Tsis yog, zoo, tau kawg, kuv tsis mob hnyav. Yuav tsum muaj kev txwv rau qhov uas nws muaj peev xwm ua kom yooj yim rau ib qho kev kawm. Tab sis rau thawj theem, nkag siab cov ntsiab lus yooj yim thiab nrawm " nkag mus" cov ncauj lus, nws yuav ua tau. Peb mam li sib tham txog yuav ua li cas kom raug lub npe cov ntaub ntawv no (kev xaiv: "Kev kawm tshuab rau dummies", "Cov ntaub ntawv tsom xam los ntawm cov pawm", "Algorithms rau cov me nyuam") thaum kawg.
Rau qhov taw tes. Sau ob peb daim ntawv thov kev pab cuam hauv MS Excel rau kev pom thiab pom kev sawv cev ntawm cov txheej txheem uas tshwm sim hauv ntau txoj kev kawm tshuab thaum txheeb xyuas cov ntaub ntawv. Pom yog kev ntseeg, tom qab tag nrho, raws li cov neeg coj ntawm kab lis kev cai hais tias, uas tsim feem ntau ntawm cov txheej txheem no (los ntawm txoj kev, tsis yog txhua tus ntawm lawv. Qhov muaj zog tshaj plaws "txhawb vector tshuab", lossis SVM, txhawb vector tshuab yog lub invention ntawm peb compatriot Vladimir Vapnik, Moscow lub koom haum ntawm kev tswj.
Peb cov ntaub ntawv rau kev tshuaj xyuas
1. K-txhais tau tias pawg
Cov teeb meem ntawm hom no xa mus rau "kev kawm tsis muaj kev saib xyuas," thaum peb yuav tsum faib cov ntaub ntawv thawj zaug rau hauv qee yam ntawm pawg paub ua ntej, tab sis peb tsis muaj tus lej ntawm "cov lus teb raug"; peb yuav tsum muab rho tawm ntawm cov ntaub ntawv nws tus kheej. . Qhov tseem ceeb classical teeb meem ntawm nrhiav subspecies ntawm iris paj (Ronald Fisher, 1936!), uas yog suav hais tias yog thawj kos npe rau ntawm qhov kev paub no, tsuas yog qhov xwm.
Txoj kev yog heev yooj yim. Peb muaj ib pawg ntawm cov khoom uas sawv cev ua vectors (sets ntawm N tus lej). Hauv irises, cov no yog pawg ntawm 4 tus lej uas qhia lub paj: qhov ntev thiab qhov dav ntawm sab nrauv thiab sab hauv lobes ntawm perianth, feem (). Qhov kev ntsuas Cartesian ib txwm raug xaiv raws li qhov kev ncua deb, lossis ntsuas qhov sib thooj ntawm cov khoom.
Tom ntej no, pawg chaw raug xaiv randomly (lossis tsis randomly, saib hauv qab), thiab qhov kev ncua deb ntawm txhua yam khoom mus rau pawg chaw raug suav. Txhua yam khoom ntawm ib kauj ruam iteration tau raug cim tias yog nyob rau qhov chaw ze tshaj plaws. Tom qab ntawd qhov chaw nruab nrab ntawm txhua pawg yog pauv mus rau tus lej lej ntawm kev sib koom ua ke ntawm nws cov tswv cuab (los ntawm kev sib piv nrog physics, nws tseem hu ua "center of mass"), thiab cov txheej txheem rov ua dua.
Cov txheej txheem converges sai heev. Hauv cov duab hauv ob qhov ntev nws zoo li no:
1. Thawj qhov kev faib tawm ntawm cov ntsiab lus ntawm lub dav hlau thiab tus naj npawb ntawm pawg
2. Txheeb xyuas cov chaw hauv pawg thiab muab cov ntsiab lus rau lawv pawg
3. Hloov cov kev sib koom ua ke ntawm cov chaw hauv pawg, rov suav cov kev koom tes ntawm cov ntsiab lus kom txog thaum cov chaw ruaj khov. Lub trajectory ntawm pawg pawg txav mus rau nws txoj haujlwm kawg yog pom.
Thaum twg los xij, koj tuaj yeem teeb tsa cov chaw tshiab (tsis muaj kev faib cov ntsiab lus tshiab!) thiab pom tias cov txheej txheem muab faib tsis yog ib txwm tsis meej. Kev ua lej, qhov no txhais tau hais tias rau kev ua haujlwm tau zoo (cov lej ntawm cov squared nrug ntawm cov ntsiab lus mus rau qhov chaw ntawm lawv pawg), peb pom tsis yog lub ntiaj teb, tab sis qhov tsawg kawg nkaus hauv zos. Qhov teeb meem no tuaj yeem kov yeej los ntawm qhov kev xaiv tsis yog qhov kev xaiv thawj zaug ntawm pawg thawj coj, lossis los ntawm kev suav cov chaw muaj peev xwm (qee zaum nws yog qhov zoo los tso lawv raws nraim ntawm ib qho ntawm cov ntsiab lus, ces tsawg kawg yog qhov lav tias peb yuav tsis tau khoob. pawg). Nyob rau hauv txhua rooj plaub, ib tug finite set ib txwm muaj ib tug infimum.
(tsis txhob hnov ββββqab qhib kev txhawb nqa macro. Cov ntaub ntawv tau raug txheeb xyuas cov kab mob)
Kev piav qhia ntawm txoj kev ntawm Wikipedia -
2. Kwv yees los ntawm polynomials thiab cov ntaub ntawv tawg. Kev cob qhia
Cov kws tshawb fawb zoo tshaj plaws thiab nrov ntawm cov ntaub ntawv tshawb fawb K.V. Vorontsov piav qhia luv luv txog kev kawm tshuab raws li "kev tshawb fawb ntawm kev kos duab nkhaus los ntawm cov ntsiab lus." Hauv qhov piv txwv no, peb yuav pom tus qauv hauv cov ntaub ntawv siv qhov tsawg tshaj plaws squares txoj kev.
Cov txheej txheem ntawm kev faib cov ntaub ntawv los rau hauv "kev cob qhia" thiab "tswj" tau qhia, nrog rau cov xwm txheej xws li kev rov ua haujlwm, lossis "rov kho dua" rau cov ntaub ntawv. Nrog rau qhov tseeb kwv yees, peb yuav muaj qee qhov yuam kev ntawm cov ntaub ntawv kev cob qhia thiab qhov yuam kev me ntsis ntawm cov ntaub ntawv tswj. Yog tias tsis yog, nws ua rau muaj kev hloov pauv rau cov ntaub ntawv kev cob qhia thiab ua yuam kev loj ntawm cov ntaub ntawv xeem.
(Nws yog qhov paub zoo tias los ntawm N cov ntsiab lus ib tus tuaj yeem kos ib qho nkhaus ntawm N-1th degree, thiab cov qauv no hauv cov ntaub ntawv dav dav tsis muab qhov xav tau. )
1. Teem thawj qhov kev faib tawm
2. Peb faib cov ntsiab lus mus rau "kev cob qhia" thiab "tswj" hauv qhov piv ntawm 70 txog 30.
3. Peb kos qhov kwv yees nkhaus raws cov ntsiab lus kev cob qhia, peb pom qhov yuam kev nws muab rau cov ntaub ntawv tswj
4. Peb kos ib qho nkhaus nkhaus los ntawm cov ntsiab lus kev cob qhia, thiab peb pom qhov yuam kev loj heev ntawm cov ntaub ntawv tswj (thiab xoom ntawm cov ntaub ntawv kev cob qhia, tab sis dab tsi yog qhov point?).
Qhia, tau kawg, yog qhov kev xaiv yooj yim tshaj plaws nrog ib qho kev faib ua "kev cob qhia" thiab "tswj" subsets; nyob rau hauv rooj plaub no, qhov no yog ua ntau zaus rau qhov zoo tshaj plaws kev hloov ntawm cov coefficients.
Qhib macros rau kev ua haujlwm raug
3. Gradient qhovntsej thiaj tsis mob thiab kev hloov pauv ntawm kev ua yuam kev
Yuav muaj 4-dimensional rooj plaub thiab linear regression. Linear regression coefficients yuav txiav txim ib kauj ruam los ntawm kauj ruam siv txoj kev gradient qhovntsej thiaj tsis mob, thaum pib tag nrho cov coefficients yog xoom. Ib daim duab cais qhia qhov kev hloov pauv ntawm kev ua yuam kev txo qis vim cov coefficients raug kho ntau dua thiab raug. Nws muaj peev xwm mus saib tag nrho plaub 2-dimensional projections.
Yog tias koj teem lub gradient qhovntsej thiaj tsis mob loj heev, koj tuaj yeem pom tias txhua zaus peb yuav hla qhov tsawg kawg nkaus thiab yuav tuaj txog ntawm qhov tshwm sim ntawm ntau cov kauj ruam, txawm hais tias thaum kawg peb tseem yuav tuaj txog (tshwj tsis yog tias peb ncua lub qhovntsej thiaj tsis mob. ntau - ces lub algorithm yuav mus "hauv spades"). Thiab daim duab ntawm qhov yuam kev nyob ntawm qhov iteration kauj ruam yuav tsis du, tab sis "jerky".
1. Tsim cov ntaub ntawv, teem caij gradient qhovntsej thiaj tsis mob
2. Nrog rau kev xaiv kom raug ntawm gradient qhovntsej thiaj tsis mob, peb smoothly thiab sai ncav cuag qhov tsawg kawg nkaus
3. Yog hais tias qhov gradient qhovntsej thiaj tsis raug xaiv, peb overshoot qhov siab tshaj plaws, qhov yuam kev graph yog "jerky", convergence yuav siv sij hawm ntau ntawm cov kauj ruam.
ΠΈ
4. Yog tias peb xaiv qhov gradient qhovntsej thiaj tsis mob, peb txav deb ntawm qhov tsawg kawg nkaus
(Yuav kom rov tsim dua cov txheej txheem siv cov txheej txheem gradient qhovntsej thiaj tsis mob tshwm sim hauv cov duab, kos lub thawv "cov ntaub ntawv siv").
Raws li cov zej zog hwm, puas yog qhov yooj yim thiab kev nthuav qhia cov khoom siv tau? Puas tsim nyog txhais cov lus ua lus Askiv?
Tau qhov twg los: www.hab.com