyaQdOpcwMxFK6N6Nr+LGhSJufQF3vo3TNAtt/WHg4z/ncOb8UcqoNp737aysrq1vbFa2qts7u3v77sFhR8tMYdLGkknVi5AmjArSNtQw0ksVQTxipBuNb2b17gNRmkpxbyYpCTgaCppQjIy1Qrc2YLY5RqEPr+CAxdLoAuYmt+yFbt1reIXgMvgl1EGpVuh+DWKJM06EwQxp3fe91AQ5UoZiRqbVQaZJivAYDUnfokCc6CAvbpnCE+vEMJHKPmFg4f6eyBHXesIj28mRGenF2sz8r9bPTHIZ5FSkmSECzxclGYNGwlkwMKaKYMMmFhBW1P4V4hFSCBsbX9WG4C+evAyds4bvNfy783rzuoyjAo5BDZwCH1yAJrgFLdAGGDyCZ/AK3pwn58V5dz7mrStOOXME/sj5/AHZk5j+</latexit><latexit sha1_base64="KwPfTpSGDOCgPJNEw2zpRNa+MCg=">AAACC3icbZDLSsNAFIYnXmu9RV26GVoEVyURQTdC0Y3LCvYCbQiTyaQdOpcwMxFK6N6Nr+LGhSJufQF3vo3TNAtt/WHg4z/ncOb8UcqoNp737aysrq1vbFa2qts7u3v77sFhR8tMYdLGkknVi5AmjArSNtQw0ksVQTxipBuNb2b17gNRmkpxbyYpCTgaCppQjIy1Qrc2YLY5RqEPr+CAxdLoAuYmt+yFbt1reIXgMvgl1EGpVuh+DWKJM06EwQxp3fe91AQ5UoZiRqbVQaZJivAYDUnfokCc6CAvbpnCE+vEMJHKPmFg4f6eyBHXesIj28mRGenF2sz8r9bPTHIZ5FSkmSECzxclGYNGwlkwMKaKYMMmFhBW1P4V4hFSCBsbX9WG4C+evAyds4bvNfy783rzuoyjAo5BDZwCH1yAJrgFLdAGGDyCZ/AK3pwn58V5dz7mrStOOXME/sj5/AHZk5j+</latexit><latexit sha1_base64="KwPfTpSGDOCgPJNEw2zpRNa+MCg=">AAACC3icbZDLSsNAFIYnXmu9RV26GVoEVyURQTdC0Y3LCvYCbQiTyaQdOpcwMxFK6N6Nr+LGhSJufQF3vo3TNAtt/WHg4z/ncOb8UcqoNp737aysrq1vbFa2qts7u3v77sFhR8tMYdLGkknVi5AmjArSNtQw0ksVQTxipBuNb2b17gNRmkpxbyYpCTgaCppQjIy1Qrc2YLY5RqEPr+CAxdLoAuYmt+yFbt1reIXgMvgl1EGpVuh+DWKJM06EwQxp3fe91AQ5UoZiRqbVQaZJivAYDUnfokCc6CAvbpnCE+vEMJHKPmFg4f6eyBHXesIj28mRGenF2sz8r9bPTHIZ5FSkmSECzxclGYNGwlkwMKaKYMMmFhBW1P4V4hFSCBsbX9WG4C+evAyds4bvNfy783rzuoyjAo5BDZwCH1yAJrgFLdAGGDyCZ/AK3pwn58V5dz7mrStOOXME/sj5/AHZk5j+</latexit><latexit sha1_base64="KwPfTpSGDOCgPJNEw2zpRNa+MCg=">AAACC3icbZDLSsNAFIYnXmu9RV26GVoEVyURQTdC0Y3LCvYCbQiTyaQdOpcwMxFK6N6Nr+LGhSJufQF3vo3TNAtt/WHg4z/ncOb8UcqoNp737aysrq1vbFa2qts7u3v77sFhR8tMYdLGkknVi5AmjArSNtQw0ksVQTxipBuNb2b17gNRmkpxbyYpCTgaCppQjIy1Qrc2YLY5RqEPr+CAxdLoAuYmt+yFbt1reIXgMvgl1EGpVuh+DWKJM06EwQxp3fe91AQ5UoZiRqbVQaZJivAYDUnfokCc6CAvbpnCE+vEMJHKPmFg4f6eyBHXesIj28mRGenF2sz8r9bPTHIZ5FSkmSECzxclGYNGwlkwMKaKYMMmFhBW1P4V4hFSCBsbX9WG4C+evAyds4bvNfy783rzuoyjAo5BDZwCH1yAJrgFLdAGGDyCZ/AK3pwn58V5dz7mrStOOXME/sj5/AHZk5j+</latexit>

(16)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

16

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

minw

1

2w>w

s.t.8i, yi(w>xi + b) 1

minw

1

2w>w + C X

i

i

s.t.8i, yi(w>xi + b) 1 ⇠i

i 0

minw,b

1

2w>w + C X

i

max 0, 1 yi(w>xi + b)

LHinge(y, x, w, b) = max 0, 1 y(w>x + b)

minw

1

2w>0 w0 + C X

i

max(0, 1 yiw>xi)

Jt(w) = 1

2w>0 w0 + C · N max(0, 1 yiw>xi)

rJt =

⇢[w0; 0] if max(0, 1 yiwix) = 0 [w0; 0] C · Nyixi otherwise

f (x) f (u) + rf(u)>(x u)

minx f (x)

s.t. g1(x)  0, . . . , gm(x)  0

L(x, ) = f (x) + Xm

i=1

igi(x)

minx max L(x, ) = f (x) +

Xm i=1

igi(x) s.t. 1 0, . . . , m 0

minx max

0 L(x, ) = f (x) + Xm

i=1

igi(x)

(2)

max0 f (x) +

Xm i=1

igi(x) =

⇢f (x) if x satisfy all the constraints 1 otherwise

5

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

0>

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

ix

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

min

x

max L(x, ) = f (x) +

X

m i=1

i

g

i

(x) s.t.

1

0, . . . ,

m

0

min

x

max

0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

(2)

5

Outer minimization:

Search x over constraint space To minimize f(x)

(17)

Primal and dual forms: constraint optimization

Primal: general constraint optimization problem

17

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

5

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

ix

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

min

x

max L(x, ) = f (x) +

X

m i=1

i

g

i

(x) s.t.

1

0, . . . ,

m

0

min

x

max

0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

(2)

5

Totally equivalent!

It is very trial to incorporate equality constraints! How?

(18)

Primal and dual forms: constraint optimization

Wo focus on this primal form

18

216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269

min

w

1

2 w

>

w

s.t. 8i, y

i

(w

>

x

i

+ b) 1

min

w

1

2 w

>

w + C X

i

i

s.t. 8i, y

i

(w

>

x

i

+ b) 1 ⇠

i

i

0

min

w,b

1

2 w

>

w + C X

i

max 0, 1 y

i

(w

>

x

i

+ b)

L

Hinge

(y, x, w, b) = max 0, 1 y(w

>

x + b)

min

w

1

2 w

0>

w

0

+ C X

i

max(0, 1 y

i

w

>

x

i

)

J

t

(w) = 1

2 w

>0

w

0

+ C · N max(0, 1 y

i

w

>

x

i

)

rJ

t

=

⇢ [w

0

; 0] if max(0, 1 y

i

w

xi

) = 0 [w

0

; 0] C · Ny

i

x

i

otherwise

f (x) f (u) + rf(u)

>

(x u)

min

x

f (x)

s.t. g

1

(x)  0, . . . , g

m

(x)  0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

min

x

max L(x, ) = f (x) +

X

m i=1

i

g

i

(x) s.t.

1

0, . . . ,

m

0

min

x

max

0

L(x, ) = f (x) +

X

m i=1

i

g

i

(x)

(2)

5

References

Related documents

[r]

P Big Bass Lake: There is no minimum length limit on largemouth and smallmouth bass, but bass from 14&#34; through 18&#34; may not be kept, and only one fish over 18&#34; is

FORMS AND FINISHES Sheet, strip, plate, and bar forms are shown in Table 1. Sheet finishes are shown in Table 2. Bar product conditions and finishes are shown in Table 3; Plate is

including the RIEs for breaking the isolation of elementary and secondary teacher education from the mainstream and integrating it with higher education The Schools will have Centres

Background Multidetector computed tomography (MDCT) angiography represents the standard of reference in the follow-up of patients after endovascular abdominal aortic aneurysm

In the 30 states that do not have genuine alternative certification, only 5 per- cent of the newly certified teachers chose the alternative route in 2004–05 (the most recent year

To compare parasite burdens obtained by real-time PCR assay with those of limiting dilution assay, twelve 8-fold serial dilutions of the lymph node

The services of MPN consist of the collecting service (there are four methods, i.e. the online method, the information link method, the direct method and the direct debit data