Slide 1

Slide 1 text

DNEFWJPDNEFWJP) %FFQ3BDFSͰ࢝ΊΔػցֶश σʔλΞφϦςΟΫεࣄۀຊ෦ΠϯςάϨʔγϣϯ෦ େᖒ༐ే

Slide 2

Slide 2 text

ࣗݾ঺հ େᖒ༐ే σʔλΞφϦςΟΫεࣄۀຊ෦ΠϯςάϨʔγϣϯ෦ ػցֶशνʔϜΤϯδχΞ!େࡕ %FFQ3BDFSྺ൒೥͘Β͍

Slide 3

Slide 3 text

ࠓ೔࿩͢͜ͱ w"84%FFQ3BDFSͱ͸ wػցֶशͱ͸ w"84ʹ͓͚Δػցֶश w"84%FFQ3BDFSͰԿΛͲ͏ֶ΂Δͷ͔ wػցֶशΛΑΓֶͿͨΊʹ w"84%FFQ3BDFSΛ΍ͬͯΈΔ

Slide 4

Slide 4 text

DNEFWJPDNEFWJP) "84%FFQ3BDFSͱ͸

Slide 5

Slide 5 text

DNEFWJPDNEFWJP) https://aws.amazon.com/jp/deepracer/

Slide 6

Slide 6 text

"84%FFQ3BDFS "84্ͰϞσϧΛֶशͤͯ͞ɺϥδίϯͰࣗಈӡస wϚωδϝϯτίϯιʔϧ͔ΒϞσϧΛֶश wεέʔϧͷϥδίϯʹϞσϧΛࡌͤΔ͜ͱͰࣗಈӡస wόʔνϟϧαʔΩοτʹϞσϧΛఏग़͠ɺϥοϓλΠϜΛڝ͏ w"84SF*OWFOU΍"844VNNJUͰͷ%FFQSBDFS-FBHVFͰ࣮ ࡍʹ૸ΒͤͯϥοϓλΠϜΛڝ͏

Slide 7

Slide 7 text

DNEFWJPDNEFWJP) https://aws.amazon.com/jp/deepracer/

Slide 8

Slide 8 text

Ϛωδϝϯτίϯιʔϧ

Slide 9

Slide 9 text

"84%FFQ3BDFS-FBHVF

Slide 10

Slide 10 text

Ϟσϧͷֶश

Slide 11

Slide 11 text

γϛϡϨʔγϣϯ ֶशॳظ

Slide 12

Slide 12 text

γϛϡϨʔγϣϯ ͋Δఔ౓ֶशͨ͠ஈ֊

Slide 13

Slide 13 text

DNEFWJPDNEFWJP) ػցֶशͱ͸

Slide 14

Slide 14 text

ػցֶश֓ཁ ͬ͘͟Γݴ͏ͱنଇੑΛֶशͤ͞Δ͜ͱɻ wڭࢣ͋Γֶश౴͑Λ༩͑ͯೖྗͱ౴͑ͷنଇੑΛֶशͤ͞Δ wڭࢣͳֶ͠शɿ౴͑Λ༩͑ͣʹೖྗͷنଇੑΛֶशͤ͞Δ w൒ڭࢣ͋Γֶशɿجຊతʹ͸౴͑Λ༩͑ͣʹೖྗͷنଇੑΛֶशͤ͞ Δ͕ɺ෦෼తʹ౴͑Λ༩͑ͯิॿ͢Δ wڧԽֶशɿ͋Δ؀ڥͰใु͕࠷େԽ͢Δߦಈબ୒ͷنଇੑΛֶशͤ͞Δ ͳͲ

Slide 15

Slide 15 text

ػցֶशྫ ҟৗݕग़

Slide 16

Slide 16 text

ػցֶशྫ ୯ޠͷຒΊࠐΈ

Slide 17

Slide 17 text

ػցֶशྫ ΫϥελϦϯά

Slide 18

Slide 18 text

ڧԽֶश֓ཁ ؀ڥ͔ΒΤʔδΣϯτʹঢ়ଶ ΤʔδΣϯτ͕ঢ়ଶʹج͍ͮͯ ߦಈ ؀ڥ͕ߦಈ݁Ռʹج͍ͮͨใु ΛΤʔδΣϯτ΁ dΛ܁Γฦ͢ ௕ظతʹ࠷େใु͕΋Β͑ΔߦಈΛબͿΑ͏ʹֶश͢Δ

Slide 19

Slide 19 text

%FFQ3BDFSͰ࢖ΘΕ͍ͯΔख๏ 1SPYJNBM1PMJDZ0QUJNJ[BUJPO https://nervanasystems.github.io/coach/components/agents/policy_optimization/ppo.html

Slide 20

Slide 20 text

DNEFWJPDNEFWJP) "84ʹ͓͚Δػցֶश

Slide 21

Slide 21 text

"84ʹ͓͚Δػցֶश

Slide 22

Slide 22 text

"84%FFQ3BDFS RoboMakerͰͷγϛϡϨʔγϣϯͱ SageMakerͰͷϞσϧͷֶशΛ܁Γฦ͢

Slide 23

Slide 23 text

%FFQ3BDFSʹؔ܎͢Δ"84αʔϏε ࠓճؔ܎͢Δͷ͸ʜ

Slide 24

Slide 24 text

%FFQ3BDFSʹؔ܎͢Δ"84αʔϏε ࠓճؔ܎͢Δͷ͸ʜ Ϟσϧͷֶश

Slide 25

Slide 25 text

"NB[PO4BHF.BLFS ػցֶशͷϑϧϚωʔδυαʔϏε w+VQZUFSϊʔτϒοΫͷϗεςΟϯά wػցֶशϞσϧͷֶश wػցֶशϞσϧͷϗεςΟϯά wϥϕϧ͚ͮ Ξϊςʔγϣϯ ͷ؅ཧ ͳͲ DeepRacerͰ࢖༻

Slide 26

Slide 26 text

%FFQ3BDFSʹؔ܎͢Δ"84αʔϏε ࠓճؔ܎͢Δͷ͸ʜ Ϟσϧͷֶश AWS RoboMaker γϛϡϨʔγϣϯ

Slide 27

Slide 27 text

"843PCP.BLFS ϩϘοτΞϓϦέʔγϣϯͷ։ൃΛαϙʔτ͢ΔαʔϏε wϑϧϚωʔδυͳγϛϡϨʔγϣϯαʔϏε w$MPVEͱ࿈ܞͨ͠։ൃ؀ڥ w304ͷΫϥ΢υ֦ு wϑϦʔτϚωδϝϯτ ͳͲ DeepRacerͰ࢖༻

Slide 28

Slide 28 text

"84%FFQ3BDFS RoboMakerͰͷγϛϡϨʔγϣϯͱ SageMakerͰͷϞσϧͷֶशΛ܁Γฦ͢

Slide 29

Slide 29 text

DNEFWJPDNEFWJP) %FFQ3BDFSͰ ԿΛͲ͏ֶ΂Δͷ͔

Slide 30

Slide 30 text

ػցֶश΍ڧԽֶशͷྲྀΕ΍֓ཁ ʹ͍ͭͯମݧ&ֶ΂Δ

Slide 31

Slide 31 text

ػցֶश΍ڧԽֶशͷྲྀΕ΍֓ཁͬͯʜ wػցֶशͱڧԽֶशͷงғؾ wͲͷΑ͏ʹͯ͠ղܾ͢Δ͔ߟ͑ΔػցֶशλεΫͷઃܭ wԿ౓΋܁Γସ͑͢Ϟσϧͷվળ wᷚʹΑ͘ฉ͘ϋΠύʔύϥϝʔλ ͳͲ

Slide 32

Slide 32 text

ػցֶशλεΫͷઃܭ wػցֶशΛ࢖ͬͨ໰୊ͷղ͖ํ΁ͷམͱ͠ࠐΈ wใुؔ਺ͷઃܭ wͲ͏͢Ε͹ϥοϓλΠϜ͕଎͘ͳΔͷ͔ wͲͷΑ͏ʹใुΛ༩͑Ε͹ɺϥοϓλΠϜ͕଎͍૸ΓํΛֶ श͢Δͷ͔

Slide 33

Slide 33 text

ใुؔ਺

Slide 34

Slide 34 text

ߦಈύλʔϯ

Slide 35

Slide 35 text

Ϟσϧͷվળ wγϛϡϨʔγϣϯϩάΛ࢖ͬͨϞσϧͷಛ௃෼ੳ wҙਤͨ͠௨Γʹֶशͯ͠Δ͔ɺ૸ͬͯΔ͔Λ֬ೝ͠ɺվળʹ ͭͳ͛Δ wΑΓ଎͘૸ΔΑ͏ʹ͢ΔͨΊʹ࣮ݧΛ܁Γฦ͢

Slide 36

Slide 36 text

ใु

Slide 37

Slide 37 text

ͲͷΑ͏ʹ૸ͬͨͷ͔

Slide 38

Slide 38 text

ϋΠύʔύϥϝʔλͷௐ੔ wϞσϧͷֶशͷਐΊํΛௐ੔͢Δ wχϡʔϥϧωοτϫʔΫ΍ޯ഑߱Լ๏ɺ110ͳͲͷಛੑʹԠ ͨ͡࠷దԽ wεϐʔυΛ্͛ͨΓ͢Δͱɺֶश͕೉͘͠ͳΓɺֶश཰ͳͲ ͷύϥϝʔλΛௐ੔͢Δඞཁ͕Ͱͯ͘Δ

Slide 39

Slide 39 text

%FFQ3BDFSʹ͓͚ΔϋΠύʔύϥϝʔλ

Slide 40

Slide 40 text

ԿΛֶ΂Δͷ͔ʜ wػցֶशλεΫͷઃܭ wػցֶशΛ࢖ͬͨ໰୊ͷղ͖ํ΁ͷམͱ͠ࠐΈ w଎͍ϥοϓλΠϜ͕ग़͍ͨ͠Ͳ͏૸ΒͤΕ͹ɺͲ͏ֶशͤ͞Ε͹͍͍ͷ͔ʜ wϞσϧͷվળ wγϛϡϨʔγϣϯϩάΛ࢖ͬͨϞσϧͷಛ௃෼ੳ wҙਤͨ͠௨Γʹֶशͯ͠Δ͔ɺ૸ͬͯΔ͔֬ೝ͠ɺվળʹͭͳ͛Δ wϋΠύʔύϥϝʔλͷௐ੔ wϞσϧͷֶशͷਐΊํΛௐ੔͢Δ

Slide 41

Slide 41 text

ԿΛֶ΂Δͷ͔ʜ wػցֶशλεΫͷઃܭ wػցֶशΛ࢖ͬͨ໰୊ͷղ͖ํ΁ͷམͱ͠ࠐΈ w଎͍ϥοϓλΠϜ͕ग़͍ͨ͠Ͳ͏૸ΒͤΕ͹ɺͲ͏ֶशͤ͞Ε͹͍͍ͷ͔ʜ wϞσϧͷվળ wγϛϡϨʔγϣϯϩάΛ࢖ͬͨϞσϧͷಛ௃෼ੳ wҙਤͨ͠௨Γʹֶशͯ͠Δ͔ɺ૸ͬͯΔ͔֬ೝ͠ɺվળʹͭͳ͛Δ wϋΠύʔύϥϝʔλͷௐ੔ wϞσϧͷֶशͷਐΊํΛௐ੔͢Δ

Slide 42

Slide 42 text

৭ʑ΍ͬͯΈΔͱɺΑ͘෼͔Βͳ͍ͱ͜Ζ͕ग़ͯ͘Δ wֶश཰ͬͯ w//΍ޯ഑߱Լ๏ʹ͍ͭͯௐ΂ͯΈΔ wใुͬͯͲ͏༩͑Ε͹͍͍ͷ͔ wใुؔ਺ͷઃܭํ๏΍110ʹ͍ͭͯௐ΂ͯΈΔ ͳͲͳͲ

Slide 43

Slide 43 text

ͳΜ͔Α͘෼͔ΒΜ͚Ͳ ৭ʑ஌Γͨ͘ͳͬͯ͘Δ

Slide 44

Slide 44 text

DNEFWJPDNEFWJP) ػցֶशΛΑΓֶͿͨΊʹ

Slide 45

Slide 45 text

ػցֶशΛΑΓֶͿͨΊͷ͍Ζ͍Ζ w"84%FFQ3BDFS%SJWFOCZ3FJOGPSDFNFOU-FBSOJOH wIUUQTXXXBXTUSBJOJOH%FUBJMTF-FBSOJOH JE w$IBJOFS5VUPSJBM wIUUQTUVUPSJBMTDIBJOFSPSHKBJOEFYIUNM w"NB[PO4BHF.BLFS&YBNQMFT wIUUQTHJUIVCDPNBXTMBCTBNB[POTBHFNBLFSFYBNQMFT

Slide 46

Slide 46 text

"84%FFQ3BDFS%SJWFOCZ3FJOGPSDFNFOU-FBSOJOH https://www.aws.training/Details/eLearning?id=35393

Slide 47

Slide 47 text

$IBJOFSνϡʔτϦΞϧ https://tutorials.chainer.org/ja/index.html

Slide 48

Slide 48 text

$IBJOFSνϡʔτϦΞϧ https://tutorials.chainer.org/ja/tutorial.html

Slide 49

Slide 49 text

"NB[PO4BHF.BLFS&YBNQMFT https://github.com/awslabs/amazon-sagemaker-examples

Slide 50

Slide 50 text

"NB[PO4BHF.BLFS&YBNQMFT https://github.com/awslabs/amazon-sagemaker-examples

Slide 51

Slide 51 text

"NB[PO4BHF.BLFS&YBNQMFT https://github.com/awslabs/amazon-sagemaker-examples

Slide 52

Slide 52 text

%JTUSJCVUFE%FFQ3BDFS3-USBJOJOHXJUI4BHF.BLFSBOE3PCP.BLFS https://github.com/awslabs/amazon-sagemaker-examples/blob/master/reinforcement_learning/ rl_deepracer_robomaker_coach_gazebo/deepracer_rl.ipynb

Slide 53

Slide 53 text

DNEFWJPDNEFWJP) %FFQ3BDFSΛ΍ͬͯΈΔ

Slide 54

Slide 54 text

Ϛωδϝϯτίϯιʔϧ

Slide 55

Slide 55 text

ϞσϧҰཡ

Slide 56

Slide 56 text

Ϟσϧ࡞੒

Slide 57

Slide 57 text

Ϧιʔεͷ࡞੒

Slide 58

Slide 58 text

Ϟσϧ໊

Slide 59

Slide 59 text

ίʔεબ୒

Slide 60

Slide 60 text

ߦಈઃఆ

Slide 61

Slide 61 text

ߦಈύλʔϯ

Slide 62

Slide 62 text

ใुؔ਺

Slide 63

Slide 63 text

ใुؔ਺ͷαϯϓϧ

Slide 64

Slide 64 text

ϋΠύʔύϥϝʔλ

Slide 65

Slide 65 text

ֶश࣌ؒ

Slide 66

Slide 66 text

ֶश։࢝

Slide 67

Slide 67 text

ϞσϧҰཡ

Slide 68

Slide 68 text

ֶशը໘

Slide 69

Slide 69 text

3PCP.BLFSͷγϛϡϨʔγϣϯπʔϧ

Slide 70

Slide 70 text

(B[FCPͰͷγϛϡϨʔγϣϯը໘

Slide 71

Slide 71 text

ֶश׬ྃ

Slide 72

Slide 72 text

όʔνϟϧαʔΩοτ

Slide 73

Slide 73 text

όʔνϟϧαʔΩοτ

Slide 74

Slide 74 text

όʔνϟϧαʔΩοτʹొ࿥͢Δ

Slide 75

Slide 75 text

όʔνϟϧαʔΩοτʹొ࿥͢Δ

Slide 76

Slide 76 text

όʔνϟϧαʔΩοτʹొ࿥͢Δ

Slide 77

Slide 77 text

Ϟσϧαϒϛοτޙ

Slide 78

Slide 78 text

ϞσϧධՁத

Slide 79

Slide 79 text

ϞσϧධՁ׬ྃ ׬૸ͤͣ…

Slide 80

Slide 80 text

DNEFWJPDNEFWJP) ·ͱΊ

Slide 81

Slide 81 text

·ͱΊ w%FFQ3BDFSͰ͸ػցֶशɺڧԽֶशʹ৮ΕΔ͜ͱ͕Ͱ͖Δ wࣄલ஌ࣝ͸΄ͱΜͲඞཁͳ͘ɺखܰʹ࢝ΊΒΕΔ wແྉ࿮΋͋ΔͨΊɺͱΓ͋͑ͣ΍ͬͯΈΔͷ͕͓͢͢Ί wͨͩ͠ɺແྉ࿮Λ௒͑Δͱ҆͘͸ͳ͍ͷͰɺ΍Γ͗͢஫ҙ ڵຯΛ࣋ͬͨ಺༰ΛͲΜͲΜ۷ΓԼ͍͛ͯͬͯ Ώ͘Ώ͘͸ػցֶशڧऀʹʜ

Slide 82

Slide 82 text

͋Γ͕ͱ͏͍͟͝·ͨ͠