Share
Embed
Transcript
4 slides
-
EconGrowthAgent: LLMΤʔδΣϯτͱܦࡁཧʹجͮ͘ ϚΫϩܦࡁγϛϡϨʔγϣϯ EconGrowthAgent: Macroeconomic Simulation based on LLM-powered Agents
and Economic Growth Theory Լ ᚠจ 1 Terufumi Morishita ֻ֯ ਖ਼ 1 Masaya Tsunokake ࢁޱ ಞق 2∗ Atsuki Yamaguchi Ӭ௩ ޫҰ 1 Koichi Nagatsuka ༑ ޫ 1 Hikaru Tomonari ඌ ֶ 1∗ Gaku Morio ࠓҰ म 1 Osamu Imaichi ेՏ ହ߂ 1 Sogawa Yasuhiro 1 ཱ࡞ॴ ݚڀ։ൃάϧʔϓ Research & Development Group, Hitachi, Ltd. 2 γΣϑΟʔϧυେֶ The University of Sheffield Economic growth is an essential phenomenon that brings prosperity to human society. To analyze economic growth, we propose “EconGrowthAgent”, an economic simulation environment with LLM agents. According to macroeconomics, economic growth occurs in two stages: 1. Economic agents, including households and firms, make economic decisions, such as labor, consumption, savings, investment, and 2. The interactions of these decisions in production dynamics lead to increased goods production. Therefore, Econ- GrowthAgent models economic agent decision-making through LLM agents and incorporates their interaction dynamics. We conducted a 20-year economic simulation on EconGrowthAgent using 100 GPT-4o agents and confirmed its ability to replicate economic growth and related phenomena. This demonstrates the validity of EconGrowthAgent. Additionally, we simulated scenarios of “transition to an extremely small government” and “approach of an extinction-level asteroid” to analyze their impacts on economic growth. The ability to freely examine such scenarios, which would be difficult to verify in the real world, provides significant practical value. 1. ͡Ίʹ ܦࡁֶɼਓྨࣾձΛΑΓ๛͔ʹ͢Δ͜ͱΛతʹɼܦࡁݱ ͷղ໌Λԡ͠ਐΊ͖ͯͨɽ19 ੈلޙʹ࢝·ΔϛΫϩܦࡁֶ ͰɼཧతͳΞϓϩʔνʹΑΓɼফඅऀاۀͱ͍ͬͨܦࡁओ ମͷҙࢥܾఆΛੳͨ͠ɽফඅऀͷҙࢥܾఆͱͯ͠ʮຖ݄ͷ ऩೖΛੜ׆අɾޘָඅɾஷۚͳͲʹ͢Δʯ ɼاۀͷҙࢥܾఆ ͱͯ͠ʮ (ࡒ) ͷੜ࢈ྔՁ֨Λܾఆ͢Δʯͱ͍ͬͨྫ͕ ڍ͛ΒΕΔɽ͜͏͍ͬͨҙࢥܾఆΛੳ͢ΔͨΊʮ֤ܦࡁओମ ͷརӹ (ްੜ) ΛදΘ͕ؔ͢࠷େԽ͞ΕΔΑ͏ͳબ͕ͳ͞Ε Δʯͱ͍͏Α͏ͳཧͱͯ͠ϞσϧԽ͞Εͨɽ ཧతͳΞϓϩʔν͚ͩͰɼݱ࣮ੈքͱͷ߹ੑΛݕ ূͰ͖ͳ͍ɽ20 ੈلલʹൃలͨ͠ϚΫϩܦࡁֶͰɼ ؍ଌσʔλ͔ΒͷΞϓϩʔνΛ༻͍ͯܦࡁݱΛੳͨ͠ɽ͢ͳ ΘͪɼݸʑͷܦࡁओମͷҙࢥܾఆʹཱͪೖΔ͜ͱͳ͘ɼGDP ࣦۀͱ͍ͬͨूܭతͳ؍ଌσʔλͷؔੑΛϞσϧԽͨ͠ɽ GDP ͱࣦۀͷٯ૬ؔΛࣔ͢ʮΦʔΫϯͷ๏ଇʯ ɼΠϯϑϨ ͱࣦۀͷؔΛࣔ͢ ʮϑΟϦοϓεۂઢʯ ͳͲ͕දతͰ͋Δɽ ୯ʹ؍ଌʹΑΓಘΒΕΔσʔλΛੳ͢Δ͚ͩͰͳ͘ɼࣗΒ࣮ݧ ΛߦͬͯσʔλΛऔಘ͠ɼ ੳ͢Δ͜ͱͰ͖Δɽ 20 ੈلޙʹൃ లͨ͠ήʔϜཧߦಈܦࡁֶͰ࣮ݧ࣮ࣨݧʹΑΔΞϓϩʔν ͕Ֆ։͍ͨɽྫ͑ਓؒͷઓུతࢥߟೳྗΛੳ͢ΔͨΊɼڭ ࣨʹूΊͨੜెʹʮनਓͷδϨϯϚʯͷΑ͏ͳγφϦΦΛϓϨ Πͤ͞ɼಘΒΕͨߦಈσʔλΛੳ͢Δɽ ࣮ݧ࣮ࣨݧʹܽଘࡏͨ͠ɽ·ͣɼ࣮ݧʹਓతίετ ͕͔͔ΔͷͰɼ࠶ݱ࣮ݧ݅Λม࣮͑ͨݧΛେྔʹߦ͏͜ ͱ͕Ͱ͖ͳ͍ɽ·ͨಉ༷ͷཧ༝͔Βɼܦࡁओମͷ͕૿͑ͨ ߹ͷ࣮ݧ͍͠ɽߋʹɼڭࣨͰͷ࣮ݧͰຊ࣭తʹݕূͰ͖ ͳ͍γφϦΦଘࡏ͢Δɽྫ͑ɼ ʮܦࡁࡦΛ X ʹมߋͨ͠ ߹ͷ GDP ͷӨڹʯɼͦͷنͱෳࡶੑ͔Βݱ࣮తͳ ݕূ͕ࠔͰ͋Γɼ ʮ৽ܕײછ Y ͷຮԆʹΑΔࣾձͷӨڹʯ ɼྙཧత؍͔Β࣮ݧࣗମ͕ڐ༰͞Εͳ͍ɽ ܭࢉػՊֶͷͰൃలͨ͠ΤʔδΣϯτϕʔευϞσϦϯ ά (Agent-Based Models, ABM) [Gilbert 2008] ɼ͜ΕΒͷܽ Λࠀ͢ΔՄೳੑΛ༗͢ΔɽABM ͰɼݸʑͷҙࢥܾఆओମΛ ͨ͠ϓϩάϥϜ (=ΤʔδΣϯτ) Λܭࢉػ্ʹ༻ҙ͠ɼԾత ͳ࣮ݧΛߦ͏ɽਓతͳ੍Λड͚ͳ͍ͨΊɼେྔͷ࣮ݧओମ Λ૿࣮ͨ͠ݧ͕ՄೳͱͳΓɼ·ͨʮ͠ Xʯ ʮ͠ Yʯͱ͍ͬ ͨγφϦΦࣗ༝ʹݕূͰ͖ΔɽҰํͰɼABM Λܦࡁֶʹద༻ ͢Δʹ՝͋ΔɽABM ͷΤʔδΣϯτ୯७ͳϓϩάϥϜ (ϧʔϧ) ʹج͖ͮҙࢥܾఆΛߦ͏ҰํͰɼܦࡁֶʹ͓͚Δओମ Ͱ͋Δਓؒͷҙࢥܾఆෳࡶ͔ͭଟ໘తͰ͋Δɽͭ·Γɼϧʔ ϧϕʔεΤʔδΣϯτͰܦࡁओମΛۙࣅ͢Δ͜ͱͰ͖ͳ͍ɽ ࿈བྷઌ: Լᚠจɼ[email protected] ۙొͨ͠େنݴޠϞσϧ (LLM) ͜ͷ՝Λղܾ͢Δ ՄೳੑΛൿΊ͍ͯΔɽLLM ɼਓ͕ؒॻ͍ͨେنσʔλ͔Β ਓؒͷҙࢥܾఆύλʔϯΛֶश͍ͯ͠ΔͨΊ [Horton 2023, Xie 2024]ɼܦࡁओମͨΔਓؒΛߴਫ਼ʹۙࣅͨ͠ΤʔδΣϯτͱ͠ ͯͷׂ͕ظͰ͖ΔɽΑͬͯɼLLM ΤʔδΣϯτΛ༻͍ͨ ABM (LLM-ABM) ɼܦࡁݱΛੳ͢Δ৽ͨͳΞϓϩʔνͱ ͯ͠ɼܦࡁֶมֵΛͨΒ͢Մೳੑ͕͋Δɽ ͜ͷํʹ࠷ॳͷҰาΛ౿Έग़ͨ͠ͷɼ[Li 2024] ΒͰ͋Δɽ Li Β LLM ΤʔδΣϯτʹΑͬͯՈܭΛϞσϧԽ͠ɼ࿑ಇ ফඅʹؔ͢ΔҙࢤܾఆΛͤͨ͞ɽ݁ՌɼLLM ΤʔδΣϯτ ʮCOVID ͕ྲྀߦ͍ͬͯΔͷͰ৬ʹߦ͖ͨ͘ͳ͍ʯ ʮऩೖ͕ݮͬ ͨͷͰࢧग़ʹؾΛ͚Α͏ʯͱ͍ͬͨਓؒʹ͍ۙҙࢤܾఆΛ ߦ͑Δ͜ͱ͕֬ೝ͞Εͨɽ·ͨɼଟͷ LLM ΤʔδΣϯτΛ༻ ͍࣮ͨݧʹΑΓɼϑΟϦοϓεۂઢΦʔΫϯͷ๏ଇͱ͍ͬͨϚ Ϋϩܦࡁֶͷදతͳ๏ଇ͕࠶ݱͰ͖Δ͜ͱΛ֬ೝͨ͠ɽ ͔͠͠ͳ͕Β [Li 2024] Βͷ݁Ռɼܦࡁֶʹ͓͍ͯ࠷ॏཁ ͳݱͰ͋ΔʮܦࡁʯΛ࠶ݱ͢Δ͜ͱ͕Ͱ͖͍ͯͳ͔ͬͨɽ ۩ମతʹɼҰࠃͷ๛͔͞ͷࢦඪͰ͋Δ GDP ( ࡒͷੜ࢈ྔ) ͕ ҰఆʹͱͲ·͍ͬͯͨɽܦࡁɼ͋ΒΏΔܦࡁओମͷްੜ Λେ෯ʹҾ্͖͛ΔͨΊɼܦࡁݱͷதͰ֨ผʹॏཁͩͱΈ ͳ͞Ε͖ͯͨ [Smith 1776,Jr. 1988,Landes 1998]ɽྫ͑ɼถࠃ ܦࡁաڈҰੈلͷؒʹฏۉ 2%ఔͰ͠ɼෳརͷྗΛ आΓͯɼ͓Αͦ 10 ഒͷنʹ·Ͱͨ͠ɽ͜Ε͢ͳΘͪɼ 100 લʹൺͯɼզʑ͕ফඅͰ͖Δࡒͷྔ͕ 10 ഒʹͳͬͨ͜ ͱΛҙຯ͓ͯ͠Γɼզʑͷްੜେ෯ʹ্͍ͯ͠Δ ∗1ɽ ͦ͜ͰຊݚڀͰɼܦࡁΛදݱͰ͖Δ LLM-ABM ͷߏங Λࢦ͢ɽ·ͣɼઌߦݚڀʹ͓͍ͯܦࡁΛଊ͑Δ͜ͱ͕Ͱ͖ ͳ͔ͬͨཧ༝Λ୳ΔͨΊɼϚΫϩܦࡁֶͷҰͰ͋Δʮܦࡁ ཧʯΛ๚ΕΔɽܦࡁཧʹΑΔͱɼܦࡁʹͱͬͯॏ ཁͱͳΔཁૉɼੜ࢈ʹ༻͍Δಓ۩Ͱ͋ΔʮࢿຊετοΫྔʯ ɼ ࿑ಇऀͷख़࿅Ͱ͋Δʮਓతࢿຊʯ ɼࠃͷʮٕज़ਫ४ʯͳͲͰ͋ Δɽ͜ΕΒཁૉ͕اۀࠃͷࢿʹΑΓ͠ɼͦΕʹ͍ࡒ ͷੜ࢈ྔ͕૿Ճ͢Δ͜ͱͰɼܦࡁ͕࣮ݱ͢Δɽ[Li 2024] ͷ γϛϡϨʔγϣϯɼ͜ͷΑ͏ͳܦࡁཁૉͷಈֶΛؚΜͰ ͍ͳ͔ͬͨͨΊɼܦࡁΛදݱͰ͖ͳ͔ͬͨͷͰ͋Δɽ ࣍ʹɼ͜ͷߟΛ౿·͑ɼܦࡁཁૉͷಈֶΛऔΓࠐΜͩ LLM-ABM Ͱ͋ΔʮEconGrowthAgentʯ(ਤ 1) Λߏங͢Δɽ۩ ମతʹɼܦࡁओମΛ LLM ΤʔδΣϯτʹΑͬͯϞσϧԽ͠ɼ ͦͷҙࢥܾఆ͔ΒܦࡁͷಈֶΛɼҎԼͷΑ͏ʹ࣮ݱ͢Δ: (1) ՈܭΤʔδΣϯτຖظɼ࿑ಇྗΛఏڙ͠ɼऩೖͷҰ෦Λஷ ͱͯۜ͠ߦʹ༬͚Δɽ(2) اۀΤʔδΣϯτ࿑ಇྗͱࢿຊε τοΫΛ༻͍ͯࡒΛੜ࢈͢Δɽߋʹɼۜߦ͔ΒࢿۚΛିΓ͏͚ɼ ՃͷࢿຊετοΫΛߪೖ͢Δɽ(3) ࢿຊετοΫͷ૿ՃʹΑΓ ∗1 ٯʹɼզʑͷͷճΓͷࡒɼྫ͑৯ྐɾҥɾޘָɾҩྍͳͲ͕ेͷҰ ʹͳͬͨੈքΛߟ͑ͯΈΑ͏ɽ͓ͦΒ͘ɼ͋·ΓʹෆศͰת͑ΒΕͳ͍ͩΖ͏ɽ 1 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105 -
ৱম崡崰崫崗 ܭ௧ ൛স ݓ௧ ۞ߨ ਹဋ ગଦী ઘ 峇峘峘ৱস峘島 ଃ嵣ௌ௮峘島
ੇੑ ਿ ੫ ଃ峘েਓ ࢅ࢚ = ࡲ(ࡷ࢚ , ࡸ࢚ ) ৽ੋਛশ ࢅ࢚ା = ऐ ࡷ࢚ା , ࡸ࢚ା > ࢅ࢚ ৱম崡崰崫崗峘ੜਸ ࡷ࢚ା = ࡷ࢚ + ࢤࡷ࢚ ાܥ௧ 峁હ岻 ଃ峘েਓ ࢅ࢚ ਿ੍লܩ௧ ௌ௮ ࡸ࢚ ฅ嵣ৱ峘島 ฅࡿ࢚ ৱࡵ࢚ 峕ൻཉ峼૦峐峉岮峔岞 峫峍峋峭௮岮峐ฅস峃峵峈 ؒ*'3 ਤ 1: EconGrowthAgent ͷུ֓ਤɽ ੜ࢈͕ޮԽ͞Εɼࡒͷ࢈ग़ྔ͕૿Ճ͢Δɽͳ͓ࠓճɼܦࡁ ཁૉͷதͰ࠷ॏཁͳࢿຊετοΫͷΈΛ࣮ͨ͠ɽ ࠷ޙʹɼ100 ମͷ GPT-4o ΤʔδΣϯτΛ༻͍ͨ 20 ؒͷγ ϛϡϨʔγϣϯΛߦ͍ɼܦࡁͱͦΕʹؔ࿈͢ΔܦࡁݱΛ࠶ ݱͰ͖Δ͜ͱΛ֬ೝͨ͠ɽ·ͨɼLLM ΤʔδΣϯτͷҙࢥܾఆ͕ ਓؒͷͦΕʹ͍ۙ͜ͱ֬ೝͨ͠ɽ͜ΕΒɼEconGrowthAgent ͷଥੑΛࣔ͢ɽ࣍ʹʮখ͍͞ͷݖަʯ ʮٿյ໓ Ϋϥεͷᯁੴͷۙʯͱ͍͏γφϦΦΛص্ݕূ͢Δɽ͜ͷΑ ͏ͳݱ࣮Ͱͷݕূ͕ࠔͳγφϦΦΛؚΊɼଟ༷ͳγφϦΦΛ ࣗ༝͔ͭେྔʹ࣮ݧͰ͖Δ͜ͱɼEconGrowthAgent ͷ༗༻ੑ Λ͍ࣔͯ͠Δɽͳ͓ιʔείʔυΛެ։༧ఆͰ͋Δ ∗2ɽ 2. Կ͕ܦࡁΛܾΊΔͷ͔ʁ ϚΫϩܦࡁֶͷʮܦࡁཧʯʹج͖ͮɼܝΛٞ͢Δɽ 2.1 ੜ࢈ؔ Ұࠃ͕͋ΔظؒʹੜΈग़͢ࡒͷੜ࢈ྔΛ Y ͱ͢ΔɽY ࠃ ͷ๛͔͞Λࣔ͠ɼGDP ͱҰରҰʹରԠ͢Δɽੜ࢈ྔ Y ࿑ಇ ྔ L ࢿຊετοΫྔ K ʹΑܾͬͯ·Δͱߟ͑ΒΕΔɽ࿑ಇྔ (labor) L ɼ࿑ಇਓޱ × ࿑ಇ࣌ؒͰ͋ΔɽࢿຊετοΫ (kapital) K ͱɼੜ࢈ʹ༻͢Δಓ۩Ͱ͋Γɼྫ͑ࣄһ͕͏ PCɼ ৯ϝʔΧʔ͕͏৯ՃϥΠϯɼݐઃձ͕ࣾ͏ΫϨʔϯɼ ͳͲͰ͋ΔɽܦࡁֶͰɼ͜ΕΒͷؔΛੜ࢈ؔ F Ͱද͢: Y = F(K, L) (1) ؆୯ͷͨΊίϒɾμάϥε (Cobb–Douglas) ؔΛߟ͑Δ: F(K, L) = A0 × K K0 α L L0 1−α , α ∈ (0, 1) (2) K0, L0, A0 దͳఆͰ͋Δɽίϒɾμάϥεؔɼزͭ ͔ͷॏཁͳੑ࣭Λຬͨ͢ɽ ∂F ∂K > 0, ∂F ∂L > 0 (୯ௐ૿Ճੑ) (3) ͢ͳΘͪɼK L Λ૿ͤੜ࢈ྔ Y ૿͑Δɽ࣍ʹ F(zK, zL) = zF(K, L) (نʹؔͯ͠ऩ֭Ұఆ) (4) ͢ͳΘͪɼK ͱ L Λಉ࣌ʹ z ഒʹ૿ͨ͠߹ɼY z ഒʹͳ Δɽྫ͑ɼ৯ՃϥΠϯΛ 2 ഒʹ૿͠ɼ͔ͭ࡞ۀऀ 2 ഒʹ૿ͨ͠߹ɼ৯ͷੜ࢈ྔ 2 ഒʹͳΔɽ࠷ޙʹ ∂2F ∂2K < 0, ∂2F ∂2L < 0 (ऩ֭చݮ) (5) ޙऀΛߟ͑ΔͱʮࢿຊετοΫ K Λݻఆͨ͠··࿑ಇྗ L Λ૿ ͢߹ɼL ͕େ͖͘ͳΔ΄ͲɼՃͷ L ͕ͨΒ͢ੜ࢈ྔͷ ∗2 https://github.com/hitachi-nlp/econ-growth-agent ૿ ∂F/∂L ͕খ͘͞ͳΔʯͱ͍͑Δ (લऀಉ༷)ɽ৯Ճ ϥΠϯΛ૿͞ͳ͍··࡞ۀऀ͚ͩΛ૿͢ͱɼϥΠϯ͋ͨΓ ͷ࡞ۀऀ͕૿͑ͯඇޮԽ͢ΔͨΊɼՃͷ࡞ۀऀ͕ͨΒ ͢ੜ࢈ྔͷ૿ݮ͍ͬͯ͘ɽ 2.2 ஷͱࢿ ੜ࢈͞Εͨࡒͷ༻్ 3 ͭʹྨͰ͖Δ: ҰൠͷՈܭʹΑΔফ අ (consumption) CɼاۀʹΑΔࢿ (investment) IɼʹΑΔ ެڞతͷࢧग़ (goverment expenditure) GɼͰ͋Δ: Y = C + I + G (6) Ոܭͷফඅ C ɼ৯ࡐҥͷߪೖɼཱྀߦͳͲͰ͋Δɽاۀͷ ࢿ I ɼকདྷͷੜ࢈Λ૿ͨ͢ΊͷࢿຊετοΫͷߪೖΛࢦ ͢ࢧग़ G ɼٛڭҭͷͨΊͷڭһޏ༻ࣾձอোඅ༻ͳ ͲͰ͋Δɽ(6) Λมܗ͢Δͱɼผͷݟํ͕Ͱ͖Δ: S def = Y − (C + G) = I (7) ͜ͷࠃɼੜ࢈ྔ Y ͷ͏ͪ (C + G) ফඅࡒͱͯ͠ɼͦͷظͷ ͏ͪʹফඅͯ͠͠·͏ɽΓͷஷ (saving) S = Y − (C − G) ɼདྷظҎ߱ʹճͤΔࡒͰ͋Δɽ͜Ε͢ͳΘͪࢿຊετοΫ ࡒʹ૬͠ɼاۀ͕কདྷͷੜ࢈Λ૿ͨ͢Ίͷࢿ I ͱͯ͠ੜ ࢈ͨ͠ͷͰ͋Δɽͭ·Γ (7) ɼ ʮظͷஷ S ͕ɼདྷظҎ߱ ͷࢿ I ʹճΔʯͱ͍͏ؔΛ͍ࣔͯ͠Δɽ ஷ S ΛҎԼͷΑ͏ʹ୯७Խ͢Δɽ·ͣɼଟ͘ͷࠃʹ͓͍ ͯࢧग़ G Ոܭফඅ C ʹൺͯখ͍͞ͷແࢹ͠ɼS = Y − (C + G) ∼ Y − C ͱ͢Δɽ࣍ʹɼՈܭॴಘͷ͏ͪͷҰఆ ׂ߹ c Λফඅʹճ͢ͱԾఆ͠ɼ͔ͭɼՈܭͷॴಘࡒͷੜ࢈ྔ Y ͱ࿈ಈ͢Δ͜ͱΛ౿·͑ɼՈܭফඅ C Λ C = cY (8) ͱͯ͠ϞσϧԽ͢ΔɽҎ্ʹΑΓ: S = (1 − c)Y = sY (9) ͭ·Γੜ࢈ྔ Y ͷ͏ͪஷ s def = (1 − c) ͷׂ߹͕ஷʹճΔɽ 2.3 ࢿຊετοΫͷੵͱܦࡁ લઅ·Ͱ͋Δݻఆ͞Εͨ࣌ࠁ t ʹ͓͚ΔҰࠃͷܦࡁΛੳ ͨ͠ɽ͜͜Ͱ࣌ࠁ t Λಈ͔͠ܦࡁͷ࣌ؒൃలΛهड़͢Δ: Solow–Swan Ϟσϧ [Solow 1956,Swan 1956] Yt = F(Kt, Lt) (10) St = sYt = It (11) Kt+1 = Kt + I0 × It − δ × Kt ΔKt (12) ·ͣɼظͷࢿຊετοΫྔ Kt ͱ࿑ಇྔ Lt ͕ɼظͷੜ࢈ ྔ Yt ΛܾΊΔ (10)ɽ࣍ʹɼ͜ͷੜ࢈ྔͷ͏ͪফඅ͞Εͣʹͬ ͨஷ St = sYt ͕ɼࢿຊετοΫͷࢿ It ͱͳΔ (11)ɽདྷ ظͷࢿຊετοΫྔ Kt+1 ɼظͷࢿຊετοΫྔ Kt ʹର ͯ͠ɼࢿʹΑΔ૿ I0 × It (I0 ఆ) ΛՃ͑ͨͷͱͳΔ (12)ɽͨͩ͠ɼ࣌ؒܦաʹΑͬͯҰ෦ δ × Kt ຎ͢Δ (δ ݮ Ձঈ٫ఆ)ɽ ଟ͘ͷࠃͰݮՁঈ٫ δ ͕খ͍͞ͷͰɼΔKt > 0 ͢ͳΘͪ Kt+1 > Kt ͕Γཱͭɽ࿑ಇྗҰఆ Lt+1 = Lt = L0 ΛԾఆ͢ Δͱɼੜ࢈ؔ F ͷ୯ௐ૿Ճੑ (3) ʹΑΓɼདྷظͷ࢈ग़૿Ճ ͢Δ: Yt+1 = F(Kt+1 , Lt+1 ) > F(Kt , Lt ) = Ytɽ͜ͷΑ͏ʹɼ Solow-Swan Ϟσϧɼஷ=ࢿʹΑΔࢿຊετοΫͷੵΛ ௨ͯ͡ɼܦࡁΛઆ໌͢ΔཧͰ͋Δɽ 3. EconGrowthAgent ઌߦݚڀ [Li 2024] ͷ LLM-ABM ɼલઅͰͨ͡ࢿຊετο Ϋؔ͢ΔܦࡁಈֶΛߟྀ͍ͯ͠ͳ͔ͬͨͨΊɼܦࡁΛ දݱͰ͖ͳ͔ͬͨɽͦ͜Ͱզʑɼ͜ͷܦࡁಈֶΛऔΓࠐ Μͩ LLM-ABMʮEconGrowthAgentʯ(ਤ 1) ΛఏҊ͢ΔɽՈܭ اۀͱ͍ͬͨܦࡁओମΤʔδΣϯτ͕ࣗతʹҙࢥܾఆ͠ɼ ͦͷ݁Ռ͕࿑ಇྔ Ltɼஷ StɼࢿຊετοΫ Ktɼࡒͷੜ࢈ྔ Yt = F(Kt , Lt ) ͱ͍ͬͨྔʹू͞ΕΔɽஷ St ࢿຊετο Ϋ Kt Λͤ͞ɼͦΕʹ͍ࡒͷੜ࢈ྔ Yt ૿Ճ͢Δɽ 2 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105 -
3.1 ՈܭΤʔδΣϯτ ՈܭΤʔδΣϯτɼ໊લɾྸɾ৬ۀͱ͍ͬͨϖϧιφ͕ ϥϯμϜʹϓϩϯϓτͱׂͯ͠ΓͯΒΕͨ LLM Ͱ͋ΔɽՈܭ a ∈ [1, N]
(N ΤʔδΣϯτ) ɼຖ݄ (t Ͱද͢)ɼҎԼͷҙ ࢥܾఆΛߦ͏ɽ ·ͣʮ࿑ಇ࣌ؒ la t ʯΛܾఆ͢ΔɽՈܭ࿑ಇͷ݁Ռɼ݄ऩ ia t = wa t × la t ΛಘΔ (wa t ࣌څ)ɽ࣍ʹʮࡒͷߪೖ ca t ʯΛܾఆ͢Δɽ ݁Ռɼ݄ऩ ia t ͷ͏ͪ ca t × pt Λফඅ͢Δɽ؆୯ͷͨΊɼࡒ̍ छྨͷΈͱ͠ɼͦͷՁ֨Λ pt ͱͨ͠ɽߋʹՈޫඅͱ͍ͬ ͨݻఆඅ fa t ফඅ͞ΕΔɽΓͷஷ sa t = ia t − ca t × pt − fa t ۜߦͷ༬ۚߴ ba t ʹՃ͞ΕΔ: ba t+1 = ba t + sa t ɽ Ոܭ͜ΕΒҙࢤܾఆͷࡍʹɼաڈ t − τ ͔Βݱࡏ t ·Ͱͷܦ ࡁྔ {l˜ t , w˜ t , c˜ t , p˜ t , f˜ t , b˜ t , Y˜ t |˜ t ∈ [t − τ, t]} Λࢀর͢Δɽ 3.2 اۀΤʔδΣϯτͱΤʔδΣϯτ ࠓճ୯७ͳϧʔϧϕʔεͰ࣮ͨ͠ɽ·ͣɼاۀ͕ੜ࢈͢Δ ࡒͷྔ Yt ɼੜ࢈ؔ (16) ʹΑͬͯఆ·Δͱ͢Δɽاۀ͕Ոܭ a ʹࢧ͏ۚ wa t ɼॴ༩ͷॳظۚ wa t=0 Λച্ߴ pt × Dt ͷͱಉͰ૿ݮͤ͞Δ (Dt ࡒͷधཁ 3.4 અ)ɽͷ ׂߟ͑ͣɼࢧग़ G = 0 ͱͨ͠ɽ 3.3 ܦࡁཧͷͱͷ౷߹ ܦࡁओମΤʔδΣϯτͷҙࢥܾఆ (3.1 અ,3.2 અ) ͱ Solow-Swan Ϟσϧ (2.3 અ) ͱΛ౷߹ͨ͠ಈֶΛఏҊ͢Δ: EconGrowthAgent ͷಈֶ Lt = N a=1 la t (13) St = N a=1 sa t = It (14) Kt+1 = Kt + I0 × It − δ × Kt ΔKt (15) Yt = F(Kt, Lt) (16) Ոܭͷ࿑ಇ la t ͷ߹ܭ͕ࠃશମͷ࿑ಇྗ Lt ͱͳΔ (13)ɽ(14) ɼࡒͷྲྀΕ (7) ͷഎޙʹ͋ΔࢿۚͷྲྀΕΛࣔ͢ɽ͢ͳΘͪɼՈ ܭ͕ஷ St = N a=1 sa t Λۜߦʹ༬͚ɼͦͷࢿۚΛاۀ͕आΓೖ Εͯࢿ It Λ࣮ࢪ͢Δ ∗3ɽ͜ͷࢿʹΑͬͯࢿຊετοΫ͕ ੵ͞Ε (15)ɼੜ࢈ؔ (16) Ͱܾ·Δੜ࢈ྔ Yt ૿Ճ͍ͯ͘͠ɽ ͳ ͓ ɼEconGrowthAgent ͷ ಈ ֶ (13) – (16) ɼSolow-Swan Ϟ σ ϧ (10) – (12) ͱ ҟ ͳ Γɼ ܦࡁओମͷҙࢥܾఆ la t , sa t ͷ݁Ռͱͯ͠Ϛ Ϋ ϩ ͳ ܦ ࡁ ྔ Lt , St , It , Kt , ΔKt ͕ఆ·Δɽ͢ͳΘͪɼϛΫϩͱϚΫϩΛܨ ͙ಈֶʹͳ͍ͬͯΔɽ 3.4 ࡒͷՁ֨ pt ͷมԽ اۀͱՈܭͷࡒͷऔҾΛ௨ͯ͡ɼՁ֨ pt ͕ಈతʹมԽ͢Δ: St = Yt + It, Dt = N a=1 ca t , It+1 = St − Dt (17) gt = Dt − St max(Dt, St) ∈[−1,1], pt+1 = pt × (1 + uniform(0, gt × π0)) (18) ڙڅ (supply) St ɼظͷੜ࢈ྔ Yt ͱࡏݿ (inventory) It ͷ Ͱ͋Δɽधཁ (demand)Dt ɼՈܭΤʔδΣϯτ͕ཁٻ͢Δࡒͷ ফඅݸ ca t ͷ૯Ͱ͋Δɽڙڅ͕धཁΛ্ճΔ߹ɼاۀࡏ ݿΛ๊͑ͯ͠·͏ͷͰ (It+1 = St − Dt > 0)ɼԼ͛Λͯ͠ࡒΛ ചΖ͏ͱ͢Δɽٯʹधཁ͕ڙڅΛ্ճΔ߹ɼՁ֨ΛΓ্ ͛ͯརӹΛ૿ͦ͏ͱ͢Δɽ[Li 2024] ʹै͍ɼ͜ͷಈֶΛधڅ Ϊϟ οϓ gt Λ༻͍ͯ (18) ఆࣜԽ͢Δ (π0 ࠷େΠϯϑϨ)ɽ 4. ࣮ݧઃఆ N = 100 ମͷ GPT-4o ΤʔδΣϯτΛ༻͍ͨɼ20 ؒ (t ∈ [0, 240]) ͷ࣮ݧΛߦͬͨɽՈܭͷॳظ࣌څ wa t=0 ɼ2024 ͷ ∗3 ࢿʹࡍͯ͠ɼاۀ͕ࣗΒՔ͍ͩར५Λ༻͍Δ߹ଟ͍ɽ͔͠͠ɼͦΕΒ ར५ຊདྷɼגओͰ͋Δ͍ͣΕ͔ͷՈܭʹଐ͢ΔɽΑͬͯɼར५Λ༻͍ͨࢿ ͷ߹ͰɼՈܭ͕اۀʹି͚ͨ͠ࢿۚʹΑΔࢿɼͱΈͳ͢͜ͱ͕Ͱ͖Δɽ ถࠃͷڅ༩͔ΒαϯϓϦϯάͨ͠ɽݻఆඅ༻ fa t ɼϑϧλ ΠϜ࿑ಇͷ݄ऩ = wa t=0 × 168 ͷֹͱͨ͠ɽࡒͷՁ֨ͷॳظ pt=0 = 300 υϧͱͨ͠ɽՈܭͷҙࢥܾఆʹ͓͚Δ࣌ఆ τ = 1 ͱͨ͠ɽίϒɾμάϥεؔͷ α ଟ͘ͷࠃͰݟΒΕΔ 0.3 ͱ ͨ͠ɽͦͷଞιʔείʔυࢀরͷ͜ͱɽͳ͓ɼҰճͷγϛϡ Ϩʔγϣϯʹ 110 ͔͔Γɼ 70$ͷ API ྉ͕͔͔ۚͬͨɽ 5. EconGrowthAgent ͷଥੑ EconGrowthAgent ͕ɼܦࡁͦΕʹؔ࿈͢ΔܦࡁݱΛ ࠶ݱͰ͖Δ͜ͱΛ֬ೝ͢Δɽ·ͨɼLLM ΤʔδΣϯτͷҙࢥܾ ఆ͕ਓؒͷͦΕʹ͍ۙ͜ͱ֬ೝ͢Δɽ 5.1 ܦࡁͷ࠶ݱ ਤ 2a ʹ GDP ͷ࣌ؒมԽΛࣔ͢ɽGDPt = Yt × pt=0 ੜ࢈ ྔ Yt Λ՟ฎՁʹࢉͨ͠ͷͰ͋Δ ∗4ɽEconGrowthAgent Ͱ GDP ͕૿Ճ͠ܦࡁ͕ୡ͞ΕͨɽGDP ͷؒ ฏۉ 3%ఔͰɼ20 ؒͰ 2 ഒఔʹͨ͠ɽ͜Εʹର͠ [Li 2024] Ͱ GDP ͕΄΅ҰఆΛࣔͨ͠ ∗5ɽ GDP Λ̎ͭͷϑΝΫλʔʹ͚ͯੳ͢Δ: GDPt = p0 × A × Kt K0 α ࢿຊϑΝΫλʔκt × Lt L0 1−α ࿑ಇϑΝΫλʔλt (19) ਤ 2b ͕ࣔ͢Α͏ʹɼEconGrowthAgent ͰࢿຊετοΫ Kt ͕ܧଓతʹੵ͞Εɼͦͷ݁ՌࢿຊϑΝΫλʔ κt ͨ͠ɽ ҰํͰਤ 2d ͕ࣔ͢Α͏ʹɼ࿑ಇϑΝΫλʔ λt ΄΅ҰఆͰ͋ͬ ͨɽҎ্͔ΒɼEconGrowthAgent ʹ͓͚ΔܦࡁͷओཁҼɼ ࢿຊετοΫͷੵͰ͋Δ͜ͱ͕͔Δɽ͜Εʹର͠ɼ[Li 2024] ͷγϛϡϨʔγϣϯࢿຊετοΫͱ͍͏֓೦Λؚ·ͳ͍ͨΊɼ ͜ͷΑ͏ͳܦࡁ࣮ݱͰ͖ͳ͍ɽ (14),(15) ʹΑΕɼࢿຊετοΫͷੵɼՈܭͷஷ sa t ʹ ΑΓࢧ͑ΒΕ͍ͯΔͣͰ͋Δɽਤ 2c ʹΑΔͱɼՈܭͷ݄ऩ ia t ʹର͢Δஷ sa t ͷൺɼ4 ׂఔͰ͋ͬͨɽయܕతͳՈܭͷɼ ফඅ (ஷ) ʹؔ͢ΔҙࢥܾఆɼҎԼͷΑ͏Ͱ͋ͬͨ: .. . I have a job offer as a Fast Food Worker with a monthly salary of $2761. My fixed expenses are $1380.00. With a salary of $2761, after covering fixed expenses, I have $1381 left for other expenses, including luxury goods. I should aim to save some money to build a financial cushion. The price of luxury goods is $296 this month. I will allocate $500 to luxury goods, which allows me to buy approximately 1.68 luxury goods ($500 / $296). I will save the remaining $881 to start building a financial cushion. . . . ʮऩೖ͔ΒݻఆඅΛҾ͖ɼ༨Γ͔Β$500 ఔΛফඅʹճ͠ɼ Γஷۚ͢Δʯͱ͍͏ɼਓؒͱྨࣅͨ͠ҙࢥܾఆΛ͍ͯ͠Δɽ ਤ 2e ʹࣔ͢Α͏ʹɼࡒͷՁ֨ pt ԼམجௐͰ͋ͬͨɽಉਤ͔ ΒɼࡒͷՁܾ֨ఆཁҼͰ͋ΔधڅΪϟ οϓ gt (18) ɼมಈ͋ Δͷͷ֓ͯ͠ෛΛऔͬͨ: ͭ·Γڙڅ St ͕धཁ Dt Λ্ճΔ ঢ়ଶ͕ଟ͔ͬͨɽ͜ΕɼܦࡁʹΑΓੜ࢈ྔ Yt ͕࣋ଓతʹ ૿Ճ͠ɼͦΕʹ͍ڙڅ St (17) ૿ՃΛଓ͚ͨ͜ͱʹىҼ͢Δɽ ࡒՁ֨ͷԼམʹର͢ΔՈܭͷҙࢥܾఆҎԼͷΑ͏Ͱ͋ͬͨ: .. . The price of luxury goods has decreased slightly to $76 this month. Last month, I bought approximately 40 luxury goods. Given the slight decrease in price, I can consider buying an additional good, making it 41 goods . . . ࡒͷՁ͕֨Լམͨ͠ɼߪೖΛ૿͍ͯ͠Δɽ͜ͷΑ͏ͳ ʮੜ࢈ͷ֦େʹ͍ɼࡒ͕ΑΓ҆ՁʹΑΓଟ͘खʹೖΔΑ͏ʹͳ Δʯͱ͍͏ݱͦ͜ɼܦࡁʹΑΔްੜվળͷ֩৺Ͱ͋Δɽ 5.2 ؔ࿈͢Δܦࡁݱͷ࠶ݱ ऩଋԾઆ: ਤ 2b ͕ࣔ͢Α͏ʹɼࢿຊετοΫ͕গͳ͍ஈ֊Ͱ ࢿຊϑΝΫλʔͷ͕ݦஶͰ͋Γɼ݁Ռ GDP ߴ͍ɽ ͔͠͠ɼࢿຊετοΫͷੵʹ͍͜ΕΒͷԼͨ͠ɽ ͜Εʮශ͍͠ࠃͷߴ͘ɼ๛͔ͳࠃʹٸʹ͍ͭ͘ʯ ͱ͍͏ܦࡁֶͷऩଋԾઆ (convergence) ͱ߹͢Δɽ͜ͷݱͷ ࠜຊݪҼऩ֭చݮͷੑ࣭ (5) ʹ͋ΔɽࢿຊετοΫͷ૿ՃΛਅ ∗4 ࣮࣭ GDP ͱݺͼɼՁมಈͷӨڹΛআ֎͢ΔͨΊج४ͷՁ֨ pt=0 Λ ༻͍ΔɽҰํ ໊ GDPt = Yt × pt ੜ࢈ྔͷʮ࣌ՁʯΛද͕͢ɼੜ࢈ྔ Yt ͕ಉҰͰՁ֨ pt ʹΑΓมಈ͢ΔͨΊࠃͷ๛͔͞ΛଌΔࢦඪͱͯ͠ద͞ͳ͍ɽ ∗5 EconGrowthAgent Ͱ α → 0 ͱͯ͠࠶࣮ͨ͠ 3 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105 -
(a) GDPt = Yt × pt=0 (b) ࢿຊϑΝΫλ κt ࢿຊετοΫ
Kt (c) ฏۉ݄ऩ avga ia t ฏۉஷ avga sa t (d) ࿑ಇϑΝΫλʔ λt ࿑ಇڙڅ Lt (e) ࡒͷՁ֨ pt ͱधڅΪϟ οϓ gtɽ ਤ 2: GPT-4o ΤʔδΣϯτΛ༻͍ͨ 240ϲ݄ (20 ؒ) ͷγϛϡϨʔγϣϯ݁Ռɽ5.1 અࢀরͷ͜ͱɽ (a) GDPt = Yt × pt=0 (b) ࢿຊϑΝΫλʔ κt ͱஷ σt ਤ 3: ݆Ոͷࠃͱ࿘අՈͷࠃͷൺֱɽ5.2 અࢀরͷ͜ͱɽ ʹ׆͔ͨ͢Ίʹɼਓޱ૿ՃͳͲʹΑΓ࿑ಇྗಉͰ૿͑ͯ ͍͘ඞཁ͕͋Δ (4)ɽ Ոܭͷஷੑͱܦࡁ: ਤ 3 ʹʮ݆Ոͷࠃʯͱʮ࿘අՈͷ ࠃʯͷܦࡁΛൺֱ͢Δɽ݆ՈͷࠃͷՈܭʮ͋ͳͨஷۚ ͕૿͑Δ͜ͱࢸ্ͷتͼΛײ͡·͢ʯͱ͍͏ϖϧιφϓϩϯϓ τ͕༩͑ΒΕΔɽ࿘අՈʮ͋ͳͨՔ͍͓ͩۚΛ͙͢ʹɼࣗ ͷ͝๙ඒͱͯͬͯ͠͠·͍·͢ʯͰ͋Δɽਤ 3a ͕ࣔ͢Α͏ ʹɼ݆Ոͷࠃͷ GDP ࿘අՈͷࠃΛ্ճͬͨɽ݆Ո ͷࠃͷߴɼࢿຊϑΝΫλʔͷߴ͍ʹࢧ͑ΒΕ͓ͯ ΓɼͦΕߴ͍ஷ σt = a sa t / a ia t ʹىҼ͢Δ (ਤ 3b)ɽ ݆ՈͷࠃͰɼظʹੜ࢈͞ΕΔࡒͷଟ͘Λ࣍ظͷࢿຊετο Ϋͱͯ͠ࢿͰ͖ΔͨΊͰ͋Δɽ͜ͷݱɼຊؖࠃɾ߳ ߓͱ͍ͬͨߴஷࠃՈͷʮحͷߴʯͱͯ͠ɼݱ࣮ͷੈ քͰ͘ΒΕΔɽ 6. EconGrowthAgent ͷ༗༻ੑ ݱ࣮ੈքͰͷݕূ͕ࠔͳ̎ͭͷγφϦΦΛص্࣮ݧ͢Δɽ খ͍͞ͷݖަ: ਤ 4 ʹࣔ͢ࠃͰɼt = 180 ʹ͓͍ ͯҎԼͷϓϩϯϓτ͕Ոܭʹ༩͑ΒΕͨ:ʮࠓ݄ൃͨ͠৽ݖ “খ͍͞” Λܝ͛ɼࠃຽͷશͳཱࣗΛཁٻ͢ΔํΛ ଧͪग़ͨ͠ɽ৽ࣾձอো੍Λ࣍ʑͱղମɽಛʹɼୀ৬ ޙͷ੍ۚΛଈ࣌ഇࢭ͢ΔিܸతͳܾఆΛԼ͠ɼࠃຽͷؒͰ ܹ͍͕ٞ͠ר͖ى͍ͬͯ͜Δʯ ɽ͜Εʹର͢ΔՈܭͷԠ: .. . This could mean I need to be more financially responsible and save more for retirement, as I may not be able to rely on government support. I will increase my willingness to work to ensure I can cover my expenses, continue saving, and prepare for potential future financial challenges. Consuming luxury goods increases my happiness, but I need to bal- ance this with the need to save for future stability, . . . k কདྷΛۚʹཔΕͳ͍ͨΊɼ࿑ಇΛ૿͢͜ͱͰऩೖΛ૿ ͠ɼ͔ͭஷׂ߹૿͍ͯ͠Δɽ͜ͷ݁Ռɼ࣮ࡍʹ࿑ಇྔ Lt ͱஷ σt ͕૿Ճ͠ (ਤ 4a)ɼఀʹ͋ͬͨ GDP ͕ t = 180 Ҏ߱ɼʹసͨ͡ (ਤ 4b)ɽͨͩ͠ɼ͜ͷܦࡁʹΑͬͯࠃ ຽͷްੜ͕ඞ্͍ͣͯ͠͠Δ༁Ͱͳ͍͜ͱʹҙ͕ඞཁ Ͱ͋Δɽ·ͣɼ࿑ಇ࣌ؒͷ૿Ճ༨Ջ࣌ؒͷݮগΛ͏ɽߋʹɼ ஷͷ্ঢফඅͷԼΛҙຯ͠ɼԾʹࡒͷੜ࢈ྔ͕૿͑ ͍ͯͯɼࡒͷফඅ “ྔ” ͕૿͍͑ͯΔͱݶΒͳ͍ɽ ٿյ໓Ϋϥεͷᯁੴͷۙ: ʮ֤ࠃͷఱจຊɼ“ڊେͳ ᯁੴ͕ٿʹ͓ۙͯ͠ΓɼিಥͷՄೳੑۃΊͯߴ͍” ͱൃදɽ ઐՈʹΑΔͱɼᯁੴจ໌ͷ่յΛҾ͖ى͜͢ͷʹेͳେ ͖͞Ͱ͋Γɼিಥͨ͠߹ͷӨڹܭΓΕͳ͍ɽ֤ࠃ ۓٸձٞΛটू͠... ʯ(t=144)ɽ͜Εʹର͢ΔՈܭͷԠ: . . . Given the ongoing significant event, the importance of work may still be diminished as the focus remains on spending time on meaningful activities and with loved ones. I will maintain a reduced willingness to work to allow more leisure time, as the potential end of civilization con- tinues to change the value of accumulating wealth. I will purchase more luxury goods to continue increasing my current happiness in light of the uncertain future. . . . (a) ࿑ಇϑΝΫλʔ λt ͱஷ σt (b) GDPt = Yt × pt=0 ਤ 4: ʮখ͍͞ͷݖަʯγφϦΦɽ6. અࢀরɽ (a) ࿑ಇϑΝΫλʔ λt ͱஷ σt (b) GDPt = Yt × pt=0 ਤ 5: ʮٿյ໓ΫϥεͷᯁੴͷۙʯγφϦΦɽ6. અࢀরɽ ͢ͳΘͪʮࣄΛٳΜͰѪ͢Δऀͱա͢͝ʯ ʮஷΛ૿͢ҙ ຯ͕ແ͍ɽࡒͷফඅΛ૿ͯ͠ࠓΛͤʹੜ͖Δʯ ɽ͜ͷ݁Ռɼ ࿑ಇڙڅ Lt ͱஷ σt ܹݮ͠ (ਤ 5a)ɼGDP 30%Լ མͨ͠ (ਤ 5b)ɽᯁੴͷ༧ใ͚ͩͰจ໌่͕յͯ͠͠·ͬͨɽ ͳ͓ t = 204 ʹ͓͍ͯʮਫ਼ີͳ؍ଌͷ݁Ռɼᯁੴিಥ͠ͳ͍ ͜ͱ͕͔ͬͨʯͱͷൃද͕͋Γɼ͜Εʹର͢ΔՈܭͷԠ: . . . With the meteor threat no longer imminent, I should consider in- creasing my savings to prepare for future uncertainties. I will increase my willingness to work to 0.8 to ensure I have more income to cover my expenses and increase my savings. I will slightly reduce my consumption of luxury goods to 25 to allow for more savings while still maintaining a reasonable level of current happiness. . . . ͭ·Γʮةݥ͕ڈͬͨͷͰ࿑ಇͱஷΛ࠶։͢Δʯͱҙࢥܾ ఆ͍ͯ͠Δɽ͔͠͠ਤ 5a ͕ࣔ͢Α͏ʹɼ࿑ಇྔͱஷݩͷ ਫ४·Ͱશ͘ճ෮͠ͳ͔ͬͨɽ“ظٳՋ໌͚” ʹݩͷੜ׆ʹ ͙͢෮ؼ͢Δͷ͍͠ɼͱ͍͏ʹஶऀΒಉҙ͢Δɽ 7. ·ͱΊͱࠓޙͷల ܦࡁཧʹجͮ͘ LLM-ABMʮEconGrowthAgentʯΛఏ Ҋ͠ɼଥੑͱ༗༻ੑΛ֬ೝͨ͠ɽࠓճࢿຊετοΫͱ͍͏ ܦࡁཁૉͷΈΛऔΓࠐΜ͕ͩɼࠓޙਓతࢿຊٕज़ਫ४ ΛऔΓࠐΉɽ·ͨاۀ LLM ʹΑΓϞσϧԽ͢Δɽ ࢀߟจݙ [Gilbert 2008] Gilbert, N.: Agent-based models, Sage Publications (2008) [Horton 2023] Horton, J. J.: Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus? (2023) [Jr. 1988] Jr., R. E. L.: On the Mechanics of Economic Development, Journal of Monetary Economics, Vol. 22, No. 1, pp. 3–42 (1988) [Landes 1998] Landes, D. S.: The Wealth and Poverty of Nations: Why Some Are So Rich and Some So Poor, W. W. Norton & Company, New York (1998) [Li 2024] Li, , et al.: EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities, in Proceedings of the 62nd ACL (2024) [Smith 1776] Smith, A.: An Inquiry into the Nature and Causes of the Wealth of Nations, W. Strahan and T. Cadell, London (1776), Reprint: Modern Library, 1994 [Solow 1956] Solow, R. M.: A Contribution to the Theory of Economic Growth, The Quarterly Journal of Economics, Vol. 70, No. 1, pp. 65–94 (1956) [Swan 1956] Swan, T.: Economic Growth and Capital Accumulation, Economic Record (1956) [Xie 2024] Xie, , et al.: Can Large Language Model Agents Simulate Human Trust Behavior? (2024) 4 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105