7
In elimination style algorithms data is not reused round G round This is b c of difficulties w deputerce Input X for f 1,2 Player Choong x ex Xe closer using cxs.gs Nature reveals yet 209kt tEe vN 0,1 A iid consider 15 Fe xsxstYfE xsy Ys xstO.ee o Fixx X.IT xsstlP cz E o.sse EEkIELexp hLz.E o.s e Elexplx ftp.xstjt t xses WEIR W L x xi Yxscz.o o.s I.in sites ie 4 Fuse L Ts It'm exPHlz.Iw.es un.io o sheTt EEIELiI expHsz.wsesD Cannot the IEC 3 inside product b c It depends on SES sct

xsy Fixx - courses.cs.washington.edu

  • Upload
    others

  • View
    8

  • Download
    0

Embed Size (px)

Citation preview

Page 1: xsy Fixx - courses.cs.washington.edu

In elimination style algorithms data is not reused

round G round This is b c of difficulties w deputerce

Input X

for f 1,2Player Choong x ex Xe closer using cxs.gs

Nature reveals yet 209kt tEe vN 0,1A iid

consider 15

Fe xsxstYfE xsy Ys xstO.ee

o Fixx X.ITxsstlP

cz E o.sse EEkIELexp hLz.E o.s

e Elexplx ftp.xstjtt

xsesWEIR

W L x xi Yxscz.o o.s

I.insites ie

4 Fuse LTs It'm

exPHlz.Iw.es

un.io o

sheTtEEIELiI expHsz.wsesD

Cannot the IEC 3 inside product b c It depends

on SES sct

Page 2: xsy Fixx - courses.cs.washington.edu

Xsisindepe.de ofEs

Lea Peta formalized cleaned up literature

on self normalized bonds

Fix Recall from last time E Exacs'T sys

Z G O Fx z.LI xsxstlcE 09

E kztlcz.es i iHO O.llizxsxsy9

This process Td JIMI Highrmb

self normalized because coff 0 12 Essex

Lemming Fix 8 0 Let ye LXt 07 Et

where Et N 0,1 It chosen based on

Ius Ys If GEUM Se where

Ve D Xsxst VI

St XsYs

Page 3: xsy Fixx - courses.cs.washington.edu

then

PH t cN oe o.lt VrH04htEHogTNtffIT es

Note to lNt E dlogfdfj.TLif 113411 L for all t

Now we have confidence bound for

arbitrary sequence of Meas Opens

door to new algorithms

Recall UCB from MAB MAB x eitoria

Idea For each XEX algorithm

construct confidence bond arund each X I

a d then pull arm w highest uppercut

bond Intuit If UCB is very high Her

that arm either has very large reward

OR has not been sample enough finesand sampling shrinks co 6 bond

Page 4: xsy Fixx - courses.cs.washington.edu

Input Xfor 1,2

construct confidence set Ct OaECE W.p I S htt

UCB FEE LO K

Play Xea UCB x

Nature reveal Yei LKe 07 TEE

Idea Let G 40 110 0711 1E Bt

whee 13 rrllo.lt osEogTNtffIT

Let fld to III xx't VIIGCN EEE Hall a HIT

a d t.eu YoxffH There

fill fld

Page 5: xsy Fixx - courses.cs.washington.edu

gal gun

ga Tr ftp.T.xxttrIYIZa xxtH

aeffective dimension

D it 8 0 ow d

Regret of UCB X ears Logo

Therein If KO xe let 11041cL At the

iii II ix ke O t.FILTRem

h Since 13 2 EggsD

Rfa DfTPoot Xe a E mo Lx 07 957 UCBCH

Let It _ofEEE Lae 07

Page 6: xsy Fixx - courses.cs.washington.edu

O X I UCB X E UCBCH L Ie Et

O X Xt L 09 7 Lot x

i LO Ke Ie Et

0 01 Xt

Veh O 071,4185 xt

E HO Ethan 11THlyin

Ello felly.int lOt EeHutHE Z Be

E 2 Be 1 24 4185

Recall Kke 071 El htt O Xt X Ez

O X Xt Emin 213 117411yup 2

I 213T min llxtlly.int I

Page 7: xsy Fixx - courses.cs.washington.edu

if It Lo so A

E E 2 Rt min l l l Xt llyin I

t T F 72 A min 111 heap I IIT

ftp.FImiF.llxtHvI

FREIHEITNel8 I Ne I 8 l f I t t II