Federal gosudarstvennoe uchebnoe predpriyatie Chair of the System of Artificial Intelligence

3 Neuroimitators
yazyke. For bolshinstva zarubejnyx imitators or package programs, v
vÿÿÿvali problemy pri rabote - nedostatochna sovmestimost s
93
Machine Translated by Google

---

Pervonachalno vÿpolnyayutsya operatsii, svyazannÿe so sborom dannyx, perevod
ix v ÿÿÿÿÿÿÿÿÿÿÿ ÿÿÿ (ÿÿÿÿ ÿÿÿ ÿÿÿÿÿÿÿÿÿÿÿÿ ÿÿ ÿÿÿÿÿÿÿÿ
neuroimitators. Neuroimitator NeuroPro version 0.25 in nastoyashchee vremya
razrabotchika.
ispolzovaniya bolee gibkix i moshchnyx metodov. Tem ne menee, dannyy
program bÿla vnedrena vo mnogix uchebnyx zavedeniyax i legko dostupna v
Here is a brief description of the methods used by the general public
i pozvolyaet bÿstro poluchit predvaritelnÿe rezultaty nachinayushchimi
vozmojno in Paradox). Field file dannyx doljny bit chislovogo tipa
posobiya rekomenduyut obratitsya v Internet, gde programma vylojena na
reshaet mnogie zadachi daje s ustanovkoy parametrov po umolchaniyu. Vybora
used neuroimitator NeuroPro v. 0.25 (razrabotchik V.G. Tsaregorodtsev). Dannÿy
poryadok ne yavlyaetsya edinstvenno vozmojnÿm, v
processing, but can be useful for identification records.
predstaviteley nevozmojno izbejat.
v sebya bolee polnuyu i adekvatnuyu (s tochki zreniya matematicheskoy
deystviy, poleznÿy v pervuyu ochered dlya studentov ili nachinayushchix
obshchego file dannyx dve vÿborki: obuchayushchuyu i testovuyu, kotorye
schitaet, chto neyroimitator NeuroPro ustarel i ukazÿvaet na neobxodimost
ispolzuemyx neyronnyx setey. Nije
predlozheno tolko kratkoe opisanie nekotoryx iz ukazannyx
system. The description of this imitator is based on the most important text
setey, i pozvolyayushchiy bÿstro poluchit predvaritelnÿe rezultaty.
ne podderjivaetsya i ne soprovojdaetsya razrabotchikom. Of course, this
3.1 Features of working technologies with NeuroPro
programmnyy product shiroko rasprostranen, realizuet osnovnÿe operatsii
nositele). Chashche predpolagaetsya ispolzovanie format DBase (hotya)
issledovatelyami. To receive distributive programs by the author of uchebnogo
On the Internet on many sites, has an interface in Russian, good
tehnologii obrabotki dannyx s pomoshchyu neyrosetey s uchitelem. For this
(tselye or deystvitelnye). Simvolnÿe polya ne ispolzuyutsya pri
Pered zapuskom neuroimitatora nujno sformirovat na osnove
and the use of NeuroPro as one of the most convenient and available
chastnosti, sushchestvuet avtorskiy «ritual» obrabotki dannyx, vklyuchayushchiy
desyatkax saytov ÿÿÿ ÿÿÿÿÿÿÿÿÿÿÿ ÿÿÿÿÿÿÿ. In
dannom sluchae rassmatrivaetsya samÿy uproshchennÿy poryadok
The choice of Neurogenesis is based on the interesting architecture vstraivaemoy
udoben, no k nemu predÿyavlyayut pretenzii iz-za polnosvyaznoy arxitektury
statistics) rekomenduemuyu posledovatelnost obrabotki dannyx. Author
issledovateley, weakly acquainted with iskusstvennyx neuronnyx apparatus
94
Machine Translated by Google

---

zapisey (chetnÿe - in one file, nechetnÿe - in drugoy). More for convenience
nepodgotovlennogo user:
rekomenduetsya, t. k. mojet potrebovatsya ispolzovat testovuyu vyborku v
testovoy vyborke.
V. Perekrestnoe obuchenie i testirovanie. G.
Dopolnitelnÿe operatsii (determination of the parameters of parameters,
uproshchenie seti, verbalization). Pri kajdom iz etix podkhodov pervonachalnÿy
poryadok deystviy po
Sleduet predvaritelno vybrat pole otveta - odno iz poley
sluchae metki t i o budut skoree sbivat s tolku, chem pomogat. Tags 1 and 2
indifferent (not opredelyayut konkretnÿy vid vyborki); b) for the maintenance of the
network and the results of the use of the department
rekomenduetsya sozdavat otdelnÿe catalogs (folders). Prichem, for
otveta vÿbirayut s uchetom mneniya spetsialista v predmetnoy oblasti).
Recommendations:
chego bÿl sozdan i ispolzovalsya proekt AVS - testirovaniya, libo obucheniya
sootvetstvii s obshchey sovokupnostyu dobavleniem posle nazvaniya tsifr 1
(obuchayushchaya) ili 2 (testovaya). For example, if the file is named dannyx
odnogo vyxodnogo polya sozdannÿe katalogi doljny byt vlojennymi. For example,
for the preservation of results with the use of external fields
proektov setey. Bukvy t i o v nazvaniya faylov-vÿborok dobavlyat ne
vÿbranÿ polzovatelem, dlya nachala rekomenduetsya otbor chetnyx / nechetnyx
oshibok i utochnit rezultaty v sluchae neobxodimosti. Let's
consider three simple options of methods for
otdelnyx parameters (otdelnÿe polya ne ispolzuyutsya pri obuchenii) pri
opisaniya nazovem obuchayushchuyu vÿborku - vÿborka 1, testovuyu - vÿborka 2.
A. Obuchenie po obshchey
vyborke. B. B. Prostoe obuchenie po obuchayushchey vyborke i testirovanie po
kachestve obuchayushchey (naprimer, when perekrestnom testirovanii). In etom
vyxodnogo polya, libo provedeniya dopolnitelnyx eksperimentov
(parameters) file dannyx, kotoroe trebuetsya prognozirovat (obÿchno pole
sozdaniyu i soxraneniyu nachalnogo sostoyaniya neyronnyx setey (proekta)
yavlyaetsya odinakovÿm (p. 40, 41). In dannoy metodike predlagaetsya ispolzovat
sleduyushchie
Rekomenduetsya vesti kratkiy protokol vsex deystviy. For example, for
a) obuchayushchaya i testovaya vyborki (otdelnÿe file dannyx) nazvanÿ v
dopolnitelnyx eksperimentov libo izmenenii struktury seti v predelax
on 1-y vyborke. Such a protocol will allow in the future to avoid many
pomeshchayutsya in otdelnÿe files. Criteria otbora zapisey mogut byt
svod.dbf, to pervichno obuchayushchaya vÿborka poluchit nazvanie svod1.dbf,
testovaya - svod2.dbf. It okajetsya poleznÿm v dalneyshem pri soxranenii
X6, sozdan katalog «X6», esli nuzhno provesti eksperimenty s udaleniem
95
Machine Translated by Google

---

dobavlyaetsya letter «b». For example, the project n_ob2 contains a set of obuchennye po
vxodnyx). Esli nujno provesti analogichnye eksperimenty, ispolzuya
kolichestvo setey nikak ne obosnovano, no esli net ogranicheniy po vremeni,
rekomenduetsya ispolzovat ne menee 10. Otkrÿvaem isxodnÿy file dannyx, sozdaem
novuyu set. Pri etom dlya kajdoy seti opredelyayutsya vxodnÿe polya i
v) the project after the first creation of a network (set without obuchenÿ) called
n_ish (isxodnÿy), zdes «n» - uslovnoe oboznachenie reshaemoy zadachi; g) proekt s
obuchennymi po odnoy iz vÿborok setyami nazvan v
j) vozmojny bolee slozhnÿe variant, sootvetstvuyushchie kombinatsii
answer - the last. Esli nujno ispolzovat drugoe pole, to poslednee pole
neskolko poley v kachestve polya otveta. Sushchestvuet two settings
vyborke);
d) for kontrastirovannyx setey v naimenovanie dobavlena bukva «k», when
kontrastirovanii vnutrenney struktury dopolnitelno ispolzuyutsya
vxodov i binarizatsii.
Primernyy plan vÿpolneniya raboty: 1.
Sozdat neskolko neyrosetey odinakovoy struktury
kontrastirovaniya sinapsov, nazvan n_oks1, obuchennÿy po obshchey vyborke
Kolichestvo setey opredelyaetsya issledovatelem, vliyaet na stepen
umolchaniyu ustanovlen «kolichestvennÿy», this type sootvetstvuet resheniyu
ogranicheniyax rekomenduetsya ne menee 3. Statisticheski trebuemoe
directory «VIS_X7-9» (parameters X7, X8, X9 are excluded from the number
e) pri binarizatsii vesov sinapsov v naimenovanie proekta
tochnost resheniya zadachi (by default ustanovleno znachenie «0,1»). V
structure set, otlichnuyu ot zadannoy po umolchaniyu (3 layers of 10 neurons), for
example, 1 layer of 30 neurons, sozdaem catalog «1_30» i t. p .;
vyborke svod2 posle binarizatsii;
field answer (vyxodnoe field). In NeuroPro on umolchaniyu naznachaetsya field
sleduet otmetit vxodnym, t. k. neuroimitator can be used
sootvetstvii s nomerom vyborki: n_o1 (seti obucheny po vyborke svod1), n_o2 (seti
obucheny po vyborke svod2), n_o (seti obucheny po obshchey
neskolkix operatsiy, naprimer: n_okb2 - proekt, soderjashchiy seti, obuchennÿe po
vÿborke svod2, posle posledovatelnogo sokrashcheniya chisla
pervÿe bukvÿ sootvetstvuyushchix elements: «s» - synapses, «n» - neurons, «u» -
neodrorodnyx inputs, «r» - ravnomernoe uproshchenie seti. For example, proekt,
soderjashchiy obuchennÿe po vyborke svod1 seti posle
(File / Create).
vyxodnogo field: «kolichestvennÿy (neprerÿvnÿy)» and «kachestvennÿy». Po
posok sokrashcheniya chisla vxodov - n_ok;
tom je vyxodnom pole X6, sleduet vnutri kataloga «X6» zavesti
reliability of results. Minimalnoe quantity at vremennyx
assignments prediktsii (predskazaniya). Pri etom mojno dopolnitelno vybrat
96
Machine Translated by Google

---

donor, stroke number, plasmapheresis number, student code i t. p.), ispolzovat na vxode
bessmÿslenno, t. k. oni ne nesut sushchestvennoy
(dispersions, srednekvadratichnogo otkloneniya i t. p.). Dannaya velichina
yavlyaetsya vxodnÿm dlya seti i pole yavlyaetsya vyxodnÿm dlya seti (pole otveta).
Nachinayushchim issledovatelyam mojno nachinat s nastroek seti i haraktera
samostoyatelnymi i imeyut otdelnÿe znacheniya znachimostey.
ogranicheniy i osobennostey zadachi, rekomenduetsya ne menyat dannoe
sluchae daje zadachi klassifikatsii mojno uproshchenno traktovat kak zadachi
NeuroPro neskolko otlichaetsya from prediktsii: rezultaty testirovaniya
konkretnoy task i parameter trebuetsya vÿpolnyat dopolnitelnÿy
otdelno dlya kajdogo klassa. Krome togo, pomimo traditsionnyx protsentov
diapazone, chto na praktike sootvetstvuet zadacham klassifikatsii. In etom
settings typov polya. When choosing a type «kachestvennÿy» poyavlyaetsya stolko
ukazat otnoshenie predshestvovaniya znacheniy dannyx sostoyaniy. For
nastroyke. Naimenovaniya dopolnitelnyx poley obrazovany iz isxodnogo s
quantity and percentage of primerov, opredelennyx neuverenno.
Sleduet otmetit, chto sushchestvuyut polya, kotorye ne rekomenduetsya
uchityvat pri obrabotke dannyx, tak kak dannÿe polya schitayutsya
znacheniy dannogo polya libo rascheta osnovnyx statisticheskix characteristic
ispolzuetsya setyu (avtomaticheski pomechayutsya simvolnÿe polya), pole
a) polya, predstavlyayushchie soboy poryadkovÿe nomera (naprimer, kod
vliyaet na kachestvo testirovaniya seti. Esli net dopolnitelnyx
vyxodnogo polya (kolichestvennoe), ustanovlennyx po umolchaniyu. In general
Otmetim, chto realizatsiya operatsiy dlya zadach klassifikatsii v
information; sushchestvuyut situatsii osmÿslennosti takix poley, naprimer, esli oni nesut
informatsiyu o vremennom poryadke, poetomu dlya kajdoy
predstavleny razvernuto - ne tolko srednie kolichestvo i protsent, no i
znachenie tochnosti (ostavit znachenie po umolchaniyu). Type «kachestvennÿy» implies
discretion of changing the value of the parameter in nebolshom
predictions.
Dlya vxodnyx poley takje sushchestvuyut analogichnÿe vyxodnÿm polyam
analysis;
sluchae dopolnitelno poyavlyaetsya vkladka «diskretnÿe sostoyaniya polya X», where
ukazanÿ chislo diskretnyx sostoyaniy i ix konkretnÿe znacheniya, mojno
dopolnitelnyx poley, skolko sostoyaniy polya (klassov) ukazano v
i kolichestva verno i neverno opredelennyx primerov privodyatsya
lyubogo polya vÿbiraetsya odno iz vozmojnyx sostoyaniy: pole ne
obshchem sluchae tochnost vÿbiraetsya s uchetom haraktera raspredeleniya
dobavleniem nomera klassa (naprimer, isxodnoe pole X23 imeet 3 sostoyaniya,
poyavyatsya sleduyushchie polya: X23_1, X23_2, X23_3). This follows
ispolzovat setyu pri obuchenii:
97
Machine Translated by Google

---

kopiy odnogo proekta) nachinayut obuchenie s odnoy i toy je tochki
parameters menshe 5 - field answer 0, inache - 1), ne ispolzuyutsya setyu, t. k.
pri etom obÿchno poluchaetsya lineynaya zadacha, kotoraya reshaetsya trivialno
trebuetsya, i obÿchno vÿbirayut structure, zadannuyu po umolchaniyu.
Characteristics of neurons affects the speed of learning the set, chem bolshe
poryadkom raspolojeniya setey in proekte. Identichnost nachalnogo
mojet davat luchshee kachestvo prognoza. Only for bolshinstva sluchaev
spaces): gelatelno libo zapolnit ix (esli takoe vozmojno), libo ne
operational identichnost nedopustima, seti doljny byt raznymi. Obÿchno, otdelnÿe
proekty soderjat seti, sformirovannÿe dlya odnix usloviy (kak
obucheniya, budet ne odinakovÿm, esli ispolzovat razlichnÿe naborÿ
obshchem sluchae eto ne vliyaet na rezultat i proizvoditelnost, no inogda
NeuroPro vÿrezanie pustyx zapisey proisxodit v avtomaticheskom rejime. Zapis
ne ispolzuetsya pri obuchenii i testirovanii, esli ne zapolneno
X6). Under the project ponimaetsya file in the internal format NeuroPro (* .npp), v
users in dependence on konkretnoy zadachi. By umolchaniyu c
Pervonachalnoe soxranenie proekta v neskolkix kopiyax (seti ne
obespechivaetsya eshche pri sozdanii (in obshchem sluchae odin proekt mojet
(tojdestvennost) nachalnogo sostoyaniya prostranstva setey v sootvetstvii s
opredelyayushchie pole otveta (for example, if pole summÿ opredelennyx
bolshinstva prikladnyx issledovaniy bolshix neyronnyx setey ne
sostoyanie setey budet izmenyatsya, no isxodno «parallelnÿe» seti (iz
and does not require the use of neural networks.
Sleduet uchityvat i factor nepolnotÿ dannyx (v dannyx imeyutsya
znachenie characteristics, then medlennee mojet obuchatsya set, no pri etom
sostoyaniya proektov ne obyazatelna. For some specific tasks or
mnogomernogo prostranstva. Konechnoe sostoyanie, dostignutoe posle
in dannoy metodike), naprimer, obuchennÿe po odnoy i toy je vyborke. V
ispolzovat v kachestve vxodnyx pustÿe (polupustye) zapisi i polya. V
mojno ispolzovat po umolchaniyu ustanovlennoe v NeuroPro znachenie, ravnoe
0,1. 2. Save this project (for example, n_ishx6 - for exit field
vxodnyx dannyx. Sootvetstvenno rezultaty, poluchennÿe pri dannom
hotya by odno znachenie polya.
The structure of the neural network (the number of layers and neurons) is determined
kotorom xranyatsya vse sozdannÿe seti.
gives some preimushchestva. Razumeetsya, seti in odnom proekte razlichny,
daje po nachalnomu sostoyaniyu. In NeuroPro uniqueness kajdoy seti
NeuroPro uses 3 layers of 10 neurons in each layer. For
b) polya, vÿchislyaemÿe na osnove drugix poley i odnoznachno
obuchenÿ) i ego dalneyshee ispolzovanie obespechivaet identichnost
soderjat seti, raznÿe po strukture i dannym). V dalneyshem pri obuchenii
98
Machine Translated by Google

---

primerov for neuroset. If this indicator is about 60%, then
results in vide tekstovyx files (save from main menu), svodnÿe
It is necessary to pay attention to the percentage correctly defined
with obuchennymi setyami on vyborke
2. 12. Provesti perekrestnoe testirovanie, for this:
osnovaniem.
udovletvoritelnye. Esli 70–80% - good, 90% bolee - excellent. Esli protsent
correctly opredelennyx primerov menshe 50%, to sleduet
rezultaty testirovaniya i, sootvetstvenno, kachestvo obucheniya seti
podxode dlya raznyx vÿborok, mojno sravnivat i sopostavlyat s bolshim
ix preobrazovaniyu, i tolko zatem povtorno primenit dannuyu metodiku
obshchey
vyborke. For variant B vÿpolnit
sleduyushchee: 3. Soxranit proekt pod drugim imenem
(naprimer, n_o1). 4. Otkrÿt vÿborku 1 (obuchayushchuyu). 5.
Obuchit vse neyroseti. 6. Save the project under that name.
7. Provesti testirovanie, for this: 7.1. Open selection 2 (test).
7.2. Protest all vseroseti. 7.3. Save all test results:
predvaritelnye
(libo obratitsya k bolee kvalifitsirovannomu spetsialistu za pomoshchyu). Dlya
varianta V nujno vÿpolnit sleduyushchee: 3. Soxranit proekt pod drugim
imenem (naprimer, n_o1). 4. Open the selection 1. 5. Obuchit vse
neyroseti. 6. Soxranit proekt pod tem je imenem - imeem soxranennyy
proekt s
Dlya varianta A dalee vÿpolnit sleduyushchee:
3. Soxranit proekt pod drugim imenem, esli dalee predpolagaetsya
use variant B (e.g., n_o). 4. Open obshchuyu
vyborku. 5. Obuchit vse neyroseti. 6.
Save the project under that name. V
rezultate imeem soxranennyy proekt s
obuchennymi neyrosetyami po
bolee detail issledovat dannÿe, provesti dopolnitelnÿe operatsii po
results in tablitsy (primernyy vid tablitsy - tabl. 3.1).
7. Otkrÿt isxodnyy proekt, poluchennyy pri vÿpolnenii p. 2. 8. Save
the project under the drug name (for example, n_o2). 9. Open the
selection 2. 10. Obuchit vse neyroseti. 11. Save the project under
that name - save the saved project
obuchennymi setyami on vyborke 1.
99
Machine Translated by Google

---

93
77
5.50
5.30
kol-vo% kol-vo%
results in vide tekstovyx files (save from main menu), svodnÿe
1.50
Set 3
TV2
1.52
vyborke, protestirovannÿe po vtoroy - v pervoy chasti tablitsy (pervye N setey do jirnoy
linii), obuchennÿe po vtoroy vÿborke, protestirovannÿe
Set 5
35.50
68.70
29,39
TV1
1.58
Set 2
primerov
Set 1
Table 3.1. - Test results
5.44
Set 6
85
results in vide tekstovyx files (save from main menu), svodnÿe
Max.
oshibka
12.1. Open the project with obuchennymi setyami on vyborke 1 (see p. 3).
12.2. Open selection 2. 12.3. Protestirovat po ney vse neyroseti. 12.4. Save
all test results: predvaritelnye
64.50
1.55
72.90
testirovaniya mojno predstavit v edinoy tablitse: obuchennÿe po pervoy
32,44
85
7.44
Neverno reshennyx
78.4
OV1
OV2
5.79
5.10
12.5. Open the project with obuchennymi setyami on vyborke 2 (see p. 8).
12.6. Open selection 1. 12.7. Protestirovat po ney vse neyroseti. 12.8. Save
all test results: predvaritelnye
70.61
32,44
180
1.79
1.60
29.92
ispolzuetsya dlya udobstva identifikatsii testirovaniya vÿborok.
69
73.66
5.09
191
68
169
oshibka
67.56
74.05
67.56
82
5.70
26.34
je tablitse, no otdeliv jirnoy liniey). Results perekrestnogo
185
results in tablitsy (primernyy vid tablitsy - tabl. 3.1).
25.95
Set 8
1.63
primerov
on the first - on the second frequency (N setey posle jirnoy linii). Jirnaya line
71
31.30
193
5.47
177
Srednyaya
Set 7
1.65
27.10
rezultaty v vide tablitsy (primernyy vid tablitsy - tabl. 3.1, mojno v toy
Srednee 183.6 70.08
Set 4
1.61
177
Verno reshennyx
194
ÿ
ÿ
100
Machine Translated by Google

---

1. Open the project with obuchennymi setyami on obshchey vyborke (variant
A, p. 3). 2. Determine the values of the input parameters for each set
testovoy dlya setey 5–8 (TV2); vÿborka 2 ispolzuetsya tolko dlya obucheniya
Delenie
zapisi)
in kachestve testovoy (TV1).
Dlya perekrestnogo testirovaniya ispolzuyutsya dve vÿborki, poluchennÿe iz
obshchey sovokupnosti dannyx. The most simple criterion
vyborka
1
3. Save all values for each set of text files. Otmetim, chto dannÿe
pokazateli (znachimost vxodnyx parametrov), kak pravilo, net smÿsla
opredelyat, esli pri testirovanii set pokazÿvaet
in two groups: prosteyshie (definition of parameters of parameters, verbalization)
and operatsii, svyazannÿe s uproshcheniem seti (kontrastirovanie, binarizatsiya).
G1. Definition of parameters parameters. After obucheniya mojno
formiruetsya dva podmnojestva: podmnojestvo 1 (naprimer, chetnÿe zapisi),
podmnojestvo 2 (nechetnÿe zapisi). Ispolzovanie dannyx podmnojestv dlya
vyborka
Table 3.2. - Predstavlenie ispolzovaniya i deleniya vyborok pri
1
(naskolko kajdyy iz nix znachim dlya prognoza vyxodnogo polya).
Podmnojestvo 2 (naprimer, nechetnÿe
obucheniya pervyx 4-x setey ispolzuetsya vÿborka 1 (OV1), kotoraya yavlyaetsya
Ispolzovanie
vyborki. For this sleduet vÿpolnit sleduyushchee:
setey 5–8 (OV2), sootvetstvenno dlya pervyx chetyrex setey ona ispolzuetsya
Obuchayushchaya
2
(Neuroset / Znachimost vxodnyx signalov seti).
G. Dopolnitelnÿe operatsii v svoyu ochered mojno uslovno razdelit
otbora yavlyaetsya vydelenie chetnyx / nechetnyx zapisey. At this
Testovaya
perekrestnogo testirovaniya analogichno opisannomu dlya vyborok (tabl. 3.2).
Podmnojestvo 1 (naprimer, chetnye zapisi)
provesti opredelenie pokazateley znachimostey vxodnyx parameters
perekrestnom testirovanii
Primechanie: in dannyx eksperimentax OV1 = TV2, TV1 = OV2 - for
2
Pokazateli znachimosti opredelyayut tolko dlya obuchennyx setey, poetomu
logichno opredelyat ix znacheniya posle obucheniya i dlya obshchey
101
Machine Translated by Google

---

sluchaev rangi sovpadayut so srednimi znacheniyami znachimostey parameters, no
inogda otlichayutsya. Srednee znachenie mojet byt nebolshim