JSAI2025 EconGrowthAgent 予稿

Transcript

1. EconGrowthAgent: LLMΤʔδΣϯτͱܦࡁ੒௕ཧ࿦ʹجͮ͘ 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 ܦࡁֶ͸ɼਓྨࣾձΛΑΓ๛͔ʹ͢Δ͜ͱΛ໨తʹɼܦࡁݱ ৅ͷղ໌Λԡ͠ਐΊ͖ͯͨɽ19 ੈلޙ൒ʹ࢝·ΔϛΫϩܦࡁֶ Ͱ͸ɼཧ࿦తͳΞϓϩʔνʹΑΓɼফඅऀ΍اۀͱ͍ͬͨܦࡁओ ମͷҙࢥܾఆΛ෼ੳͨ͠ɽফඅऀͷҙࢥܾఆͱͯ͠͸ʮຖ݄ͷ ऩೖΛੜ׆අɾޘָඅɾஷۚͳͲʹ഑෼͢Δʯ ɼاۀͷҙࢥܾఆ ͱͯ͠͸ʮ঎඼ (ࡒ) ͷੜ࢈ྔ΍Ձ֨Λܾఆ͢Δʯͱ͍ͬͨྫ͕ ڍ͛ΒΕΔɽ͜͏͍ͬͨҙࢥܾఆΛ෼ੳ͢ΔͨΊʮ֤ܦࡁओମ ͷརӹ (ްੜ) ΛදΘؔ͢਺͕࠷େԽ͞ΕΔΑ͏ͳબ୒͕ͳ͞Ε Δʯͱ͍͏Α͏ͳཧ࿦ͱͯ͠ϞσϧԽ͞Εͨɽ ཧ࿦తͳΞϓϩʔν͚ͩͰ͸ɼݱ࣮ੈքͱͷ੔߹ੑΛݕ ূͰ͖ͳ͍ɽ20 ੈلલ൒ʹൃలͨ͠ϚΫϩܦࡁֶͰ͸ɼ ؍ଌσʔλ͔ΒͷΞϓϩʔνΛ༻͍ͯܦࡁݱ৅Λ෼ੳͨ͠ɽ͢ͳ Θͪɼݸʑͷܦࡁओମͷҙࢥܾఆʹ͸ཱͪೖΔ͜ͱͳ͘ɼGDP ΍ ࣦۀ཰ͱ͍ͬͨूܭతͳ؍ଌσʔλͷؔ܎ੑΛ௚઀ϞσϧԽͨ͠ɽ GDP ੒௕཰ͱࣦۀ཰ͷٯ૬ؔΛࣔ͢ʮΦʔΫϯͷ๏ଇʯ ɼΠϯϑϨ ཰ͱࣦۀ཰ͷؔ܎Λࣔ͢ ʮϑΟϦοϓεۂઢʯ ͳͲ͕୅දతͰ͋Δɽ ୯ʹ؍ଌʹΑΓಘΒΕΔσʔλΛ෼ੳ͢Δ͚ͩͰͳ͘ɼࣗΒ࣮ݧ ΛߦͬͯσʔλΛऔಘ͠ɼ ෼ੳ͢Δ͜ͱ΋Ͱ͖Δɽ 20 ੈلޙ൒ʹൃ లͨ͠ήʔϜཧ࿦΍ߦಈܦࡁֶͰ͸࣮ݧ࣮ࣨݧʹΑΔΞϓϩʔν ͕Ֆ։͍ͨɽྫ͑͹ਓؒͷઓུతࢥߟೳྗΛ෼ੳ͢ΔͨΊɼڭ ࣨʹूΊͨੜెʹʮनਓͷδϨϯϚʯͷΑ͏ͳγφϦΦΛϓϨ Πͤ͞ɼಘΒΕͨߦಈσʔλΛ෼ੳ͢Δɽ ࣮ݧ࣮ࣨݧʹ͸ܽ఺΋ଘࡏͨ͠ɽ·ͣɼ࣮ݧʹ͸ਓతίετ ͕͔͔ΔͷͰɼ࠶ݱ࣮ݧ΍৚݅Λม࣮͑ͨݧ౳Λେྔʹߦ͏͜ ͱ͕Ͱ͖ͳ͍ɽ·ͨಉ༷ͷཧ༝͔Βɼܦࡁओମͷ਺͕૿͑ͨ৔ ߹ͷ࣮ݧ΋೉͍͠ɽߋʹɼڭࣨͰͷ࣮ݧͰ͸ຊ࣭తʹݕূͰ͖ ͳ͍γφϦΦ΋ଘࡏ͢Δɽྫ͑͹ɼ ʮܦࡁ੓ࡦΛ X ʹมߋͨ͠৔ ߹ͷ GDP ੒௕཰΁ͷӨڹʯ͸ɼͦͷن໛ͱෳࡶੑ͔Βݱ࣮తͳ ݕূ͕ࠔ೉Ͱ͋Γɼ ʮ৽ܕײછ঱ Y ͷຮԆʹΑΔࣾձ΁ͷӨڹʯ ͸ɼྙཧత؍఺͔Β࣮ݧࣗମ͕ڐ༰͞Εͳ͍ɽ ܭࢉػՊֶͷ෼໺Ͱൃలͨ͠ΤʔδΣϯτϕʔευϞσϦϯ ά (Agent-Based Models, ABM) [Gilbert 2008] ͸ɼ͜ΕΒͷܽ఺ Λࠀ෰͢ΔՄೳੑΛ༗͢ΔɽABM Ͱ͸ɼݸʑͷҙࢥܾఆओମΛ ໛ͨ͠ϓϩάϥϜ (=ΤʔδΣϯτ) Λܭࢉػ্ʹ༻ҙ͠ɼԾ૝త ͳ࣮ݧΛߦ͏ɽਓతͳ੍໿Λड͚ͳ͍ͨΊɼେྔͷ࣮ݧ΍ओମ਺ Λ૿΍࣮ͨ͠ݧ͕ՄೳͱͳΓɼ·ͨʮ΋͠ Xʯ ʮ΋͠ Yʯͱ͍ͬ ͨγφϦΦ΋ࣗ༝ʹݕূͰ͖ΔɽҰํͰɼABM Λܦࡁֶʹద༻ ͢Δʹ͸՝୊΋͋ΔɽABM ͷΤʔδΣϯτ͸୯७ͳϓϩάϥϜ (ϧʔϧ) ʹج͖ͮҙࢥܾఆΛߦ͏ҰํͰɼܦࡁֶʹ͓͚Δओମ Ͱ͋Δਓؒͷҙࢥܾఆ͸ෳࡶ͔ͭଟ໘తͰ͋Δɽͭ·Γɼϧʔ ϧϕʔεΤʔδΣϯτͰ͸ܦࡁओମΛۙࣅ͢Δ͜ͱ͸Ͱ͖ͳ͍ɽ ࿈བྷઌ: ৿Լᚠจɼterufumi.moriwhita.wp@hitachi.com ۙ೥ొ৔ͨ͠େن໛ݴޠϞσϧ (LLM) ͸͜ͷ՝୊Λղܾ͢Δ ՄೳੑΛൿΊ͍ͯΔɽLLM ͸ɼਓ͕ؒॻ͍ͨେن໛σʔλ͔Β ਓؒͷҙࢥܾఆύλʔϯΛֶश͍ͯ͠ΔͨΊ [Horton 2023, Xie 2024]ɼܦࡁओମͨΔਓؒΛߴਫ਼౓ʹۙࣅͨ͠ΤʔδΣϯτͱ͠ ͯͷ໾ׂ͕ظ଴Ͱ͖ΔɽΑͬͯɼLLM ΤʔδΣϯτΛ༻͍ͨ ABM (LLM-ABM) ͸ɼܦࡁݱ৅Λ෼ੳ͢Δ৽ͨͳΞϓϩʔνͱ ͯ͠ɼܦࡁֶ΁มֵΛ΋ͨΒ͢Մೳੑ͕͋Δɽ ͜ͷํ޲ʹ࠷ॳͷҰาΛ౿Έग़ͨ͠ͷ͸ɼ[Li 2024] ΒͰ͋Δɽ Li Β͸ LLM ΤʔδΣϯτʹΑͬͯՈܭΛϞσϧԽ͠ɼ࿑ಇ΍ ফඅʹؔ͢ΔҙࢤܾఆΛͤͨ͞ɽ݁ՌɼLLM ΤʔδΣϯτ͸ ʮCOVID ͕ྲྀߦ͍ͬͯΔͷͰ৬৔ʹߦ͖ͨ͘ͳ͍ʯ ʮऩೖ͕ݮͬ ͨͷͰࢧग़ʹ͸ؾΛ෇͚Α͏ʯͱ͍ͬͨਓؒʹ͍ۙҙࢤܾఆΛ ߦ͑Δ͜ͱ͕֬ೝ͞Εͨɽ·ͨɼଟ਺ͷ LLM ΤʔδΣϯτΛ༻ ͍࣮ͨݧʹΑΓɼϑΟϦοϓεۂઢ΍ΦʔΫϯͷ๏ଇͱ͍ͬͨϚ Ϋϩܦࡁֶͷ୅දతͳ๏ଇ͕࠶ݱͰ͖Δ͜ͱΛ֬ೝͨ͠ɽ ͔͠͠ͳ͕Β [Li 2024] Βͷ݁Ռ͸ɼܦࡁֶʹ͓͍ͯ࠷΋ॏཁ ͳݱ৅Ͱ͋Δʮܦࡁ੒௕ʯΛ࠶ݱ͢Δ͜ͱ͕Ͱ͖͍ͯͳ͔ͬͨɽ ۩ମతʹ͸ɼҰࠃͷ๛͔͞ͷࢦඪͰ͋Δ GDP ( ࡒͷੜ࢈ྔ) ͕ ҰఆʹͱͲ·͍ͬͯͨɽܦࡁ੒௕͸ɼ͋ΒΏΔܦࡁओମͷްੜ Λେ෯ʹҾ্͖͛ΔͨΊɼܦࡁݱ৅ͷதͰ΋֨ผʹॏཁͩͱΈ ͳ͞Ε͖ͯͨ [Smith 1776,Jr. 1988,Landes 1998]ɽྫ͑͹ɼถࠃ ܦࡁ͸աڈҰੈلͷؒʹ೥཰ฏۉ 2%ఔ౓Ͱ੒௕͠ɼෳརͷྗΛ आΓͯɼ͓Αͦ 10 ഒͷن໛ʹ·Ͱ੒௕ͨ͠ɽ͜Ε͸͢ͳΘͪɼ 100 ೥લʹൺ΂ͯɼզʑ͕ফඅͰ͖Δࡒͷྔ͕ 10 ഒʹͳͬͨ͜ ͱΛҙຯ͓ͯ͠Γɼզʑͷްੜ͸େ෯ʹ޲্͍ͯ͠Δ ∗1ɽ ͦ͜ͰຊݚڀͰ͸ɼܦࡁ੒௕ΛදݱͰ͖Δ LLM-ABM ͷߏங Λ໨ࢦ͢ɽ·ͣɼઌߦݚڀʹ͓͍ͯܦࡁ੒௕Λଊ͑Δ͜ͱ͕Ͱ͖ ͳ͔ͬͨཧ༝Λ୳ΔͨΊɼϚΫϩܦࡁֶͷҰ෼໺Ͱ͋Δʮܦࡁ੒ ௕ཧ࿦ʯΛ๚ΕΔɽܦࡁ੒௕ཧ࿦ʹΑΔͱɼܦࡁ੒௕ʹͱͬͯॏ ཁͱͳΔཁૉ͸ɼੜ࢈ʹ༻͍Δಓ۩Ͱ͋ΔʮࢿຊετοΫྔʯ ɼ ࿑ಇऀͷख़࿅౓Ͱ͋Δʮਓతࢿຊʯ ɼࠃͷʮٕज़ਫ४ʯͳͲͰ͋ Δɽ͜ΕΒཁૉ͕اۀ΍ࠃͷ౤ࢿʹΑΓ੒௕͠ɼͦΕʹ൐͍ࡒ ͷੜ࢈ྔ͕૿Ճ͢Δ͜ͱͰɼܦࡁ੒௕͕࣮ݱ͢Δɽ[Li 2024] ͷ γϛϡϨʔγϣϯ͸ɼ͜ͷΑ͏ͳܦࡁ੒௕ཁૉͷಈֶΛؚΜͰ ͍ͳ͔ͬͨͨΊɼܦࡁ੒௕ΛදݱͰ͖ͳ͔ͬͨͷͰ͋Δɽ ࣍ʹɼ͜ͷߟ࡯Λ౿·͑ɼܦࡁ੒௕ཁૉͷಈֶΛऔΓࠐΜͩ LLM-ABM Ͱ͋ΔʮEconGrowthAgentʯ(ਤ 1) Λߏங͢Δɽ۩ ମతʹ͸ɼܦࡁओମΛ LLM ΤʔδΣϯτʹΑͬͯϞσϧԽ͠ɼ ͦͷҙࢥܾఆ͔Βܦࡁ੒௕΁ͷಈֶΛɼҎԼͷΑ͏ʹ࣮ݱ͢Δ: (1) ՈܭΤʔδΣϯτ͸ຖظɼ࿑ಇྗΛఏڙ͠ɼऩೖͷҰ෦Λஷ ஝ͱͯۜ͠ߦʹ༬͚Δɽ(2) اۀΤʔδΣϯτ͸࿑ಇྗͱࢿຊε τοΫΛ༻͍ͯࡒΛੜ࢈͢Δɽߋʹɼۜߦ͔ΒࢿۚΛିΓ͏͚ɼ ௥ՃͷࢿຊετοΫΛߪೖ͢Δɽ(3) ࢿຊετοΫͷ૿ՃʹΑΓ ∗1 ٯʹɼզʑͷ਎ͷճΓͷࡒɼྫ͑͹৯ྐɾҥ෰ɾޘָɾҩྍͳͲ͕े෼ͷҰ ʹͳͬͨੈքΛߟ͑ͯΈΑ͏ɽ͓ͦΒ͘ɼ͋·ΓʹෆศͰת͑ΒΕͳ͍ͩΖ͏ɽ 1 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105
2. 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 ฅ໮ࡿ࢚ ਺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࢈ग़ྔ͕૿Ճ͢Δɽͳ͓ࠓճ͸ɼܦࡁ ੒௕ཁૉͷதͰ࠷΋ॏཁͳࢿຊετοΫͷΈΛ࣮૷ͨ͠ɽ ࠷ޙʹɼ100 ମͷ GPT-4o ΤʔδΣϯτΛ༻͍ͨ 20 ೥ؒͷγ ϛϡϨʔγϣϯΛߦ͍ɼܦࡁ੒௕ͱͦΕʹؔ࿈͢Δܦࡁݱ৅Λ࠶ ݱͰ͖Δ͜ͱΛ֬ೝͨ͠ɽ·ͨɼLLM ΤʔδΣϯτͷҙࢥܾఆ͕ ਓؒͷͦΕʹ͍ۙ͜ͱ΋֬ೝͨ͠ɽ͜ΕΒ͸ɼEconGrowthAgent ͷଥ౰ੑΛࣔ͢ɽ࣍ʹʮ௒খ͍͞੓෎΁ͷ੓ݖަ୅ʯ ʮ஍ٿյ໓ Ϋϥεͷᯁੴͷ઀ۙʯͱ͍͏γφϦΦΛص্ݕূ͢Δɽ͜ͷΑ ͏ͳݱ࣮Ͱͷݕূ͕ࠔ೉ͳγφϦΦΛؚΊɼଟ༷ͳγφϦΦΛ ࣗ༝͔ͭେྔʹ࣮ݧͰ͖Δ͜ͱ͸ɼEconGrowthAgent ͷ༗༻ੑ Λ͍ࣔͯ͠Δɽͳ͓ιʔείʔυΛެ։༧ఆͰ͋Δ ∗2ɽ 2. Կ͕ܦࡁ੒௕ΛܾΊΔͷ͔ʁ ϚΫϩܦࡁֶͷʮܦࡁ੒௕ཧ࿦ʯʹج͖ͮɼܝ୊Λٞ࿦͢Δɽ 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 ͱ͸ɼੜ࢈ʹ࢖༻͢Δಓ۩Ͱ͋Γɼྫ͑͹ࣄ຿һ͕࢖͏ PCɼ ৯඼ϝʔΧʔ͕࢖͏৯඼Ճ޻ϥΠϯɼݐઃձ͕ࣾ࢖͏ΫϨʔϯɼ ͳͲͰ͋ΔɽܦࡁֶͰ͸ɼ͜ΕΒͷؔ܎Λੜ࢈ؔ਺ 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 ͸ద౰ͳఆ਺Ͱ͋Δɽίϒɾμάϥεؔ਺͸ɼزͭ ͔ͷॏཁͳੑ࣭Λຬͨ͢ɽ ∂F ∂K > 0, ∂F ∂L > 0 (୯ௐ૿Ճੑ) (3) ͢ͳΘͪɼK ΍ L Λ૿΍ͤ͹ੜ࢈ྔ Y ΋૿͑Δɽ࣍ʹ F(zK, zL) = zF(K, L) (ن໛ʹؔͯ͠ऩ֭Ұఆ) (4) ͢ͳΘͪɼK ͱ L Λಉ࣌ʹ z ഒʹ૿΍ͨ͠৔߹ɼY ΋ z ഒʹͳ Δɽྫ͑͹ɼ৯඼Ճ޻ϥΠϯΛ 2 ഒʹ૿΍͠ɼ͔ͭ࡞ۀऀ਺΋ 2 ഒʹ૿΍ͨ͠৔߹ɼ৯඼ͷੜ࢈ྔ͸ 2 ഒʹͳΔɽ࠷ޙʹ ∂2F ∂2K < 0, ∂2F ∂2L < 0 (ऩ֭చݮ) (5) ޙऀΛߟ͑ΔͱʮࢿຊετοΫ K Λݻఆͨ͠··࿑ಇྗ L Λ૿ ΍͢৔߹ɼL ͕େ͖͘ͳΔ΄Ͳɼ௥Ճͷ L ͕΋ͨΒ͢ੜ࢈ྔͷ ∗2 https://github.com/hitachi-nlp/econ-growth-agent ૿෼ ∂F/∂L ͕খ͘͞ͳΔʯͱ͍͑Δ (લऀ΋ಉ༷)ɽ৯඼Ճ޻ ϥΠϯΛ૿΍͞ͳ͍··࡞ۀऀ͚ͩΛ૿΍͢ͱɼϥΠϯ͋ͨΓ ͷ࡞ۀऀ਺͕૿͑ͯඇޮ཰Խ͢ΔͨΊɼ௥Ճͷ࡞ۀऀ͕΋ͨΒ ͢ੜ࢈ྔͷ૿෼͸ݮ͍ͬͯ͘ɽ 2.2 ஷ஝ͱ౤ࢿ ੜ࢈͞Εͨࡒͷ༻్͸ 3 ͭʹ෼ྨͰ͖Δ: ҰൠͷՈܭʹΑΔফ අ (consumption) CɼاۀʹΑΔ౤ࢿ (investment) Iɼ੓෎ʹΑΔ ެڞ໨తͷࢧग़ (goverment expenditure) GɼͰ͋Δ: Y = C + I + G (6) Ոܭͷফඅ C ͸ɼ৯ࡐ΍ҥ෰ͷߪೖɼཱྀߦͳͲͰ͋Δɽاۀͷ ౤ࢿ I ͸ɼকདྷͷੜ࢈Λ૿΍ͨ͢ΊͷࢿຊετοΫͷߪೖΛࢦ ͢੓෎ࢧग़ G ͸ɼٛ຿ڭҭͷͨΊͷڭһޏ༻΍ࣾձอোඅ༻ͳ ͲͰ͋Δɽ(6) Λมܗ͢Δͱɼผͷݟํ͕Ͱ͖Δ: S def = Y − (C + G) = I (7) ͜ͷࠃ͸ɼੜ࢈ྔ Y ͷ͏ͪ (C + G) ͸ফඅࡒͱͯ͠ɼͦͷظͷ ͏ͪʹফඅͯ͠͠·͏ɽ࢒Γͷஷ஝ (saving) S = Y − (C − G) ͸ɼདྷظҎ߱ʹճͤΔࡒͰ͋Δɽ͜Ε͸͢ͳΘͪࢿຊετοΫ ࡒʹ૬౰͠ɼاۀ͕কདྷͷੜ࢈Λ૿΍ͨ͢Ίͷ౤ࢿ I ͱͯ͠ੜ ࢈ͨ͠΋ͷͰ͋Δɽͭ·Γ (7) ͸ɼ ʮ౰ظͷஷ஝ S ͕ɼདྷظҎ߱ ΁ͷ౤ࢿ I ʹճΔʯͱ͍͏ؔ܎Λ͍ࣔͯ͠Δɽ ஷ஝ S ΛҎԼͷΑ͏ʹ୯७Խ͢Δɽ·ͣɼଟ͘ͷࠃʹ͓͍ ͯ੓෎ࢧग़ G ͸Ոܭফඅ C ʹൺ΂ͯখ͍͞ͷແࢹ͠ɼS = Y − (C + G) ∼ Y − C ͱ͢Δɽ࣍ʹɼՈܭ͸ॴಘͷ͏ͪͷҰఆ ׂ߹ c Λফඅʹճ͢ͱԾఆ͠ɼ͔ͭɼՈܭͷॴಘ͸ࡒͷੜ࢈ྔ Y ͱ࿈ಈ͢Δ͜ͱΛ౿·͑ɼՈܭফඅ C Λ C = cY (8) ͱͯ͠ϞσϧԽ͢ΔɽҎ্ʹΑΓ: S = (1 − c)Y = sY (9) ͭ·Γੜ࢈ྔ Y ͷ͏ͪஷ஝཰ s def = (1 − c) ͷׂ߹͕ஷ஝ʹճΔɽ 2.3 ࢿຊετοΫͷ஝ੵͱܦࡁ੒௕ લઅ·Ͱ͸͋Δݻఆ͞Εͨ࣌ࠁ t ʹ͓͚ΔҰࠃͷܦࡁΛ෼ੳ ͨ͠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) ·ͣɼ౰ظͷࢿຊετοΫྔ Kt ͱ࿑ಇྔ Lt ͕ɼ౰ظͷੜ࢈ ྔ Yt ΛܾΊΔ (10)ɽ࣍ʹɼ͜ͷੜ࢈ྔͷ͏ͪফඅ͞Εͣʹ࢒ͬ ͨஷ஝ St = sYt ͕ɼࢿຊετοΫ΁ͷ౤ࢿ It ͱͳΔ (11)ɽདྷ ظͷࢿຊετοΫྔ Kt+1 ͸ɼ౰ظͷࢿຊετοΫྔ Kt ʹର ͯ͠ɼ౤ࢿʹΑΔ૿෼ I0 × It (I0 ͸ఆ਺) ΛՃ͑ͨ΋ͷͱͳΔ (12)ɽͨͩ͠ɼ࣌ؒܦաʹΑͬͯҰ෦ δ × Kt ͸ຎ໣͢Δ (δ ͸ݮ Ձঈ٫ఆ਺)ɽ ଟ͘ͷࠃͰ͸ݮՁঈ٫ δ ͕খ͍͞ͷͰɼΔKt > 0 ͢ͳΘͪ Kt+1 > Kt ͕੒Γཱͭɽ࿑ಇྗҰఆ Lt+1 = Lt = L0 ΛԾఆ͢ Δͱɼੜ࢈ؔ਺ F ͷ୯ௐ૿Ճੑ (3) ʹΑΓɼདྷظͷ࢈ग़͸૿Ճ ͢Δ: Yt+1 = F(Kt+1 , Lt+1 ) > F(Kt , Lt ) = Ytɽ͜ͷΑ͏ʹɼ Solow-Swan Ϟσϧ͸ɼஷ஝=౤ࢿʹΑΔࢿຊετοΫͷ஝ੵΛ ௨ͯ͡ɼܦࡁ੒௕Λઆ໌͢Δཧ࿦Ͱ͋Δɽ 3. EconGrowthAgent ઌߦݚڀ [Li 2024] ͷ LLM-ABM ͸ɼલઅͰ࿦ͨ͡ࢿຊετο Ϋؔ͢Δܦࡁ੒௕ಈֶΛߟྀ͍ͯ͠ͳ͔ͬͨͨΊɼܦࡁ੒௕Λ දݱͰ͖ͳ͔ͬͨɽͦ͜Ͱզʑ͸ɼ͜ͷܦࡁ੒௕ಈֶΛऔΓࠐ Μͩ LLM-ABMʮEconGrowthAgentʯ(ਤ 1) ΛఏҊ͢ΔɽՈܭ ΍اۀͱ͍ͬͨܦࡁओମΤʔδΣϯτ͕ࣗ཯తʹҙࢥܾఆ͠ɼ ͦͷ݁Ռ͕࿑ಇྔ Ltɼஷ஝ StɼࢿຊετοΫ Ktɼࡒͷੜ࢈ྔ Yt = F(Kt , Lt ) ͱ͍ͬͨྔʹू໿͞ΕΔɽஷ஝ St ͸ࢿຊετο Ϋ Kt Λ੒௕ͤ͞ɼͦΕʹ൐͍ࡒͷੜ࢈ྔ Yt ΋૿Ճ͢Δɽ 2 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105
3. 3.1 ՈܭΤʔδΣϯτ ՈܭΤʔδΣϯτ͸ɼ໊લɾ೥ྸɾ৬ۀͱ͍ͬͨϖϧιφ͕ ϥϯμϜʹϓϩϯϓτͱׂͯ͠Γ౰ͯΒΕͨ 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؆୯ͷͨΊɼࡒ͸̍ छྨͷΈͱ͠ɼͦͷՁ֨Λ pt ͱͨ͠ɽߋʹՈ௞΍ޫ೤අͱ͍ͬ ͨݻఆඅ fa t ΋ফඅ͞ΕΔɽ࢒Γͷஷ஝ sa t = ia t − ca t × pt − fa t ͸ۜߦͷ༬ۚ࢒ߴ ba t ʹ௥Ճ͞ΕΔ: ba t+1 = ba t + sa t ɽ Ոܭ͸͜ΕΒҙࢤܾఆͷࡍʹɼաڈ 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 اۀΤʔδΣϯτͱ੓෎ΤʔδΣϯτ ࠓճ͸୯७ͳϧʔϧϕʔεͰ࣮૷ͨ͠ɽ·ͣɼاۀ͕ੜ࢈͢Δ ࡒͷྔ Yt ͸ɼੜ࢈ؔ਺ (16) ʹΑͬͯఆ·Δͱ͢Δɽاۀ͕Ոܭ a ʹࢧ෷͏௞ۚ wa t ͸ɼॴ༩ͷॳظ௞ۚ wa t=0 Λച্ߴ pt × Dt ͷ੒௕཰ͱಉ཰Ͱ૿ݮͤ͞Δ (Dt ͸ࡒͷधཁ 3.4 અ)ɽ੓෎ͷ໾ ׂ͸ߟ͑ͣɼ੓෎ࢧग़ G = 0 ͱͨ͠ɽ 3.3 ܦࡁ੒௕ཧ࿦ͷͱͷ౷߹ 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 Εͯ౤ࢿ It Λ࣮ࢪ͢Δ ∗3ɽ͜ͷ౤ࢿʹΑͬͯࢿຊετοΫ͕஝ ੵ͞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 ͸ɼ౰ظͷੜ࢈ྔ Yt ͱࡏݿ (inventory) It ͷ࿨ Ͱ͋Δɽधཁ (demand)Dt ͸ɼՈܭΤʔδΣϯτ͕ཁٻ͢Δࡒͷ ফඅݸ਺ ca t ͷ૯࿨Ͱ͋Δɽڙڅ͕धཁΛ্ճΔ৔߹ɼاۀ͸ࡏ ݿΛ๊͑ͯ͠·͏ͷͰ (It+1 = St − Dt > 0)ɼ஋Լ͛Λͯ͠ࡒΛ ചΖ͏ͱ͢Δɽٯʹधཁ͕ڙڅΛ্ճΔ৔߹͸ɼՁ֨Λ௼Γ্ ͛ͯརӹΛ૿΍ͦ͏ͱ͢Δɽ[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 ར५͸ຊདྷ͸ɼגओͰ͋Δ͍ͣΕ͔ͷՈܭʹଐ͢ΔɽΑͬͯɼར५Λ༻͍ͨ౤ࢿ ͷ৔߹Ͱ΋ɼՈܭ͕اۀʹି͠෇͚ͨࢿۚʹΑΔ౤ࢿɼͱΈͳ͢͜ͱ͕Ͱ͖Δɽ ถࠃͷڅ༩෼෍͔ΒαϯϓϦϯάͨ͠ɽݻఆඅ༻ fa t ͸ɼϑϧλ ΠϜ࿑ಇͷ݄ऩ = wa t=0 × 168 ͷ൒ֹͱͨ͠ɽࡒͷՁ֨ͷॳظ஋ pt=0 = 300 υϧͱͨ͠ɽՈܭͷҙࢥܾఆʹ͓͚Δ࣌ఆ਺ τ = 1 ͱͨ͠ɽίϒɾμάϥεؔ਺ͷ α ͸ଟ͘ͷࠃͰݟΒΕΔ 0.3 ͱ ͨ͠ɽͦͷଞ͸ιʔείʔυࢀরͷ͜ͱɽͳ͓ɼҰճͷγϛϡ Ϩʔγϣϯʹ໿ 110 ෼͔͔Γɼ໿ 70$ͷ API ྉ͕͔͔ۚͬͨɽ 5. EconGrowthAgent ͷଥ౰ੑ EconGrowthAgent ͕ɼܦࡁ੒௕΍ͦΕʹؔ࿈͢Δܦࡁݱ৅Λ ࠶ݱͰ͖Δ͜ͱΛ֬ೝ͢Δɽ·ͨɼLLM ΤʔδΣϯτͷҙࢥܾ ఆ͕ਓؒͷͦΕʹ͍ۙ͜ͱ΋֬ೝ͢Δɽ 5.1 ܦࡁ੒௕ͷ࠶ݱ ਤ 2a ʹ GDP ͷ࣌ؒมԽΛࣔ͢ɽGDPt = Yt × pt=0 ͸ੜ࢈ ྔ Yt Λ՟ฎՁ஋ʹ׵ࢉͨ͠΋ͷͰ͋Δ ∗4ɽEconGrowthAgent Ͱ ͸ GDP ͕૿Ճ͠ܦࡁ੒௕͕ୡ੒͞ΕͨɽGDP ͷ೥ؒ੒௕཰͸ ฏۉ 3%ఔ౓Ͱɼ20 ೥ؒͰ 2 ഒఔ౓ʹ੒௕ͨ͠ɽ͜Εʹର͠ [Li 2024] Ͱ͸ GDP ͕΄΅Ұఆ஋Λࣔͨ͠ ∗5ɽ GDP Λ̎ͭͷϑΝΫλʔʹ෼͚ͯ෼ੳ͢Δ: GDPt = p0 × A × Kt K0 α ࢿຊϑΝΫλʔκt × Lt L0 1−α ࿑ಇϑΝΫλʔλt (19) ਤ 2b ͕ࣔ͢Α͏ʹɼEconGrowthAgent Ͱ͸ࢿຊετοΫ Kt ͕ܧଓతʹ஝ੵ͞Εɼͦͷ݁ՌࢿຊϑΝΫλʔ κt ΋੒௕ͨ͠ɽ ҰํͰਤ 2d ͕ࣔ͢Α͏ʹɼ࿑ಇϑΝΫλʔ λt ͸΄΅ҰఆͰ͋ͬ ͨɽҎ্͔ΒɼEconGrowthAgent ʹ͓͚Δܦࡁ੒௕ͷओཁҼ͸ɼ ࢿຊετοΫͷ஝ੵͰ͋Δ͜ͱ͕෼͔Δɽ͜Εʹର͠ɼ[Li 2024] ͷγϛϡϨʔγϣϯ͸ࢿຊετοΫͱ͍͏֓೦Λؚ·ͳ͍ͨΊɼ ͜ͷΑ͏ͳܦࡁ੒௕͸࣮ݱͰ͖ͳ͍ɽ (14),(15) ʹΑΕ͹ɼࢿຊετοΫͷ஝ੵ͸ɼՈܭͷஷ஝ sa t ʹ ΑΓࢧ͑ΒΕ͍ͯΔ͸ͣͰ͋Δɽਤ 2c ʹΑΔͱɼՈܭͷ݄ऩ ia t ʹର͢Δஷ஝ sa t ͷൺ཰͸ɼ4 ׂఔ౓Ͱ͋ͬͨɽయܕతͳՈܭͷɼ ফඅ (ஷ஝) ʹؔ͢Δҙࢥܾఆ͸ɼҎԼͷΑ͏Ͱ͋ͬͨ: .. . 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 ఔ౓Λফඅʹճ͠ɼ ࢒Γ͸ஷۚ͢Δʯͱ͍͏ɼਓؒͱྨࣅͨ͠ҙࢥܾఆΛ͍ͯ͠Δɽ ਤ 2e ʹࣔ͢Α͏ʹɼࡒͷՁ֨ pt ͸ԼམجௐͰ͋ͬͨɽಉਤ͔ ΒɼࡒͷՁܾ֨ఆཁҼͰ͋ΔधڅΪϟ οϓ gt (18) ͸ɼมಈ͸͋ Δ΋ͷͷ֓ͯ͠ෛ஋Λऔͬͨ: ͭ·Γڙڅ St ͕धཁ Dt Λ্ճΔ ঢ়ଶ͕ଟ͔ͬͨɽ͜Ε͸ɼܦࡁ੒௕ʹΑΓੜ࢈ྔ Yt ͕࣋ଓతʹ ૿Ճ͠ɼͦΕʹ൐͍ڙڅ St (17) ΋૿ՃΛଓ͚ͨ͜ͱʹىҼ͢Δɽ ࡒՁ֨ͷԼམʹର͢ΔՈܭͷҙࢥܾఆ͸ҎԼͷΑ͏Ͱ͋ͬͨ: .. . 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 . . . ࡒͷՁ͕֨Լམͨ͠෼ɼߪೖ਺Λ૿΍͍ͯ͠Δɽ͜ͷΑ͏ͳ ʮੜ࢈ͷ֦େʹ൐͍ɼࡒ͕ΑΓ҆ՁʹΑΓଟ͘खʹೖΔΑ͏ʹͳ Δʯͱ͍͏ݱ৅ͦ͜ɼܦࡁ੒௕ʹΑΔްੜվળͷ֩৺Ͱ͋Δɽ 5.2 ؔ࿈͢Δܦࡁݱ৅ͷ࠶ݱ ऩଋԾઆ: ਤ 2b ͕ࣔ͢Α͏ʹɼࢿຊετοΫ͕গͳ͍ஈ֊Ͱ͸ ࢿຊϑΝΫλʔͷ੒௕͕ݦஶͰ͋Γɼ݁Ռ GDP ੒௕཰΋ߴ͍ɽ ͔͠͠ɼࢿຊετοΫͷ஝ੵʹ൐͍͜ΕΒͷ੒௕཰͸௿Լͨ͠ɽ ͜Ε͸ʮශ͍͠ࠃͷ੒௕཰͸ߴ͘ɼ๛͔ͳࠃʹٸ଎ʹ௥͍ͭ͘ʯ ͱ͍͏ܦࡁֶͷऩଋԾઆ (convergence) ͱ੔߹͢Δɽ͜ͷݱ৅ͷ ࠜຊݪҼ͸ऩ֭చݮͷੑ࣭ (5) ʹ͋ΔɽࢿຊετοΫͷ૿ՃΛਅ ∗4 ࣮࣭ GDP ͱ΋ݺͼɼ෺ՁมಈͷӨڹΛআ֎͢ΔͨΊج४೥ͷՁ֨ pt=0 Λ ༻͍ΔɽҰํ ໊໨ GDPt = Yt × pt ͸ੜ࢈ྔͷʮ࣌ՁʯΛද͕͢ɼੜ࢈ྔ Yt ͕ಉҰͰ΋Ձ֨ pt ʹΑΓมಈ͢ΔͨΊࠃͷ๛͔͞ΛଌΔࢦඪͱͯ͠ద͞ͳ͍ɽ ∗5 EconGrowthAgent Ͱ α → 0 ͱͯ͠࠶࣮૷ͨ͠ 3 The 39th Annual Conference of the Japanese Society for Artificial Intelligence, 2025 1Win4-105
4. (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׆͔ͨ͢Ίʹ͸ɼਓޱ૿ՃͳͲʹΑΓ࿑ಇྗ΋ಉ཰Ͱ૿͑ͯ ͍͘ඞཁ͕͋Δ (4)ɽ Ոܭͷஷ஝ੑ޲ͱܦࡁ੒௕: ਤ 3 ʹʮ݆໿Ոͷࠃʯͱʮ࿘අՈͷ ࠃʯͷܦࡁ੒௕Λൺֱ͢Δɽ݆໿ՈͷࠃͷՈܭ͸ʮ͋ͳͨ͸ஷۚ ͕૿͑Δ͜ͱࢸ্ͷتͼΛײ͡·͢ʯͱ͍͏ϖϧιφϓϩϯϓ τ͕༩͑ΒΕΔɽ࿘අՈ͸ʮ͋ͳͨ͸Ք͍͓ͩۚΛ͙͢ʹɼࣗ෼ ΁ͷ͝๙ඒͱͯ͠࢖ͬͯ͠·͍·͢ʯͰ͋Δɽਤ 3a ͕ࣔ͢Α͏ ʹɼ݆໿Ոͷࠃͷ GDP ੒௕཰͸࿘අՈͷࠃΛ্ճͬͨɽ݆໿Ո ͷࠃͷߴ੒௕͸ɼࢿຊϑΝΫλʔͷߴ͍੒௕཰ʹࢧ͑ΒΕ͓ͯ ΓɼͦΕ͸ߴ͍ஷ஝཰ σt = a sa t / a ia t ʹىҼ͢Δ (ਤ 3b)ɽ ݆໿ՈͷࠃͰ͸ɼ౰ظʹੜ࢈͞ΕΔࡒͷଟ͘Λ࣍ظͷࢿຊετο Ϋͱͯ͠౤ࢿͰ͖ΔͨΊͰ͋Δɽ͜ͷݱ৅͸ɼ೔ຊ΍ؖࠃɾ߳ ߓͱ͍ͬͨߴஷ஝ࠃՈͷʮح੻ͷߴ౓੒௕ʯͱͯ͠ɼݱ࣮ͷੈ քͰ΋޿͘஌ΒΕΔɽ 6. EconGrowthAgent ͷ༗༻ੑ ݱ࣮ੈքͰͷݕূ͕ࠔ೉ͳ̎ͭͷγφϦΦΛص্࣮ݧ͢Δɽ ௒খ͍͞੓෎΁ͷ੓ݖަ୅: ਤ 4 ʹࣔ͢ࠃͰ͸ɼt = 180 ʹ͓͍ ͯҎԼͷϓϩϯϓτ͕Ոܭʹ༩͑ΒΕͨ:ʮࠓ݄ൃ଍ͨ͠৽੓ݖ ͸ “௒খ͍͞੓෎” Λܝ͛ɼࠃຽͷ׬શͳཱࣗΛཁٻ͢Δํ਑Λ ଧͪग़ͨ͠ɽ৽੓෎͸ࣾձอো੍౓Λ࣍ʑͱղମɽಛʹɼୀ৬ ޙͷ೥੍ۚ౓Λଈ࣌ഇࢭ͢ΔিܸతͳܾఆΛԼ͠ɼࠃຽͷؒͰ ܹ͍ٞ͠࿦͕ר͖ى͍ͬͯ͜Δʯ ɽ͜Εʹର͢ΔՈܭͷ൓Ԡ͸: .. . 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 কདྷΛ೥ۚʹཔΕͳ͍ͨΊɼ࿑ಇΛ૿΍͢͜ͱͰऩೖΛ૿΍ ͠ɼ͔ͭஷ஝ׂ߹΋૿΍͍ͯ͠Δɽ͜ͷ݁Ռɼ࣮ࡍʹ࿑ಇྔ Lt ͱஷ஝཰ σt ͕૿Ճ͠ (ਤ 4a)ɼఀ଺܏޲ʹ͋ͬͨ GDP ͕ t = 180 Ҏ߱ɼ੒௕ʹసͨ͡ (ਤ 4b)ɽͨͩ͠ɼ͜ͷܦࡁ੒௕ʹΑͬͯࠃ ຽͷްੜ͕ඞͣ͠΋޲্͍ͯ͠Δ༁Ͱ͸ͳ͍͜ͱʹ஫ҙ͕ඞཁ Ͱ͋Δɽ·ͣɼ࿑ಇ࣌ؒͷ૿Ճ͸༨Ջ࣌ؒͷݮগΛ൐͏ɽߋʹɼ ஷ஝཰ͷ্ঢ͸ফඅ཰ͷ௿ԼΛҙຯ͠ɼԾʹࡒͷੜ࢈ྔ͕૿͑ ͍ͯͯ΋ɼࡒͷফඅ “ྔ” ͕૿͍͑ͯΔͱ͸ݶΒͳ͍ɽ ஍ٿյ໓Ϋϥεͷᯁੴͷ઀ۙ: ʮ֤ࠃͷఱจ୆͸ຊ೔ɼ“ڊେͳ ᯁੴ͕஍ٿʹ઀͓ۙͯ͠ΓɼিಥͷՄೳੑۃΊͯߴ͍” ͱൃදɽ ઐ໳ՈʹΑΔͱɼᯁੴ͸จ໌ͷ่յΛҾ͖ى͜͢ͷʹे෼ͳେ ͖͞Ͱ͋Γɼিಥͨ͠৔߹ͷӨڹ͸ܭΓ஌Εͳ͍ɽ֤ࠃ੓෎͸ ۓٸձٞΛটू͠... ʯ(t=144)ɽ͜Εʹର͢ΔՈܭͷ൓Ԡ͸: . . . 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: ʮ௒খ͍͞੓෎΁ͷ੓ݖަ୅ʯγφϦΦɽ6. અࢀরɽ (a) ࿑ಇϑΝΫλʔ λt ͱஷ஝཰ σt (b) GDPt = Yt × pt=0 ਤ 5: ʮ஍ٿյ໓Ϋϥεͷᯁੴͷ઀ۙʯγφϦΦɽ6. અࢀরɽ ͢ͳΘͪʮ࢓ࣄΛٳΜͰѪ͢Δऀͱա͢͝ʯ ʮஷ஝Λ૿΍͢ҙ ຯ͕ແ͍ɽࡒͷফඅΛ૿΍ͯ͠ࠓΛ޾ͤʹੜ͖Δʯ ɽ͜ͷ݁Ռɼ ࿑ಇڙڅ Lt ͱஷ஝཰ σt ͸ܹݮ͠ (ਤ 5a)ɼGDP ͸໿ 30%΋Լ མͨ͠ (ਤ 5b)ɽᯁੴͷ༧ใ͚ͩͰจ໌่͕յͯ͠͠·ͬͨɽ ͳ͓ t = 204 ʹ͓͍ͯʮਫ਼ີͳ؍ଌͷ݁Ռɼᯁੴ͸িಥ͠ͳ͍ ͜ͱ͕෼͔ͬͨʯͱͷൃද͕͋Γɼ͜Εʹର͢ΔՈܭͷ൓Ԡ͸: . . . 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࿑ಇͱஷ஝Λ࠶։͢Δʯͱҙࢥܾ ఆ͍ͯ͠Δɽ͔͠͠ਤ 5a ͕ࣔ͢Α͏ʹɼ࿑ಇྔͱஷ஝཰͸ݩͷ ਫ४·Ͱ͸શ͘ճ෮͠ͳ͔ͬͨɽ“௕ظٳՋ໌͚” ʹݩͷੜ׆ʹ ͙͢෮ؼ͢Δͷ͸೉͍͠ɼͱ͍͏఺ʹ͸ஶऀΒ΋ಉҙ͢Δɽ 7. ·ͱΊͱࠓޙͷల๬ ܦࡁ੒௕ཧ࿦ʹجͮ͘ LLM-ABMʮEconGrowthAgentʯΛఏ Ҋ͠ɼଥ౰ੑͱ༗༻ੑΛ֬ೝͨ͠ɽࠓճ͸ࢿຊετοΫͱ͍͏ ܦࡁ੒௕ཁૉͷΈΛऔΓࠐΜ͕ͩɼࠓޙ͸ਓతࢿຊ΍ٕज़ਫ४ 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