[oitehe, 1997) #2
‘Poland eal, 2001
#1ig # it
113
RIT LEN ROE, ROR, IRE KLAR
HE, PB, RUE ANE. SEARS, SET MAME. Rte
SF IG, CR RB, UN A,
SOLE TA, ARNIS HIE, AO HEM REE, RT TURE
ERAT!
BBS AEM BI, HEI, BRATIKECT Ft
4S “UBF” (machine learning).
ISR BLS, RAS AAR A BA AF a A AL, fo), 2
HARIRI, ABSA. SIME, BLUR ALMERIE? AEB AE
RATE MEL LUE IS FRO, SOARES, 9B
RACH RAVE HAAR, ALARA. A, BRAT ARTE
PROVIGIL? RIEHL, REALL, POLAR, LARS ABRIL
‘APSE ITMCET LURE RAIIT. OA, BRUTAK EAM I, FL
TLR, FT Wee. ORT HSL, EARS RER. PILL, BRITE Me
THOTH, RUARNELORT HSB, MBNA, ROMER
‘RO eH PO
LIND REAL EER AZ ESSER. ESBS?
SEERA TER, EB TSO SPB, A
WAM RAAB MIERE. EHSL, “Hak” BR OL “aR” eatee
7, FBG, WLRBE SL SBT ABE a, EAT ACBL ae HE
; AU” (model) BIH, BN “424534” (learning algorithm). #7 FAS, R
ATOM MCAR SORE Es, CARE FRR Pe A ERD A
UE BIAS TE), EL A AP LI). SI
SBA UIST “HOR” (OE, BZ AOL, PTL BAUME LES
RE OBS WEE
OIE BUM” LAAN hE ASR, ATSC BL
JR (ito PRES), OH “LGR” AIRE (Ld — SB). ADIN RIE A
Ae DET
Hebe Kean — at
AEE POT pa eH
AON ELA
we
aR AR “AOS
1° (ainng instance) %
eat"
FARRAEARRE
RM, RM RPL RL
SH RA, HA
AREER.
Maw. ATR ah
mee #
12 SARE
SEU TALES, BEAT BAR. BRR T He LA, (1
MAR; Bw, MOM), (ARE; LA Mee
fa), (BRA; fe; WA Ae), BRET A —RICR,
“=” BBR “RGA”
RABRARA MAP “HHS” (data set), HPBRIARERT —
ASHE BO @ (2K Be — 4 A HE, BRI —4> “a” (instance) BR “FF
AC” (sample). RRB RN BART MIRA MOTT, fin “ee”
“HERE” “URGE” FR “RHE” (attribute) a8 “HFG” (Feature); BUH FAO
4H, Bln “FER” “GR”, HH “YEH” (attribute value). MITE MA
HK “IRELRTEY” (attribute space), “AEAR%ET8)” (sample space) sR “8A
318)” . PIUERATHE “BR” “RAR” “RoE” ES, METER
AEF GTS Hee, A AB AY PEE a PH A A
EATS PME AM — TEAR, BRINE Tao A—T
“#4 MRE” (feature vector)
TAR, & D = far, 22,...,0m} RRA m PABBA,
AN Td 7 AE A (AT NE), I A a
wy = (wins Tiny. tia) Hed HEPES A PAN, wy YX, SE os BE
22 TEE j NYE EAE Olin ER 8 NUTELLA ELE“
HE” ), d MAFEA a; (8) “#898” (dimensionality).
DSB p18 BOM Bt ERK “2% 3)” (learning)ak “ Wll¥E” (training),
TTR ET Be. RL ET I I “UNE
BEB” (training data), HP MEAKA—‘ “UIGEFEA” (training sample),
WARAA RRA A “USE” (training set). ABA NY TR TBE
PAE AERTARE, FILA RE “ABER” (hypothesis); RE CEME AE A, WEE
” 80” Be RSE” (ground-truth), #7) SAEMLAR WY T eH REA. A
PARRY “AF 2]88” (learner), AEE A RES MEM SME
fa) A SBE.
RABE AEM RATAN LATER AR ETL” BBE, 1
ANTE TR UAE ER LA. ERE “HLM” (prediction) My
BM, RNBRGUARAIN “HR” iE, FI “(RWS REM,
ht 4), AER)”. BRP aR RIE, oN I”, HON he
12” (label); WH T tricia Bia, MAK “PEGI” (example). — Ati, FH 12 eke
3
BEARD
tea HE A
ahaa) “HAS
ae RET
ae MRR OS
(testing instance) i “H
so"
Bae ELA
Tika a eA Wh
H,RL6F
Ae SIE I"
sabe"
mi, BAAR
FO (unseen instance)
ARERR SAH
SURAT AK (Hb 20
Bi, BABA 10 AT
FORGE, ARE
Re 10),
(eis) SRB 1 PEE, Je ys ©) IR a HIG, DRA IRIC A,
DSK “#RA751)” (label space) ak “ SMythi 314)”
PRAM A, a RE RL”, EES
“4b26” (classification); #4 FAM AE REE (A, #7) Wn 75 TL LIA BE 0.95. 0.37,
de eS ES BK “IIIA” (wegression). ft AW Bw + HY a
2A” (binary clasfication){E%, ALAWEIEP 7A “IER” (positive class),
A-*KA “KR” (negative class); SRS PAR MH, WMA “SY
28” (multi-class classification) f£ 9. — fs, THM (ES J Ah Bal Bt at
HE {(w1,y1), (w2,y2)s-++5 (ms Yon) ) BATA, BIL — AMA BEI] A BL
ela] Y ARM f XY. MORE, HBS Y = {1,41} M {0,1};
SI RAES, |Y| > 2; MEVAES, Y =R, R WEB.
PARAS, ERITREA “MGR” (resting), SBMA
HEA “BARFEAR” (testing sample). PUMTEFE f i, MWR o, TRI
wie y= (2).
PUNE AT DLT A “FRE” (clustering), BEWARE PAYAL Sy RAE
Al, SEAR ANAS “ABE” (cluster); SK FFE APIA EY Ai — see HES
Aalst, Pin “eM “ET”, BEE AT “oh”. ROR RUSESDt
PRAT BUT SE AEE, FE RT RA AEA
AO, RARE, RT RMT” ORAM RR AEE AES,
LF 25 L A O P AS
BUA ASA IC, FES KBAR: “ME
*@5]” (supervised learning) #1 “Zi #444)” (unsupervised learning), 428
AUG AT HS AACA, TARR IE IS MBAR.
PERL, PLES OY BRE SRE RE “RAR”
HUA ELAR FERRE ETE REE, BILGE PRA IF A EME ES, BR
A tA BB A Sh FD EW AS. A a A
BRAM ED, MH “IE1” (generalization) fe 1. AAT HAZ 1 AE Hy AYA AE
RIESE TREE ASA], HE, IS ARI AR A a) BR
HOSE, BRAG A BBE AE HB AS A, a A
SEE AEA BE HB EME Ad ABT. PPAR. HR BEA AS fa op
AREAL — “Pada “5-4” (distribution) D, BATRA MEAL EBT,
HGR Ss ERAN, BD “ih ir F]4} Ai” (independent and identically
distributed, (8 ii.d.). —RiTA, WAAR, RNAB WHT D Mite B LOMAS ES
(ANB) V(CAD) Hh
BAK
mim ee
RRS, ROP OT RL 2 A a A 1.
1.3 (BARI
244 (induction) ‘532 (deduction) AHN ABATE. HAM,
9AM “HEAL” (generalization) i2AE, BUA FLEE Se AE
‘A; FEAT —IOCBINGIRA “F546” (cpecialization) At #E, BUMMER EERE
UFR. On, ARCH I RS, AT — A HANH SA
ARMOR, LR, EAPO” BIRR — PMN RE, BURA
“J8%5)” (inductive learning).
YAR DAIL LZ, LMU A AE ARLE,
TEBE 7 ER AA SH A (concept), LARA “Hs
PU" MEAT” RATER ATS. REL ABEL, BRS
EAE Hm YO SAE IAT, BEE ARBOR ASAE
A" BL I, RELA TA, NF RO 9B AL
EAA RAA RAT RMP, RUN GE” ORR" RAE HAT Ae
9 0/1 Aa ARIEL AR ASEH], ASANO, RR TEATS
PAB
a
oon] &
SRS HY ee EI EA ET” “ERA
“RE” RSTAR EME, RAZ, ABN MAES TREAT
BUREN HEART, PAL, BUEN “ETRE, HE
HAE. SOP” ROR NOIR As, FAA ARREST “OEE TE > (B
BRS1) A (HAR?) A (MU 2)” RSL 7” Rem HEI, TRAIT
PRR 1.1 MUIRAT EA, 12” WER.
BATE G ERR, LL A— A: “(RR AR) A (HERE) A RE
PSS phOG)” ARAL TS? AEM, RAE LA, TEA
ABSA “REAL”, UML ML Un aA Hy A Ae A I 13 mR 5
ES RIS, FAT. UROL OIE UNAS I ARE" , SOTERA SAR TAL
[Cohen and Feigenbaum, Wi, (HAE, AIBEALENY ML, dm “(CPR EREL) A (HLA HE) A (RETR va)”
190 A RARER i?
were
BRUT RY DLE #23 RUPEE — “EBT I BL (hypothesis) #8 SH 22 DPA
AURAL AE, AU HAR ALAR FNMA “UCI” (fit) BL, DRAB AS
HAIR TE MMAR, HAAR EL, ERAT IK NET
RATA LR EA An“ (CREP) 0 HEE?) a (RH =2)” OT REAR
PTE RAL, LIRA RAR” EY ORE” k= ATER;
FEMANBSI, LE ORE” THOME ZUM A, BRACES “8” REA,
fli “5 > (ERE +) CAPER) (Reo) ” AIA
9h, RAY, FEA BRAT” «wh, HA AA AN sR
JO" SAMARAS ARLE, HEE HEIL” BOBO BRAT © RARIK
CUE USE (MELB GRR, a CLR” “URE Rv” amy. 2. 2 PATRI, UR
eee a Ae ATMA 4 x 3x 341 = 97. B11 AML a TR
jarasaenen, :
Moco rmacemec ton” PME BG eT,
(Gian BB+ B=)
(aR) [CR ....
(ek We HE Me.) CEN WR RE
(GRR Pe (GRA, ER ETD |...
Bi. akaaHmein
BULL ATH wR RE GTR, On AY FAA,
LE ALL. ARGO, MeL ACh mT LAS AG AM RG TE UAB BL
its BAC eae 59 7
anaes a, By BURR) SHABBIR, RAEN RAR AG UIRSR Be TAT EA
emaster, Aik: SEAT EMPIM) ELL, RETA S MMR.
Rai RHA
Re. BEM, MIP MN EO, (EN
FATRPER UIRSSIAT IN, DL, PREAH SMELLS USB, RAE AS
VERSE BLN “EAR” , BRITE 29 “ARIA” (version space). ia,
CE LDR, 5H 1.1 UR ERA 1.2 BE Acatitbskey a
PATRI ATE
ee EMS TA
-
Bf HUY Rab oat
BY REM, Bt,
Bek re Ree
46° (eature selection)
K ORAL EE
Abe eR ART
SLR MATRA,
FARRAR TH
AAR HS
EM; LEA RT
BTR RE EAA
Seth A A A
AFHCARTRUAT
sagas.
ike it
Gi RAE
Gia]
(Gia PD
12 SAMRAT
1.4 VAS as
SELF 3 A YT RS PR, PE, P12
ASMA RNR AM: REA ST SUR BR, 5 EA
SRA ET PAR AON AR, AE A. PO, A (
r= HE; A AAR, MRRNRANE “Mo (8
FE 6) A (AMER) A ROSS 4)”, MARRS MAM AEM, MORK
FAT SAME, UA RAE. HBA, RR — MAY (aR
WE?
PRA 1.1 PAUP, RTI LS MRRP IE”
Ri, ME —TAANE AEN, COMBE EMR. RN,
EARS “Uae” BRB CAR PEA. A, BARAT SER AE IK “RT BE
Fe” OAL, MEER CHE Mo (EES «) A (HEAR) A Bh)” ;
ABRAM ME TR “ROE AL” AOR, 3h AL SARS “a”
HU, MERE “HEM (BPRS «) A (HER REME) ACTS «)” . SLRRAED
SEE ER RA HM, BH “IN ARAE” (inductive bias),
NEA “tna”
AE AL BEF AS BEAK Ds SEE, 25 RUE LL
LENE
t 2 HTL he)
Rit
()ARTB (BREA
BBL BASRA. (RR: MARIE; OR: RAE)
MELE, ALAR! RRL HLT a He Sy ERE, (SASH BRI 1.4(b) EY
HR: 5A FALL, B SUISSE A — BL?
(RB, RES REL. AZ, PE of, BEER
Sei Eee MK &, OF, MRE SL, EMAL Oy He Ly oF.
BUA, AMATO RT EE TSH LM, BAA AS 8 UE Sh
WHEW Lo TOME “BLAU” HEA 2. OS? REE
He ANTE:
Py AS IR, AAA] AUER EF) 1 BLT MA. P(A|X, La)
REGGE LBP WARE X PEE h OER, HS f RRNA BEY
HSC ARML. 2, 1 “UIBRUNRIE” , BUS, ZEUS STA EB
WE
Eae(SalX,f)= >> SD P(e) (h(a) ¢ F(@)) P(h| X,£4) (1A)
W 26x
Soop 1.) aie, BF ALIA 1, AF RUA ALO.
ASSIA, ELEC EUR ACHT LRA TBC + {0,1}, CTA
{OUI ABET ATER f BAVARIA, A
DBaltalh, A= ry DY Pe) Url) ¢ f(w)) P(r | X,£2)
We aeReX
> PCa) @) EPI X84) Tle) ) # F(a)
2oXX
= ¥ Pe YEP | y 21
26x-x
pe Pe) POh| x8)
20x 14 ARSE
9
Fibs NEL Ramee
RECS ERS
=241 > P(e). (1.2)
2oXoX
SA(1.2)SE ARH, ARERR SEAL SEL PERSIE 8
£;, RATA
DY Bete GalX, f) = Y7 Borel 1X, f) 5 (13)
T 7
WARE W, KICK AME 2, SHRM, US oy SM, CMAN RE
‘RHI RRR “HEAT MRAP” EAL (No Free Lunch Theorem, fifk NFL
5258) [Wolpert, 1996; Wolpert and Macready, 1995)
SRF, ARAL ae AT EE KET: BEAR TATA
MATE ME BERNE AVN AS, MA Ht A EEO?
RIERA, NFL ERAT HSA: HA “lM” ALA
Fl, RTA LR. SMEAR. SMR, RNR EA
CE CEA Pa A Det Hl (IOS RES), AD ERB — ARR,
BPR MRT REM ME, EEEADM MLA SAMAR, RANA
ian, TRA A HSI B th, SRR IE PEAS KB A RE OR
fh. BEM RM, BA “WATE” RMT OMRTR RRM A
HERB. BRAG CBT ATT RE, (RAT RIF AL.
BIEL, LIM NFL EMEA ASEAN RE BBL Tf SSI,
Testes lt, In, (LB RAT RAAB NLA, Apa (BEB A: API + (CHR #)
1 (HRA =ABE) 0 (BFE HOE) A (BBE 2: ET o> (EAE +) A (HABE)
‘8 (BRP =SHEMR)). HA NFL 5 TJ, STM EP. RAC SN A
RIOT, MAP MCAS, A AER; apa) AEA,
SE (CR f= BRE; Wi) IB Wa,
ROPE (RITTER, “(HLA RBM; HT — BM)” HAE CLARA IL, TCH
ADEA; RIM)” CORE, BARE.
RDA, NPL se SEIN, AULA ARTA DU, BUI
ENRIGHT ZR SURLE” REERE, BATA TAT EMO, HF
TIT. BRE RITA HR, TRENT SUE I A
SOS RUA SR, EH LARA MA, FU
SUE 5 AR I, Cee se sein
1965 4, Feigenbaum =
AMT ERLE
‘$A 9 DENDRAL.
Bh pz
Me UR A
RB SHIM
mee
15 BRR
LEE ALA (artificial intelligence) HF A AI — 5 MERAY LR
1. SECT EARLE TEACH, AL SRE “HEE” GREAT
DAR EAE F OLE RHEE AED, ULAR RAE RAT WME. WERE
{EERO A. Newell fil H. Simon ft) “2481163” (Logic Theorist) MEAP I
JHE 4 “ if FA] RAF” (General Problem Solving) fF, #6 (Eee SBT
MATOMB HER. HN, “BRB” FARE 1952 FER THe
RPAH S (COCR) Pi 38 EH, 4 1963 FUER T EAB 52
KEE, HHA -RMA, EH 285 SL PRAMAS, A.
Newell #1 H. Simon BAIA MAN LRG T 1975 FORK. Mi, BARR
PLATA, METRE), OSA RE ERAT AT HEA
E. A, Feigenbaum AV 21, SEALS AAW AE, BLA BLL AA SOIR.
RISE, MHS EERIE, AREA T “STIR
Wh”. ERM, AA ASE tH, AES CA TARR. B.A.
Feigenbaum fF “SALA” ZSC7E 1994 FRG BRK. (HR, MEAT
VA, CRAIG “ATA” , MBE, BE RIE
SESH OLAS AEDS, Pe, AE ARE, ORAL SRR
ENE
WEL, ARE 1950 FRFR RMA He, ER) TLE AAT
fe; SHS ER ALE AEST, A A. Samuel H% MBE
BOLE. RAE, PME MAY “RAL” (connectionism) #7
FERED, ARztELAEA FP. Rosenblatt [484161 (Perceptron), B. Widrow ft
Adaline . ZEA tPA, SEP IB AR ath “AFSL” (symbolism) HE
AMBRE, KALA P. Winston fh) “MHF RAL” . R. S. Michalski
BAM “PO ASS RH” EB. Hunt SAM “BERR RE”
BR, CLP SB Ya MEM DARL GAR BI, FORE.
tPA N. J. Nilson 8) “ 222) 618%” 8; —+24e eA ASEH APS EE
EERIE EG RH EER AN A
1980 4.3L, 43% HRV EASRE EAST TS — ALAR Tt (WML);
FE, (RTS RAD MEMS MPL RAE SM; 1983 4, Tioga HINGE
AMET R.S. Michalski. J. G. Carbonell #'T. Mitchell 3:44 (BLIE4 31: —
ALA AEIRZE) [Michalski et al., 1983], 814i OLARED OTF L EET T
3H; 1986 4F, 8 —AHLAEE AE WIITY Machine Learning QI; 1989 4, A 15 RE
BLK Ot.
RBH LP eT.
BLES t.
n
Le TRAIT) Artificial Intelligence HARHLB SEB, WaT AB
ARERR ERATE LE, IC FESR SILZE J. G. Carbonell £44. MIT it
foett 1990 4A) (BLA): HAG ATLY (Carbonell, 1990] —F84. BMIA,
SHE ERE BLS FR — ER LE BR
EERIE ASA.
R. S. Michalski % A [Michalski et al., 1983] #261842) 7 RRIS A “WE
IPE” ARM MR Ee” MRM RET”
pea)” SMA; E. A. Feigenbaum SABAH (ALMA FA (=A)
[Cohen and Feigenbaum, 1983] , MHEGLS2 211529 “SURE” “aR
eA" REA AM eA”. SURE IRER “IIMA” |
ABS AMEE ABI FR, ZETEC Ah ah HR HH RRA, Rb
BRET REISE, (UE TERT TSR: RE AR RAL
FR.S. Michalski $ASTBLA “Mii 4e 2)" #1 “CEL ARLE” 5
VARESE “REO”, BAI Ea 2 HE a, hth
UE DOR, MURS . ARTE “MSE” (ET
WSVGA AD), CE TREE. FEMA) A, AS AB AS a.
POL BRAT AS Ay A A RE 8 EL.
FEL HEE, OEE)” ERE SE AE,
SAR 10H BEIM (decision tree) #19 F B48 3 7). SH Be BREESE DE
Bitwaemt, LU BMGs be, BRB TR eT Ae
PEMA. PMS HH BARR Rit (Inductive Logic
Programming, ff)¥5 ILP), PI @R#L 2 #7) -5 EF RH eR, EE —
Ors 9 (BUH Tali) eae AT A eon, WLS ANS FEI AR AAT (AN Prolog
FER) ITER AGE ASIAN. AES Fe SE 2] HE iL MEA A
WRRABAA ATM. BBL, ARETE BER EA
T “HEA” Fo “UR”, Ze HER” AES RR Zen,
ALAA T Ama, MOTE “RN” AT SRR RR
ROA FA GARI AEG RAR TORO, li, ZE “AES” AIR
SRIRAM AR BAR. Sb, SRE Ete ERE
BA “NRRL A le MSE” TIE EA TREE AY. DRE
ABARAT ORSH, ISK HERE AOS S ARAL — ILP AAR
ARTA Ae a AE A, AT RE Sh Hh eid HS A A, TA
RSH LE 2H AT A, HE, ILP OL A RB, LHS.
BLK OR,
PL IMES,
wim ee
SLE AY SRO EAT AN; RT, RT. A, ta Ae
FAB, HRS BCE AL PER EL ACK, SANE, FALL, fa
FRAME LIA GREAT EO, FUP AAR BR i He A
HEALER AZ BT, “OMIA S]” 055 EAE
PORN CAD). eH ORE eS TKR, (8
HARMS AL SRAM S Raw RE, Pin RAAE B.
Simon SBA A LWA “RP HRT AM Soa” , BTA RE RESE
AAA RA ERAT BETES. ARIE REEL 8 HI TA
UMC, IER IRATE M. Minsky MIS. Papert £1969 “E48, (STAY) ARE
WE FARAH POR, BR At Seah” IZ ALA ah RANT. 1983 4,
J. J, Hopfield AUF a2 A RAF “SRS HE IR RL” SKA NP ELC
HARDER, HOERE ERA AMA. 1986 4, D. E. Rumelhart FAH
PRAT HAM BP Ri, PET REM, SESE ET RE
BRAM, BREET PE RAR” EEL, ER AY ABE
FH, SHEERS; AT, TA BP ARR, He
HY CLAERS BAST ab REE. See, BP PARE LS
PUNE Z —. HERR OE MMA RUSE “REE” 5 ATID, FAS
HVORARSK, HERHRERSACMS, EERE TL “HS” ; SK-H
WH, SMMPLAZ SM, FABRE ITE,
SAHRA, ““BEEEAEDI” (statistical learning) ALFEAE SK It
HE TIE AEVOM G, (AeHEBOR ELE FO ALHL Support Vector Machine, (8%
SVM)LL RH —ARAH “BATE” (kernel methods). 7TH MWS At — tt
RCE E TEM, SETAE [Vapnik, 1998] AAPA ATE
T3ERH, lin VN. Vapnike 46 1963 REM T “KCRG” AES, AWA A. J.
Chervonenkis #1968 E861 VO Me, 2 1974 Sed T SHR dS.
A PLEU ALTAR AR SCHL AE A TERE BLA 3 EH, TY HI As FAT
SCR PARAL SCRR ZEAL TE (RHF ARE, SCORERS) ATE At MER
ASPIRE ATEDLEEDL: Ti, IE RAEI BOR BLE
2.08, MA 8 BGT DLE ES] Ee EL SCHR ETE THOR, RS
, Sih) SHE: LES AT IER, ASCH TEAL APH, Bete
B5(kernel trick) BEAMS T HUME DI LRG, BET LTR
BBEAMBAREZ—,
AMINE, —+— HHS), REE ARBLEK, MTU “REY 16 SRD
13
BESO,
Stee BK t.
AD” BRA. FRE, RITE IS i” Ee. ERT
FURAN, LKAGRGE, ARFRATR OMT, REEVE
BFS T RRL. CLE WL AR BEANE PE RE A, OT A ER
RM; WRF ALA ROR ER HE, UST REP IA “HS” ,
FSR, HERE IIT. AL, RE RZ Pe Me A, (BE
REAR T MLR T OF ATM, Lae AE LHR T BAL,
WA, Et AsO A RM? AE A A: BHAT. RE RT
REEARMRAKREM, BREAD, MURS “LIA”; le ARE
RL, UR NOIRPRAR, 2HR EBD Te, HOTEMORI. HF eat
AT “KMGRINAR” , SORE SSE BAT AAR, AAR Ce
BARRA, HALE, Hee ER, SR
Intel x86 AFH aS SPER EARS IE TE RE, AE
FRELETERADEREAEAR. REEMA, SA NS
ase or SACL.
HBL, MLE RAE OS RR KE SR, AT LL
REP, RERERAMGLATRA, MONRAARCAT SHES
HT AF,
16 SARA
FRAO HEP, ARR, TEE, fT. MIE ORE A TC
Ft, AREAS TB BUR T ABBE, BE HEART MET TAL
FHS SEALS, TOMAR E IAN TA REAR, BER aL
‘MR AMR BARR SEB IEE.
FR, ENA ENE ASR RP, EES RR. PIB, ME
PURER. AL, ERE. SHR, ORAS RD
H, RARE PM, ARASH OMAR” SR, PLE
FR A BBR RZ
MEFABAWERLLEARKT RMOBRALH. Glin, “LOTR
2" RBA ARRRTRE RRA, A Ht ey
PEF LAAT 2 RE. EN ETD BRM “EB” 3)
“SUR” (ME LAE, SURO LIOR, SORE. BENT. OF
KEOGH, RTT” PENBEDRAROR EG, SMB ENRA
BBA EG LHR. NASAUPL HA AX
AAEM GRABER
BE R26 HAS
han" FRE
YRERNREE AME,
DARPA 4k Ao
BB aa TR ITAL A,
ERA. ERESREA
BEAT DARPA Bb
eR
Aun EAT
9, aE Rae ak
BEWA, ROMO
ei
sation,
FUERTE RTH
FEL REWER
HenReemotER
EK AI IE
PREHALA I, ®
RAS OUTER TR
ARRAN, RARE
eae,
wie st
BLE, ANE ANBAF MEI “MICH” EBLE AY
“ORM” , 7B BE” RK, OLE AE
PEAR, FI “GE” ACE AER OT, RE Bt
ELT RE RAR OMA, ET th AT A AR TR RE
A AVOPSES SK, MA ALSO) OSE, 2001 4F, 3 NASA-JPL ANE
4 Science Ait £3 11983 [Mjolsness and DeCoste, 2001] #841, Lt 2] 8
REPT ERA ORAM CE, SEMPER EE LEK,
2003 4F, DARPA fa) PAL itl, HLS 72) MBE LIB RAL
GSE KE. SATAN, SE LR Re A CH ME AL ka NASA #1 DARPA #8
BEER, TOS PAPUA TPS RL SF I EE, IER A TM.
2006 4, RARE AAP ARIE LIRA ALAR” , LAE
OURS AZ—T. Mitchell SRA ERERAE, 201277 3B, REELS
BORED * MARTTI T RMR RI” , SNH AE a MERTEN IAB
SERIE RCL ns), AES KR KARR:
BUBED. RI, fe (crowdsourcing). EA, ALARA AUER LD
AALS EACH, EAL: OR, Ef Hei EAA,
TAI” MGR, ARRAS DL AREER HM, BL RUA” SEN
PINAR DORI, AS ASR “AER” (date mining), 3
SEALER — FATE So e HR, BRMBLARE— eE
AIG, CRSURS EAT, SUE WLAPETI, Se ETE
WARK (Zhou, 2003]. RIAL AN RAE p AAT, BLA BO “he
RSI” YEAR. AUR, A RM OA
LER, (OULREE VSR EOI RETRO THER. PEL
EF VST RR HE ALES TATE RCA ALNSE A, 2
AHA SAR, PASE AGK AE LN, Set we NL Ea
RIE, BLA SOR ASI eRe
SR, ABRIL SE AMER, PLEA, HIER
PR FSB, AAG PULP I A AB
FEA, A SR AOU AE, ZENE ANAL BL
PUB ERR, PMERSEAT AT, POLAT ABNER UP AERHIGARA,
ASB EBA RE HER Es FL
PRAT ASS, PK. ELBE ELI ae SME IPE ARES,
MRS ABA RT eT LIOR TA. RA 1.6 BARRA
WR eee eH
4 [Mitchel 1997] 42 4
BT apeeAT ERE
MARA METRE
#1 SOIR Hy ALVIN
Aa
WA, CS. RM MAT) SAMA AIRE: “COREE
4 RL 2S i PE FRE Ae FRI”,
JETS EMER STO, CLEP, AL aL
Ay RRERAMH, MER TMSA TMZ ORR, AI OLE
ABA. HL, ARMBRRRES, MBLILAMRBBHSH. ATS
Ry BRONR. ARAMA, EAA Ey ONE AT
“AR” WH, CAE A ME MLSE AR. Ba.
BE, MO, PRGA A CLLR ARAMA, BE
PLEA BAR MZ ETE BE, FES TE LAR BEAR A Be, BBE
Fe EI Ew HL
Fl. EREARRMNAFZ —, SHEE H LAA EE
$8, CRE RAL RTE. LE BEBE
BAAR, BAHL LAAT AT MRAM Ow. RASTER, BANE
GR, ALAA BE EP. SE AT IR
FAUDOEFE. 1k KAYE TENE a IEE Le a PT BOT
ARES EB), Bit AAI IFAS, TG A ARI Le NE
LATTA. FORE MB AEA, TEA AAS
YO 2 i UA 9 th LB HED. 2004
463 A, ZER DARPA MAM BBE, MAE ER
S. Thrunti) M1 HF Hal B AAA 6 Att 53 SPRING TE T 132 I
JE, LOSER BEE A Hei DS A AEA LL kN, BLA SE, ZERO
PRB EAT BURR Me EK OLAS PLE AE — PAR. S. Thrun REI
OLS 6 BET A ABA. (EAS RRR, A RE ALE TGR
RR, RAMS, BR, Ri. KK. VSR A MRA BT
HR, AM OHGAP BREATH. 2011 6 A, RM APART,
Hy Be — AY AI, RE I
SRW. ARE BEA AYRE RAEI PLA
PARAM T “AHL” FEF.
PSFIRARZ CHM ARHA BGI. 2012 FRA AGL,
EGR PARSE AAA, HAD SIE SORT, A a
FB RAT Bh. A AALS EBERT TH ER, HA HE
REAR —UPHEZ FER SR, TIED AT ITE Bt ME
ESI, 1 Ha i A) — A a, AAT ALE WEKA LEZ HS
RESINRARR,
M8 Walato KFA
RANRE JAVA FH
het ome tat
n/m.
mie ee
SUA SR OAL IAI, A AER LE Pt
aM, TO RO SRM AAR, SRA BCR, UR
PABA ITRI T AK. On, LRT BNA, HERE
EID ERGERRPRPS A BERTI] J), ESR VRE SO
SUE IRMR, EA UE EDR RDN ARS 1500 TSR: EE,
SALAS, SUC GT OLSEN 10 (CIR TEREARLEN. BLA BATR
ADELE EI”, HAT AY “NM” , GLARE BML
PURPA ERAT APA, ALR T Hever yO BL, MORALE
SAR em BIN ee, WERT a eam HL 2008 Ae A
T 14%; AES, CR MAST RRR ES “RR
BY. AARNE VOIR R. Chan AG MEA.
HRA, WRB SEO SAR ALP RRR BR
RFU, LOLA AE TEA AAT SH
FAA RAR RAEI “JEANIE”. Li, P. Kanerva 42—
“PALES 4-4 FUsP BEALL SDM (Sparse Distributed Memory) 84! (Kanerva,
1988] RISER ALARA RE AEE HY, (LABRET AUR, SDM HA
RRP RRSL, FE. EARN BILE Ree, INT RAH
ACSpASLOLT HLH. LIRR E RESON MRA J EERE AEA EA IE A
Riis TTA. ERA. GRASP, TE ASI)” FEE
AR ABATAM HAIL, IRAE EE, ULB ROLLE RL
WREOLAY, DICH NER OR,
17 BEAR
[Mitchell, 1997] 2235 —A La TTHESCH, [Duda et al., 2001; Al
paydin, 2004; Flach, 2012) #2 Hi 4800 A174. [Hastie et al., 2009] 407
AAEBTUY, [Bishop, 2006] HARA SA OMEL, LSE A Mat We URE
[Shalev-Shwartz and Ben-David, 2014) {sd -FS0i6(W 05-2. [Witten et al.,
2011) BF WEKA RS ATR, ATS Hii WEKA SRE
HRA Rik.
ARH 1.5 Al 6 WEBROOT (, 2007]. CULRBED: HALA
378) [Michalski et al., 1983] 1058 T 20 (c°¥-AIMS HY 16 RICH, EBL AED
UPC A SCWR. BB NUE 7 4AM, Morgan Kaufmann tht
RAUF 1986 4M 1990 FHT RBS, BOMB. CAT 17 mikes
RREILLSOF.
BFS 15
REISE 8H.
wv
‘WHE M) RIVER B. A. Feigenbaum IRMEA A HTT, He
4590=48 [Cohen and Feigenbaum, 1983] 370.8 51847 Tift, ALR
SIAMESE SCM. [Dietterich, 1997] 8 HLS] SURLY Re MET T VERA A a.
AIRS ARLE YL EAL, BETA IE AE ASR EA A
RATE, (UME RIEAT IN “EBT!” (transfer learning) [Pan and Yang,
2010), HL “24H%21” (learning by analogy) ZEEE THA RAL HFT
BIG; ACR —OY MY RIE” (deep learning) ZEAE EIA LAMAR
SHEA ESS GF OE.
MELA PRE MSE ANT RIEARE, APP EY AN > AE TS
FAGBARLEM, Hint ALRE ARR — ORR, RRRT HK
FES IE PH [Hunt and Hovland, 1963]. [Winston, 1970] 42%
Bits BAUR” FR, CMSEA SEP EAAPLORRUARAR
OK. [Simon and Lea, 1974] KPH T “FA” AERA PRAM.
(Mitchell, 197] sadeth T AASB SS. BERET PRE EME
Pa
SRE HET SUE SR SMR — BP ME, ESE RE
BE. RESP AOE, Bio PJ mE “EL
Oi" 2 REE EE HOR” MA RCT.
PRB EDL RE Sih th A 1 SIG AH [Blumer et al., 1996]. HHLEAEN
Seth ZA RN A” JA ASE AAT, Palak, Re 77
DBD HU tf AAPL FAL [Webb, 1996; Domingos, 1999]. IER
ASE, A 7) FE AE LEE HOE — A A I, So a BT
RPE HH (28 70 3414-2704 4 AY“ & RERUN” (principle of multiple
explanations), E3k(R H 52 WR — BHT A MB [Asmis, 1984], 1 SEHR
*P2] (ensemble learning) 7 HAHAH INE.
PBFA WR LB NRA WE LAE SL(ICML).
AAR BALM RRL (NIPS) A FA BRS (COLT), BEAK RES
WEB WHOLE A SVL (ECML) AW HPL SF I SiN (ACML); A HY
HARREAMTALE Journal of Machine Learning Research #1 Machine Learning.
ATH RESELL ICAL. AAAT LRTI M Artificial Intell
gence, Journal of Artificial Intelligence Research, 33 4:38 aif 3 2 WL
‘i KDD, ICDM LEBEN Fi ACM Transactions on Knowledge Discovery
from Data, Data Mining and Knowledge Discovery, it ##LL% SHAH mae #
BA) HBA Win CVPR LLAMA IEEE Transactions on Pattern
Analysis and Machine Intelligence, #425 F1S% SUR) HBEMTIiN Neural Com-
putation, IEEE Transactions on Neural Networks and Learning Systems
HERRRMBLA AMEX. Hb, Rit FARA RWS T MM Annals of
Statistica BM AART RU EA HOMME.
HARDHRAAM SE AHMAR, Bion (Bit, 1996). [4281, 2012]
FELBERT, FPL A 2) a HE —
9 BBL 88 KS (COML) UL A EAT AY “OLR RM” BRT
2 (MLA); REFRWWBE HA BARBS A MEX. ae
MOREE S ADRK
HP,
dm: aaT ERA,
(A= a)v(A=*)
5 (A= 9) th
wRawURMRAO 13
eit
14
15
ALL PSUR I LATA (PEPPER HRT.
SS UAL ARGU AeA, HE) “BER” HERS EL
BAM AAR RINRRKED. WAN
HE) c+ ((CE= ») A (HAR Rt 4 (LF= +)
V ((GR= SIR) A (HERES ») 0 (R=BTFD)),
Ha)” BAP IR A RET”
BUR RAARAMA CARI 1.1 MP
Mh), GRASS S 2 ET REM BLL.
PERL, UAL of BEAN A ZEST A BB
AABLBL. CEDIA TEP RR — AP OF Be a.
AE LA HAE RAT RIRER” RENL, BEAT“
WR” PPAR ION UBT HE, ARMA ICABIE AEE,
(LAH
Ene(SalXif)= >> D> Pl@e(h(@) sf (@)) P(h| Xe) 5
W 2ex-x
TIE “Re aR EEE” Oona.
RA BLAR Fy ECE ELI Pe DB SE ett PEA mim ee
Bax
Bite. (1996). ALAM CPAE) . BSE MAL, Ita.
FAM. (2007). “LARA SWRI” op RH SEBLAE HMR, 3(12):35-44.
ARH. (2012). BEA HE. RE HARE, ALR.
Alpaydin, E. (2004). Introduction to Machine Learning. MIT Press, Cambridge,
MA.
Asinis, E. (1984). Epicurus’ Scientific Method, Cornell University Press, Ithaca,
NY.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer,
New York, NY.
Blumer, A., A. Bhrenfeucht, D. Haussler, and M. K. Warmuth. (1996). “Oc-
cam’s razor.” Information Processing Letters, 24(6):377-880.
Carbonell, J.G., ed. (1990). Machine Learning: Paradigms and Methods. MIT
Press, Cambridge, MA.
Cohen, P. R. and E. A. Feigenbaum, eds. (1983). The Handbook of Artificial
Intelligence, volume 3. William Kaufmann, New York, NY.
Dietterich, ’. G, (1997), “Machine learning research: Four current directions.
AI Magazine, 18(4):97-136.
Domingos, P. (1999). “The role of Occam's razor in knowledge discovery.” Data
Mining and Knowledge Discovery, 3(4):409-425,
Duda, R. O., P. E. Hart, and D. G. Stork. (2001). Pattern Classification, 2nd
edition, John Wiley & Sons, New York, NY.
Flach, P. (2012). Machine Learning: The Art and Science of Algorithms that
Make Sense of Data. Cambridge University Press, Cambridge, UK.
Hand, D., H. Manila, and P. Smyth. (2001). Principles of Data Mining. MIT
Press, Cambridge, MA.
Hastie, T., R. Tibshirani, and J. Friedman. (2009). The Elements of Statistical
Learning, 2nd edition. Springer, New York, NY.
‘Hunt, E. G. and D. 1. Hovland, (1963). “Programming a model of human con-
cept formation.” In Computers and Thought (E. Feigenbaum and J, Feldman,
eds.), 310-325, McGraw Hill, New York, NY. eaxet
2
Kanerva, P. (1988). Sparse Distributed Memory. MIT Press, Cambridge, MA.
Michalski, R.S., J. G. Carbonell, and ‘T. M. Mitchell, eds. (1983). Machine
Learning: An Artificial Intelligence Approach. ‘Tioga, Palo Alto, CA.
Mitchell, T. (1997). Machine Learning. McGraw Hill, New York, NY.
Mitchell, T. M. (197). “Version spaces: A candidate elimination approach to
rule learning,” In Proceedings of the 5th International Joint Conference on
Artificial Intelligence (IJCAI), 305-810, Cambridge, MA.
Mijolsness, B. and D. DeCoste. (2001). “Machine learning for science: State of
the art and future prospects.” Science, 293(5537):2051-2056.
Pan, 8. J. and Q. Yang, (2010). “A survey of transfer learning.” IEEE Trans-
actions on Knowledge and Data Engineering, 22(10):1345-1359.
Shalev-Shwartz, 8, and 8, Ben-David. (2014). Understanding Machine Learn-
ing. Cambridge University Press, Cambridge, UK.
Simon, H. A. and G. Lea, (1974). “Problem solving and rule induction: A
unified view.” In Knowledge and Cognition (L. W. Gregg, ed.), 105-127,
Erlbaum, New York, NY.
‘Vapnik, V.N. (1998). Statistical Learning Theory. Wiley, New York, NY.
Webb, G. I. (1996). “Further experimental evidence against the utility of Oc-
cam’s razor.” Journal of Artificial Intelligence Research, 43:397-417.
‘Winston, P. H. (1970). “Learning structural descriptions from examples.” Tech-
nical Report AI-TR-231, AI Lab, MIT, Cambridge, MA,
Witten, I. H., E. Frank, and M. A. Hall. (2011). Data Mining: Practical Ma-
chine Learing Tools and Techniques, 3rd edition. Elsevier, Burlington, MA.
Wolpert, D. H. (1996). “The lack of a priori distinctions between learning al-
gorithms.” Neural Computation, 8(7):1341-1390.
Wolpert, D. H. and W. G. Macready. (1995). “No free lunch theorems for
search.” Technical Report SFI-TR-05-010, Santa Fe Institute, Sante Fe,
NM,
Zhou, Z.-H. (2003). “Three perspectives of data mining.” Artificial Intelligence,
143(1):139-146. me ee
hE-B)L
ABR: HRA” 4
1952 4, BIE + #484 (Arthur Samuel, 1901—1990)
IBM 2A FT NORE, RE RAE
PDH, TAHA APRA AAA RY SH
Ray EHS FRAY, AAR RRA, HR
ARERR GOTT RBA HB. 1950 4, RAH He
He, (John McCarthy, “ALA M#2X" , 1971 FH RKRE) Lit, A SAA
TEEPE AGERE REPAR ATE. FBR T OLR"
EAN, ARENA ALARA TEI AR
3 “Some studies in machine learning using the game of checkers” 1959 42
IBM Journal ERR RE, RAP » ALM (Edward Feigenbaum, “friRT
BAR” , 1904 FH RRAE) ARG RES Computers and Thought, # 1961
ERA EBRE RP REE RO. TE, RAC A RE
NH tt MeaR eI. Mab Rae BS OU ARS AST aa, AR RR,
AHS RRA.
SLE, PERMA A TOEALS ROME TBAB, BUMS
SAAS RR, PSMA ERA, HLA fe 5 Ow
RERRER HES, WRT Bit TI MAM. Bb, REAR
PERALTA LT EES EE 2 —, RAHAT ERA
94 T IBM HEALER, HRS He RAL ER RES HERB
= 8) x 100%,
EAT RET
ELIE,
AERORT PRR
FRR Sako
Hesse.
steak “tee
RAIA “KO
Paade tra
AD Rn Rakes
HARE MARE.
B2e RAAT
21 BBRESTUA
TER ATIE TD RARE A BPA HS RR “HEAR” (error
rate), BUMURTE m RATA a MEAD IR, MWR E = a/m; HAVE,
1—a/m 6% “BE” (accuracy), BO “RERE= 1-H” . EAR, RATE
2) BB AY Se es LM tS A HZ EERE “GRE” (error),
BE ERA A “VARA” (training error) “BHR
%” (empirical error), 26 #7 FAL RHA “i (LIR2H” (generalization
error). GAR, BATA ABE RI NOE AB. HRT, BLAS IAN
PALA, LORMAN ES DRAWER ML. ARSE, RATT
FATED, EVR ERE A, GE Ht TA
PABDRER, MPRARE HS, PIE 100%, RARE RTE
O22) BUG? SENIOR AED BES RHE FAR.
RNS, BERRA LAREN a. Ob TRB
AY, SE APE AS HAR A SE BA ER LAR”, 3
PEA EERE ALTA A at EA. PSD, FRE
BE” THAR, HEAT EEE RA A SO — He ET TEE A AD
SAAN AEM, BARS SRO TM. RAMBLES PA
“idi4” (overfitting). 45 “IMS” HHUA “KA” (underfitting), &
REN UAAN BER M ASH, 21 ST RPMS RMA
MEF ARM BRE,
ASHAR TESLA, HP RAREST AA DLT EK,
ABFA BAA TLS AAA EP BT TCH Ea
FPA REIN FIRMS. KA RE SE, PUNTER RE PT
S. EMAAR PES, TO. AE
BATRA, TALE MRE, BE AR
Pept RAO AA; RTT AUR, SAA AE ETA EA, BATT AE
RUE “BEAR” , RA He HB. PARA, TABOR: BLEED
‘i 9 LS A FNP MES EA TAR 2) RH SK TA ALRERE RARE
PERE. AAT
4h, TAREE HL
A ROTORS REL
RE.
2m RMR
notte Ra
a i
= ® (oD)
# %
# +
* RED RIK:
> abet
(DLE WER tet
M21 ithe. Kiowa
GSAT TEM, AAAS Cy, Le I CARRERA, BOR
RE RATHLEEOIEW T “PNP” ; Dik, BORIS OEEUIG, MEM, SAA RE MBINE? BIE: WALA D MATES
AOR, APE HH RSE S ADIURSE T . R TATE SO
22.1 ee
“HUNK” (hold-out) ABE BAER D AERIS, Je
SEA WER S, B—MEAMRRT, BD = SUT, SOT =2. SEV
PRUNRAU ES, FT VPA SCRRIRIE, HERE HIRE ATT
PII ARES A, BLE DW 1000 MEA, ILIA S fae 700 ME
A, T A 300 MEA, FS MEATUS, MARBLE T EA 90 RASS
BR, IU (90/300) x 100% = 30%, HAN, HABE 1 — 30% = 70%.
ERTL, UA / MRE AL OE ST Re OR BR 2 A — BE, ER
(SES a SAB A A A SR a, HP RE
BD BR RAI He DIAL WFR AFORE (sampling) ff fa ERA HR
SRA, AR FB EG BMT UB “RPE” (stratified
sampling). 103%} D HATA RA MTGE 70% RAMUS S RE
30% FRAMMIAK T, A D A 500 MERI, 500 TBI, WSR
S RAGA 350 FIED 350 TRI, MT MAA 150 TERIA 150 BES 27
Ss T PREAH LAE IK, URE EEA FUER / HUBRIS A BE
iE tmize
ATER MTR, UEC Se / UBM, Ute
RI AOE D BET OME. Blin de LTBI F sh, AT ELE D op aE
ARHEF, SACL 350 SEABED, 128) DELLS 350 +E LEH
BRR, oo MEAS a ES A F/R, A, RA
Reta oP P,P RA OT, A
AMHR, OB RAA PMOL, BUT SRA ROPE
Pr TAOS VEER. GUAT 100 YRBEBL LIS, EU — A / URE
FSRPE, 100 BG REAPR 100 55, TO AH HH SAIL FBI 100 MARAT
pay.
HS, BATA SPA A FE DUR es ARI AE BE, a AS PL 26 om MOURA SRE
WO ee ee ee ee
way na ciel, CUBCMRTREREARE TAA D VISA, (TT Heke, HERAT EAR
BEAN MORRO Bieta, OMT S1a8 AHR, RIVE S45 D AREKT, WOE
Gaxencik GMOBEMLSAE D VRRULMU BLM ANECMT RT KIER, MBH T HEE RAD ER
nas, mney Rt(Gdelity). DRA RIMM, A LM AALS 2/3 ~ 4/5 H8
Ae 20 FS Neel, REAL T Veh, AURA AT BU
997),
22.2 SRWIEE
“38 LMPIEWA” (cross validation) S:¥68C3B4E D 3) 25 bk PA MAEMO
ELRTSE, Bl D = Dy UD2U...U Dy, DAD; = @ (#5). AFR D,
FRO REAREP RB A) BORE, UIA. D PLP RAM 28a, SUCH
Ke 1 PASE HE UME, FETA UR, CEA IRA Ae
SALVAGE, IATETT EAT be UC UNGAR, FS LAY EI oe TURE
VSS. SEAR, Se RE AAO UR Re OR Us AE AP he
HAH, PBI — oi, AAAS SR ERR “TELE” (h-fold cross
SE ARIE”. Vatidation). k AVA MULL 10, ATARI 10 dE LH Ata FL eA
A 5, 205%. P22 T 10 HAE RAE A
o) aR
nnn? [49 ee
Be
[Bs [os[.[o. [Bs [BT [To] [,] > wines R10
B22 wekrERES
‘GREAT, HERES D falah oy k PPR EAES PD AK. Oy
AA EAR Aad 53 78 FD SA), PEE BE LAF AS 9 Sl
ER p th, RAAT RRR p te hE RMS, OPO AT
0K RRE w se »
en MRE “10 10 WCRI
eae (BENGE D PAL m MPA, HS k= mm, WABI BABE
49001: {Hl —ER(Leave-One-Out, fi LOO). SLA, fi RAT SEREALAR ARIS 22m
BLAH 22
NFL REAR LAF.
ATHAL ERS CH
AML ALA, RO
2,
Bootstrap hE S “HR
P ReesenaD
WRF OE
AERMIED FORE B
MER, DRASED
“Aah”. AMER
TELA OF
aba it
eR ARR.
on ae Raa.
RAFTS OS.
a
AOI, BI m AEA RATERS A mm FE FEAL
AMA, BFE FU 5 SR ALR > TR, OE
FEA S MAI F, PAS SPA LA D URL
AURRARLADL. te, REE BE, MATT, TIE
RU: PEMA SR HERA, UN a TRA A EY A A 3
PROS 1 FMEA, REL RM), RRR ARES
AHR F. SSh, RAT IR th A Ae EE; “BA He,
SAR” SEED MATT AEA
22.3 BBN
RATA D WAR. AE URE RE, a
PRD T BBS PEAR FP A, DH A A FL ARE Hs Ds, 3
ARTA — AEA IRS PRC IE, RAI,
PBR, HMB RAT. HRA ADEM MRD UE
FRSA FL REI, TPL He ET SEM TE?
“ BBbK” (bootstrapping) 2— He BAF MRK, CAR BE
VW(bootstrap sampling) HRM [Bfron and Tibshirani, 1993). #52402 m ME
ARAYABIE D , BATE MEAT IE EMER D's SEUCREAL MA D HRA
FEA, SUE IUBO. DI, AS PERE A UIA BERR D A, REAR
POOR PE IYA FT RE BOR SI; XE ST m WG, RAAB T Bm
MEAS D’, RARE A BRAS. SLR, D PAAR AR TE DI
PSUBL, 1B MAA EBL ATLA GET, PEASZE m BOR
PEPE BORSA (1 — 2)", RAED
stn (1-3) 20.368, (2)
SHE BURA, BHAA D PATA 90.8% HRP ALLER AMAR
5, PALARUTATH D! PEVIERER, DD’ APRA OPE, SERRE BAS
Sa PERU AE m PUSRPE A, TERI ASR RES 1/9, SAE
Sle ALONE MAR, RAOTRRA, INA LIHETE” (out-of bag
timate)
ERASER AN ARDY ARLYN URS LAT, oh, EAB
AAAI Pe BASE, BOATS AK
‘Ri, FABRE A RSCTA RR SAA, EBL AHR. 28 Som RETR
Se, CEA AL AB INT, BP AVA NM EH ME He FH.
224 WS SRARD
KE MAA SORA H BH (parameter) HWE, SEACMAK, HAI
AAALAC SEL. Be, AERT BUA RRPERL, kT SERRE D)
SORTA, DE MIGRSSORT Re, REGU MATIN “SRT
{98 “i” (parameter tuning).
RTH ABS, WS PC EBAY ACLs A
‘RUB, A HERE LARA A ME, SRP A A A,
fT ARE: POURS SMALE ATSIC, DUE, MEA
RECUALRURR TTA, DRE MB, REM S ROLE
HOIRLATSEAA Ke, Ltd 0, 0.2] TERIAL 0.05 22K, RUSE HORS
SHAT 5 +, REO 6 MRO EE. mR, BRE RE
C7 RCE” i, (ERATE TP ROH ARTE TT MR, NL
Rebiieh, POA RATT. HL, MILER AAT Hw LE
(ORI. ATURE: BO 3 FL, ATRL 5 MRE
(it, ARTs aa ok Ra 9° = 125 PA, ALA
WAKRSR ME, LY ISUAMES TRE, VETER ONES,
Meee R ee, SHCURERASIE ART RAMU RE ARBRE
Set m MAARTEN D , CERI tah th mE
RR RET AM, HS E RAL AA T — a, Da,
SMG, FIA SHR E CRS, NARI D ERG
BL GAVEL VURAL RPT PT ma NPA, A ERATE
pti,
a, REA, RAH I FP SR
RSA, 2 TANCE, BEE EAR PUR ER AM
ER" (validation set). Glin, #EHISUA LAN MISTRANE (GHAR, RAVER
SEAL RRO it BE ee Ia AE EE, EIR
AMSA ESE, He FATE EAE ETE T BA AAS,
2.3 tEAERE EE
YES NE ACHE RE RET P, AEA AT RIK, EG
BA BERANE HC AED MF HTM, i wR AL HE AB HE Mit (performance measure). 23 teem
RAGHRRERDE
ot.
29
EAB ERL BCT AES HGR, CERT He AS aA A A, RT AS
ESSA TARR: RRA “EK” RAT, TAR
PERE, AOR SARE, RET ES ER.
CERES, MEAIB D = {(0r, 41), (2,2), ---5 (ms ten) SE ys
FEA oy OESORI, BEI f OTE, MEIC LER f(@)
SSSR y MATHER.
BUEN RLR, “37FBLIE” (mean squared error)
E(f;D)=— (F(a) — wy - (22)
a
3E— fH, APES} AER BEM), SI RET
BUD)= [ (6@)-w? neler (238)
l~D
RAT PUM ESHA AE EE
2.3.1 RES MBE
ABA AMA TR AU AE, AE IRE he FO PE AE,
BETAS, HINTS APES. MARA LDR A APE
ARAMA EEA, FASE FE PR TE AA EA PEE, RIED, SP
RARE
BUD)=2 Ye) AW) ea)
oi
HUE LA
aec(f:D) = ZICH
=
= 1-E(f;D)
BAR, PRE AT D RA A BE OBS pl), HS RET SP
BA
BuiD)= ft) Av) payae , (26)
lo 30 om RMR IRE
acc(f;D) = LL pllf(@)=wele)de (27)
1- BUD).
23.2 SEH. B25P1
RS ALARA AL, (EASE LT ATES GR. AOL fl Ly AE
ARAB I, BRAN DAE RA EAT, NR, RR
RIASD RA BR. HAE RN AONE “MANE HS >
HORE” , RR TARE MS > MERE T OR” , ARERR
SUMBAT, my ae SE ATE.
ROMGCRET AAR. WebRRELRAPSA LB, SINE SR
Ht, RNBRAAG “MRE MP AS > eA RE” AD
FBX BANE BPH S WHR RT”. “AMER” (precision) 5 “HE
REESE CAME” (cecal) BENE FARR OER,
PFOA ALL, UDA PEALE HM B53 A
4} 8 SLIE PI (true positive), iE fi (false positive). HEAR Pi(true negative).
BR PI(false negative) MMH, & TP. FP. TN. FN SRA IR NEY
POM, ERA TP + FP+TN + FN ROBB. DRARI OME
BE” (confusion matrix)sn¥ 2.1 PAR.
HER PRA
OER
Ea SA
Eo | TP (haem) | PN (RA)
SA FP (£4) | TN (AKA)
AHA
PER P SRK RELY
rp
TP+FP?
(28)
TP
R= apyFN- (29)
PERADLRA— MF ER. — BOR, ER, REE
adi; TERE AE ENT, EMEA ECMO, OU, Ae a ab WF IR OT Re HG HH,
SY SS 38 TL A AR RS, A a, A Tt, 2.3 tiene
3
SUE AH OL
REAP RETA
Sea, TATE
ES, BAF.
a sk SPR He
“PREY
AWE SMPER, &
ABEL H MT IRC
fa eed PR a,
BAIA. CER,
ABS BK LRA.
Sa eT, QELS A aE MES we eee A, A A RE a AT FI IST
A, SY AT a PI, RRR RA 2b BEL, A
AEG. SE AEE NE BES, AY EMRE AE
LEAR ATE B, BRAT HDR 51-2 FRE HE AT ARLE AE, HELE TL
HYRAR STN “REPT AE” AIR GUOUREA, HECE RL EAE) BU “i
ARAL RE” AE TEP AREA, eb WUE A En TE SEAT SK BY LL
HCL SRSA AREA, DAME nM, RRR ARE, OS
THERES HR, (AK “PRM” , EATER “P-RAN” . FE
23 8H T TR E.
Oa
a
*oa
02
OT ss 9
eee
M28 PROS PATER
P-R PR BUMS aR 2) REARS BEE CERETT EE BE
It, AERA PR BE EO A BRE ALE”, NT
REET HN, GIMME 2.3 ESI RE A ANAL A BC 5 WR
PSS PAR RRA TR, HANH 2.3 Fe ASB, UAL — ARE T
PRBS: Fete A AY ANE ae a ae 4 ae oe ET LA. ART, HERS A
JER, MRM BAB He SM. RIN) bee a
FEUER P-R we FMA), CHEE EET A
LIA “RUE” TELL. (ERMA ACRE SSE, Blt, MABE fe
SSSSAER, HERA.
“3808” (BreakcEvent Point, fi) BEP)# LIRA —P AE BE, EE “te
WERE” MBH, IN 2.3 4 BE C AY BEP JE 0.64, TAL BEP
FRCL, UAE A TB 2 om AMR
{2 BEP BTML Te, WA PL AER:
2x PxR 2xTP
P\= "PER ~ WOER+TP-TN
(2.10)
FESS PS PERE A AE eA A HE
ALTER 4 .
ALTER ENS Rois, 7 RATHI IRA, Sa MHRA REAR A MBNA, HEB AE
moan) 3.4 HERE, WEBMARRRAD, BHR DAML, Meee
aot (3+3) HEM. Fl EM RIE — Fy, BEER ARI PEER / EE AN
nor(etn .
MH, BRERA a
Byala nant Fy = OB) xP HR an
b= ty($+2) @XP)+R *
Fe NSAP TRY” Sich 2 > 0 NER T Bae nt AER MOHTLNT EB [Van Rijsbergen, 1979]. B= 1
RAB P28 ye POE RMER PL; 8 > LRA ARM, 8 <1 AERA A
HRA VP Rin,
APM RSI REORRATHS AAMC, OAMOAT SAI / MR, MEUAEEL
AEE: BIE MRR TUN / BUR, ATT RT costso; ABAEMLANK, costor $5 costo HMR.
AID SPREE
RAL
HOR | BIR
KOK | 0 | costo
Hi | costo | 0
REKA SLANT.
"ALE (normalia-
tion) AMMEN
MOAN SHEAR EM
FILER 0,1) a
tO LIL
28,
2m MAMPI SRE
(Li +2 — eH, EAN A AB BR TART
GN (2.4) ITE ARACEAE “RL”, AAPA RRS
SEL AAS A IR. EES EARN, RAL A AAS AS BH RK
Bi, ALAR AG “REALE” (total cost). #542 2.2 PASS 0 FRE WIE
KR. WARCARK, & D* 5D” THRRAAR D HEAT RAR AF
4%, HU “PRUDHOE” (cost-sensitive) RAI
suo 3 ( 3 I(F (ws) # ui) x costor
D+
+ D Wile) Aw) vats) (223)
web
RADLAN, PD AEP A AR HER RAR, De th — SHE BE
DRO EA. 24 costs PA i, j AAR TO. 1, WATE RS
BEER YR ORR AE BE RE.
FETE SALON, ROC HAASE EL Be Re BR FE ART,
“ARPT HBR” (cost curve) KUTA BI Hi 0. Mt i AA BAH ENCE (0, 1)
ASTER ABT
px costo
PX costoy + (1 — p) X costo ”
P(t )eost = (2.28)
SCR p ALPEN Dy TE LEE; A LH AELA9 [0,1] PEALE
costo = ENRX Px costoy + FPR x (1 =p) x costo |
(2.25)
PX costoy + (1 — p) x costo J
AUP PPR 30 (2.19)F VME, PNR = 1— TPR RAR. (Ct ER
FUSE TIAEATAA: ROC Hee LE — ATL T OPT LAY ARR BE, BL ROC Ht
bb (TPR, FPR), ADAH PE ENR, ERCP a Lae
~4)\ (0, FPR) BI (1,FNR) MS Bt, AEE PAST AR AD Zea TR EP A
AMBARBT; SoutoHs ROC Heeb AEP cE He RP TERR, RS
SRT AAR BASF, FA A BR By EE EF 2) a I BR,
25 Sim. 24 Hes
Ed
AS Lp mite
58° BR Welle, 2010)
25 Rea MRE AR
2.4 bee
AAT SHAE HAE BEBE RE, EERE A ME EY HI MB HE ERE AT I
Ts HARA SHRED EIU 2) SEE REBERAER, MH
SRT HEAL. (AS RGR “HAR” WE? REARS HERE ER EI a“
Keb” 1? SiG, SLES) PHRASE RS.
BMP RIT BAR: He, RATA ROLE CHE, PTTL
TERME MRE EERE, A AR RAM; A,
SLM SRA SET RAKK, DORR AK DOMES
AAAAOAR, EAA MIR, FS OMA, MRR
HAA; BH, REBEAEAG A EMILE, OMS
REAM TARR ESM, RARWAATA. BA, ABABA TIE
AES] RPAH RESET EOE?
Sit BLM (hypothesis test) A RM BT FH HERE RAT
HE, MEPL Ah RTT AT HEMT HH, AF FEMU Eb OUR BIE Se A bb BE,
WA PREACHER ERR MER B, URI MERA S A. FL
BRAT A PP BA A PABBA, 8 SP EL SF ASL 2 ER
3k, AEP HS, ATTA CARRS ERE REM, A e eaR.
2.4.1 RA
A PS “AL” EET RRR A AAS, BH
Be = 0”. ASAEG PRAIA RET BE RR, FRAT IMR
BR 6. ZAR SMR AA, (EL, SAE AT REPEAL att AR ae
0.05. 01, 26% a ke
RAAT T I.
st a “subject to" #4
5 RRBRT ESD
fet a.
Sle MASUR
BA, FARE AREY AY RPE EAE ly PALL, ARAM BAR AA HE LH ER
Bi.
SEACH HUE] EMEA A Ec; MURR
BEE mm PREAH € xm RRSP. BLEUE AE IEA AT
PRCA, BAIN € MES BABS mn! MEARS, HK
ARREARS MBPS TE AEE (1 — e)™— 5 cE EH SHB € x m
PER A RUBE BABAR, AIA TELA m PAPAL, Ee,
PEAR ye A>] BAA PERE AY 6 HARES:
P(e) = G mena e-em (2.26)
EMAAR, RAF OP(E;€)/Oe = 0 ARI, P(e) TE BK, |e — eH
MY Plée) Web, REE (binomial) si, SHE 2.6 HAR, # = 0.3, M10
AEP 3 MERA RRR EK,
0.25:
0.20
0.05 _——
o 70
a 6
PR a
B26 =AaATEA(m = 10,¢=03)
RUTAT CHH “51428” (binomial test) FH e < 0.3" (BD “TEAR
BARAF 0.3” VAR RET, AAS, ABLE “We < eg” , SAE
1 cx ARR PF BE AN I HRC FE. HB 1 — oe ORT SIA
“HAREE” (confidence), EMME, HRT 2.6 PAR ABST,
=maxe st. oe (“)ea-ort 1
AU DUST Eo SI SLA LA,
BAVA AI MITE: SAREE AT Ah EF Nt, He
FRAME EAS FAR, BD = {(ei,y) Ly, A ay € RRL,
PRERUIUOLAEZE “FE” (order) 3H, PUMA REM LITRE BEE, BL BRAK Mik ibe,
NERS ARLE,
HERDER EE
ARE, £293 9.
AR RLRAP AAA
(savare los}
wr w feb,
ALARA MLRS,
RE AHL
BR Fay RAT oe
Diba BERT whe
De BeehR, Aw
soem.
FEM [0.0) ERK
Hans Sere e
FER AR sen OE
yesgss) © Aeztiea
Neb Ama iadikns
Co eannne
$2) = 2, EERE a,
RARE Oma,
AURA RCIA,
FaMPRAB MESA
DH Bat: Ba
SMM LAT 0, WA
AP Ret,
mom sete
RTE "AUR RE” “HR” PTE I (1.0, 0.0}, SARE “RE”
HORA “” IR” TTSEHL I {1.0,0.5, 0.0}; 25 AR HELI A FETE RE
Fi BEA kN, ULE hE, GRE “SR” EE
TL” “HTL” “SEL” BTHELH (0,0, 1), (0,1, 0), (1,0, 0).
RHA RRS
Sui) = way +b, AE F(as) =H (33)
HOT RSE w ALB WE? MLAS, KE a OTHE f(x) Sy ZIMA. 2.3 9
SPB, SAT RA (2.2) AE VES HE FE AE IA, RATT RS
TRB MG,
or) = 5 ( (ci) —
(w*o") orgmin Y (ai)
i 2
mas nin Do. wa — 6) (3.4)
SF RRA HITE UATE, CAL T AS FAL EASE A RO
FEE” (Buclidean distance). SF) 75 BEER Mt Ae sSe TAR FEN 77k
Dy “y=” (least square method). FELBHEGNT, Jeb sea LAR
RETA, (BTA PEATE LAUR EE Z FRU.
SRI w BIB AE Egy.) = Palys — way — 6)? AMG HIAERE, ALIEN
PALA “SAU” (parameter estimation). RATA Lye») 51
tw ALORS, BB)
-Ea-v2) 5 (35)
rot
Sige =2(m— Fn end) . (3.6)
REPHK(3.5)M3.6) AFA GA w A b AEM MA (closed-form) i
E wie
: (7) 32 sem
FHS RTABE
”
(38)
sea fF ay att
ATE AH FLERE D, Pea da od NE. TERA
RFA
Tay +b, HEE Flas) =U»
G06 “S7CHREENA” (multivariate linear regression).
PRADA AL FARR A) — FRE w A MATH. AMT, RAE we
Fb BAR RIE wh = (w;b), AILEY, TERRI D RAA—* m x (d+ 1)
KAMER X, PATA — a, BAT AT d TIC MT ARB a
JAE, SACRE 1, 2
f(ws)
Tn Z2 aq 1 af 1
xe fm me me alert
tmi Bat ve md 1) \@E 1
PAE PRIC SRA RIB y = (is 2; ---5 tm), WFR), A
vw" = arg min (y — Xv)" (y — Xt) (3.9)
& Ba = (y ~ Xie)" (y — Xe), Hed REGB
OF a 9x7 (xw—y) . (3.10)
on
SERWS AT ch ROMA, (Fe PE, Hee IE
BER, PRAT BT He
24 XTX yi ALA (full-rank matrix)BR IE 3 MEE (positive definite ma-
trix) Bt, @K(3.10) HST
(x™x) TxTy, (a)
Seb (XT) 1 ARAB (XTX) AMISH, ay = (wi, 1), ORE
