Kungani i-AI Namanje Ingakwazi Ukuxazulula Inkinga Yakho Yangempela Yokuthuthukisa Izibalo

ukusebenzisa i-AI ukuze wakhe imodeli yokwenza kahle yezibalo yenkinga yebhizinisi langempela, kungenzeka ukuthi uke wangena odongeni olufanayo: i-AI isebenza kahle ezibonelweni zezincwadi zokufunda futhi iyahlukana ngesikhathi uyinikeza idatha yakho yangempela kanye nenkinga yakho yangempela.

Leso sikhala asikona nje ukuqondana. Kungenxa yokuklama, futhi yisizathu esenza ngakhe i-ORPilot.

Isithembiso Sokuthuthukiswa Okunamandla e-AI

I-Operations Research (OR) ibilokhu inika amandla buthule ezinye zezinqumo ezinomthelela omkhulu ebhizinisini amashumi eminyaka – ukuhambisa amaloli ahambisa umzila, ukuhlela ukukhiqizwa kwefekthri, ukuklama amaketanga okuhlinzeka, ukwaba impahla kubathwali. Izibalo zivuthiwe futhi izixazululi zinhle kakhulu. Ibhodlela bekulokhu kuwubungcweti bomuntu obudingekayo ukuhumusha inkinga yebhizinisi ibe imodeli yezibalo.

Amamodeli Olimi Amakhulu (LLMs) abonakale eyisixazululo esiphelele. Ucwaningo olukhulayo, oluhlanganisa uchungechunge lwe-OptiMUS, OR-LLM, nezinye, lubonise ukuthi ama-LLM asezingeni eliphezulu angakhiqiza ikhodi yesixazululi elungile yezinkinga ze-linear programming (LP) kanye ne-mixed integer programming (MIP). Imiphumela ibonakale ihlaba umxhwele. Amademo ayephoqa.

Bese uzama ukusebenzisa elinye lalawa mathuluzi enkingeni yangempela, futhi imifantu ibonakala ngokushesha.

Lapho Amathuluzi Akhona Ephuka Khona

Cishe lonke ithuluzi le-LLM-for-OR elakhiwe kuze kube manje lihlanganyela umcabango ofihliwe: incazelo yenkinga iphelele, icacile, futhi inikezwa i-AI ngokwaziswa okukodwa, okufomethwe kahle ngayo yonke idatha eshumekwe kahle emgqeni.

Akuyona indlela izinkinga zangempela NOMA ezisebenza ngayo. Hhayi ngisho eduze.

Cabanga ngalokho okwenzekayo uma ithimba le-supply chain lifuna ukwakha imodeli yokuthuthukisa:

1. Incazelo yenkinga ayiphelele futhi ayicacile. Umhlaziyi webhizinisi uzothi “sifuna ukunciphisa izindleko zokuthutha” futhi akhohlwe ukusho ukuthi isikhungo ngasinye sokusabalalisa sinomkhawulo wokukhipha, ukuthi eminye imizila ayikho, noma ukuthi ukuvula indawo kuletha izindleko ezimisiwe zesikhathi esisodwa. Lokhu kweqiwe akukho
   ukunganaki. Kuyimibono umhlaziyi akubona kusobala, yingakho nje eyingozi. Isistimu ye-AI eqala ukwenza imodeli ngaphambi kokuba le mininingwane ibethelwe phansi ikhiqiza imodeli elungile ngokobuchwepheshe kodwa engalungile.
2. Idatha inkulu kakhulu ukuthi ingangena ekwazisweni. Inkinga yochungechunge lokuhlinzekwa kwempahla yangempela ingase ibandakanye amakhulukhulu ezindawo zokukhiqiza, izikhungo zokusabalalisa, amakhasimende, nezinkulungwane zemikhiqizo ngezikhathi eziningi. Ithebula lesidingo lilodwa lingase libe nezigidi zokungenela. Awukwazi ukushumeka lokho ngokushesha. Ngisho noma ungakwenza, ukugcwala kwewindi lomongo ngedatha eluhlaza kwandisa kakhulu ingcuphe yokubona izinto ezingekho.
3. Idatha onayo akuyona idatha edingwa yimodeli. Imodeli ingase idinge ibanga phakathi kwawo wonke amapheya ezindawo. Onakho itafula lezixhumanisi ze-GPS. Imodeli ingase idinge isidingo esihlanganisiwe ngomkhiqizo nesikhathi. Onakho ibhukwana lomsebenzi elinomugqa owodwa nge-oda ngalinye. Ukuvala leli gebe, okungukuthi ukwenza ikhompuyutha amapharamitha asuselwe kudatha eluhlaza, kuyisinyathelo esibalulekile sobunjiniyela okungekho thuluzi elikhona le-LLM-for-OR elisiphatha ngokuzenzakalelayo.
4. Uma usunemodeli esebenzayo, ukuphatheka nokukhiqizwa kabusha kubalulekile. Uma ufuna ukuqalisa kabusha imodeli kudatha ebuyekeziwe, shintsha usuka ku-Gurobi uye kusixazululi somthombo ovulekile, noma unikeze osebenza naye emshinini ohlukile imodeli, ubuyela esimweni sokuqala ngaphandle kwalapho ithuluzi likhiqiza i-artifact ehlala isikhathi eside, exazululi engaziwa. Amathuluzi amaningi akhiqiza ikhodi eqondene nesixazululi futhi akukho okunye.

Lawa akuwona amacala. Yizimo ezijwayelekile zanoma yimuphi umhlaba wangempela NOMA ukusetshenziswa. Amathuluzi akhona e-LLM-for-OR akhelwe umhlaba ohlukile, umhlaba wezincwadi zokufunda, futhi abonisa ama-seams awo ngenkathi eyishiya.

Sethula i-ORPilot

I-ORPilot iyi-ejenti ye-AI yomthombo ovulekile eyakhelwe kusukela phansi ukuya phezulu ngezimo zokukhiqiza. Ngokwazi kwami, ithuluzi lokuqala elisuselwa ku-LLM NOMA elidizayinelwe ngokusobala iqiniso elingcolile, elikhulu, elisindayo ledatha lokuthuthukiswa kwezimboni.

Amathuluzi amaningi e-AI okuthuthukisa agxumela aqonde ekubhaleni ikhodi lapho uchaza inkinga yakho. I-ORPilot yenza okuhlukile: ibuza imibuzo kuqala.

Leso sinqumo sokuklama, esibeka phambili ukuqonda ngaphezu kwejubane, sibonisa isimiso esisodwa esiqondisayo: i-ejenti ye-AI kufanele isebenze ngendlela efanayo nomuntu onekhono NOMA umeluleki angasebenza.

Umeluleki omuhle akahambi emhlanganweni wamakhasimende bese eqala ukubhala imodeli yezibalo ebhodini elimhlophe. Babuza imibuzo. Balalelisisa. Bahlehle uma kukhona okuthile
ayicacisi. Benza isiqiniseko sokuthi idatha isesimweni esifanele ngaphambi kokuthi kuqale ukumodela. Kuphela emva kwakho konke lokho bese becosha ipeni.

Ipayipi le-ORPilot libonisa lokhu kuziphatha ngezigaba ezinhlanu ezixhunywe ngokulandelana.

Isigaba 1: Umenzeli Wezingxoxo

I-ejenti yenhlolokhono yindawo yokungena. Ithola incazelo yakho yokuqala yenkinga yebhizinisi, okungenzeka ingacacile, ingaphelele, noma iziphikisa, futhi ikuhilele
ibhokisi elihlelekile ukuze ugcwalise izikhala. Umgomo oyinhloko wokuklama awukho ukumodela oqalayo kuze kuqedwe inhlolokhono.

I-ejenti iyalwa ukuthi ibone izikhala zolwazi encazelweni yamanje, buza okungenani umbuzo owodwa ocacisayo oqondisiwe ngesikhathi ngasinye (ukuze ugweme ukukukhungathekisa), futhi inqamule uma umsebenzi wenjongo, okuguquguqukayo kwezinqumo, izithiyo, kanye nezidingo zedatha konke kucaciswe ngokungaguquki.

Empeleni, lokhu kusho izingxoxo ezifana nalezi:

I-ORPilot: “Uma indawo isivuliwe, ingabe ihlala ivuliwe kuzo zonke izinkathi ezilandelayo, noma ingavalwa ngokuhamba kwesikhathi?”

I-ORPilot: “Ingabe le modeli iphatha uhlobo lomkhiqizo owodwa noma imikhiqizo eminingi?”

I-ORPilot: “Ukhulume ngezindleko zokuhamba.

Ngaphambi kokuqeda inhlolokhono, umenzeli wethula isifinyezo esigcwele esihlelekile esinomsebenzi wenhloso, okuguquguqukayo kwezinqumo, imikhawulo, amapharamitha, izinkomba, futhi ikunikeza ithuba lokulungisa noma yini ngaphambi kokuba leso sifinyezo sidluliselwe ezansi nomfula. Lona unogada omelene nemodi yokwehluleka evame kakhulu kumathuluzi we-LLM-for-OR: ukwenza imodeli yenkinga engalungile.

Isigaba 2: Umenzeli Wokuqoqwa Kwedatha

Lesi sigaba asinaso umlingani kumathuluzi amaningi akhona we-LLM-for-OR. Ingenye yezinto eziqanjiwe zesakhiwo ezibaluleke kakhulu ku-ORPilot.

Amathuluzi amaningi akhona e-LLM-for-OR acabanga ukuthi idatha ishumekwe embhalweni wenkinga, mancane ngokwanele ukuba angangena ekwazisweni. Ngezinkinga zezincwadi zokufunda, lokhu kuyasebenza. Ngezinkinga zangempela, ihlukana ngezindlela ezimbili. Okokuqala, amasethi edatha wangempela makhulu kakhulu. Isibonelo, inkinga ye-supply chain yamakhasimende angu-500, engu-500, yenkathi engu-12 ingaba nokufakiwe okufunwayo okungu-3,000,000. Okwesibili, ukushumeka idatha ngokushesha kukhulisa ingcuphe yokubona izinto ezingekho futhi kushise ngewindi lomongo kungenasidingo.

Impendulo ye-ORPilot ukuphatha idatha njengehlukene nokwaziswa ngokuphelele. Idatha ihlala kumafayela e-CSV. I-AI ifinyelela kuyo kuphela ngokubhala nokusebenzisa ikhodi. Umsebenzi womenzeli wokuqoqa idatha ukuthola ukuthi lawo mafayela e-CSV adinga ukubukeka kanjani.

Ngokusekelwe encazelweni yenkinga evela kumenzeli wenhlolokhono, umenzeli wokuqoqa idatha uyanquma:

• Imaphi amabhizinisi (amasethi) akhona kumodeli
• Yiziphi izibaluli (amapharamitha) ezidingwa yibhizinisi ngalinye
• I-schema enembile yethebula ngalinye elidingekayo: amagama ekholomu, izinhlobo, i-semantics

Yethula lokhu kucaciswa kuwe futhi ilinda uze unikeze wonke amafayela ngefomethi efanele. Iqinisekisa ukuphelela ngaphambi kokuqhubeka.

Okubi kakhulu, i-ejenti iyavumelana nezimo: uma ungenalo ucezu oluthile lwedatha elungele imodeli (ithi, imodeli idinga i-matrix yebanga kodwa unezixhumanisi ze-GPS kuphela), utshela umenzeli lokho onakho ngempela, futhi ibuyekeza i-schema ngokufanele — idlulisela igebe esigabeni esilandelayo ukuze isingathe.

Isigaba 3: I-Parameter Computation Agent

Cishe wonke amathuluzi akhona we-LLM-for-OR athatha amanani ezinombolo adingwa imodeli avela ngokuqondile kudatha enikezwe umsebenzisi. Empeleni, lokhu akulona neze iqiniso. Izibonelo ezimbili ezihlala zivela ezinkingeni zangempela NOMA:

• Imodeli yomzila wemoto idinga ibanga elihamba ngambili. Umsebenzisi unezixhumanisi ze-GPS. Ukubala amabanga e-Euclidean noma ezindawo kuwushintsho olungaphandle ngokuphelele kobubanzi bokwakhiwa kwe-LP/MIP.
• Imodeli yokukhiqiza yezikhathi eziningi idinga isidingo esihlanganisiwe ngesikhathi ngasinye. Umsebenzisi unencwadi yokuthenga enomugqa owodwa nge-oda ngalinye. Ipharamitha yemodeli iwukuhlanganisa isamba okufanele kubalwe kudatha eluhlaza.

I-ejenti yokubala yepharamitha ivala lesi sikhala ngokuzenzakalelayo. Ithola ukucaciswa kwenkinga namafayela e-CSV aluhlaza, bese:

1. Ikhomba ukuthi imaphi amapharamitha emodeli angakwazi ukufundwa ngokuqondile kumathebula akuhlaza
2. Ikhiqiza umbhalo wePython ukuze ubale lawo mapharamitha atholiwe
3. Isebenzisa iskripthi endaweni ene-sandboxed
4. Ibhala imiphumela njengamafayela e-CSV engeziwe, adluliselwe esinyathelweni sokumodela

Lokhu kuqinisekisa ukuthi ngesikhathi i-ejenti emodelayo ibona idatha, isuke ihlanzekile, ibhalwe ngendlela efanele, ifakwe ohlwini ngendlela efanele, futhi isilungele imodeli. Ekuhloleni kwethu, lesi sinyathelo sehlise kakhulu ukwehluleka kokukhiqiza ikhodi kanye nokubalwa kokuzama futhi.

Esinye isimo esivamile lapho i-ejenti yokubala yepharamitha ingaba usizo ukwenza ikhompuyutha amanani e-BigM. Kwezinye izivivinyo engizenzile ku-ORPilot, i-ejenti yokubala yepharamitha yenza ikhompuyutha inani le-BigM elidingekayo ukuze kube nemingcele exhumanisa okuguquguqukayo okuqhubekayo kwezinqumo zokuvula indawo kanambambili. Lena ipharamitha esuselwe engeke ikwazi ukucela umsebenzisi ukuthi ayinikeze ngokuqondile.

Isigaba 4: I-ejenti Yokukhiqiza Ikhodi

Ngokucaciswa kwenkinga ephelele, idatha eluhlaza, namapharamitha atholiwe konke kusesandleni, i-ejenti yokukhiqiza ikhodi ikhiqiza iskripthi esiphelele se-Python sokuxazulula i-backend oyikhethile. I-ORPilot okwamanje isekela ama-backends amahlanu: i-Gurobi, i-CPLEX, i-PuLP, i-Pyomo, ne-OR-Tools.

Ikhodi ekhiqiziwe isetshenziswa ngokushesha kubhokisi lesihlabathi. Uma kukhona okungahambi kahle: iphutha le-syntax, okuhlukile kwesikhathi sokusebenza, noma umphumela wesixazululi esingenakwenzeka/ esingakhawulelwe, umlayezo wephutha ogcwele kanye ne-traceback kubuyiselwa ku-LLM kanye nekhodi ekhiqizwe ngaphambilini. I-ejenti iyazama futhi, kufika enanini eliphezulu elingalungiselelwa umsebenzisi lemizamo.

Empeleni, ukwehluleka okuningi kuxazululwa phakathi nokuzama okukodwa noma okubili. Isizathu esibalulekile sokuthi i-ORPilot iphinde izame ukusebenza kahle ukuthi izigaba ezikhuphukayo sezivele zenze umsebenzi onzima: inkinga icaciswe kahle, idatha isilungele imodeli, futhi i-ejenti kuphela.
idinga ukulungisa iphutha leleveli yekhodi kunokuba icabange kabusha sonke isakhiwo semodeli.

Isigaba 5: I-ejenti yentatheli

Ngemuva kwesixazululo esiyimpumelelo, umenzeli wentatheli uhumusha imiphumela yezinombolo ngesiNgisi esilula, echaza ukuthi yiziphi izinsiza okufanele zivulwe, yiziphi izindlela okufanele zisetshenziswe, yimaphi amanani azokhiqizwa, ngolimi lwesizinda senkinga yebhizinisi yasekuqaleni, ukuze isetshenziswe umsebenzisi webhizinisi esikhundleni sochwepheshe NOMA.

Kungani Lo Myalo Ubalulekile

Ipayipi lilandelana ngamabomu. Isiteji ngasinye sifakwe kwesango esedlule siqedwa ngempumelelo. Inhlolokhono kufanele iphele ngaphambi kokuqala kokuqoqwa kwedatha. Idatha kufanele iqinisekiswe ngaphambi kokuqaliswa kokubala kwepharamitha. Amapharamitha kufanele alungele ngaphambi kokuthi ikhodi yenziwe.

Lokhu kulandelana kuvimbela imodi yokwehluleka evame kakhulu kumathuluzi asekelwe ku-LLM NOMA: amaphutha e-cascading lapho incazelo yenkinga engaqondakali isakazeka epayipini futhi ikhiqize ikhodi evumelekile ngokohlelo kodwa eyimodeli yenhloso engalungile.

Lokhu Kubukeka Kanjani Esikalini

Ngihlole i-ORPilot ezinkingeni ezimbalwa NOMA, enye yazo okuyinkinga yokuklama inethiwekhi yokuhlinzeka ngeziza ezingama-50 zokukhiqiza, izikhungo zokusabalalisa ezingama-50, amakhasimende angama-500, imikhiqizo engama-500, izikhathi eziyi-12. Imodeli eyaba umphumela yaba nezinqumo eziguquguqukayo ezingaphezu kwezigidi ezingu-9.7 kanye nezingqinamba ezingu-963,000. I-ORPilot isingathe ngempumelelo ukuphela kwepayipi eliphelele kuze kube sekugcineni, kusukela engxoxweni yokuqala ngokuqoqwa kwedatha, ukubala kwepharamitha, ukukhiqizwa kwekhodi, nokubikwa kwesixazululo, kukhiqizwe isixazululo esifanelekile ngeGurobi. Bheka iphepha lami lapha ukuze ubone imiphumela yezinkinga eziningi zokuhlola.

Ukuqalisa

I-ORPilot ingumthombo ovulekile futhi iyatholakala manje:

I-GitHub: https://github.com/GuangruiXieVT/ORPilot
Iphepha:

Ukufaka kuthatha imizuzu embalwa. I-ORPilot isekela i-OpenAI, i-Anthropic, i-Google, ne-DeepSeek njengabahlinzeki be-LLM, kanye ne-Gurobi, i-CPLEX, i-PuLP, i-Pyomo, ne-OR-Tools njenge-solver backends.

Kokuthunyelwe okulandelayo kulolu chungechunge, sizongena sijule ngokumela okuphakathi (IR) – i-artifact ye-JSON ye-solver-agnostic eyenza imiphumela ye-ORPilot iphinde ikhiqizeke futhi iphatheke kuzo zonke izindawo ezingemuva ngaphandle kokufonela i-LLM futhi. Hlala ubukele!