Slide 1

Slide 1 text

JAWS-UG AI #9
 ήʔϜӡӦͱػցֶश
 [SageMaker RL Λ࢖ͬͯΈͯ] ݹ৓ लོ mixi, Inc.

Slide 2

Slide 2 text

ݹ৓लོ ։ൃຊ෦ͨΜΆΆάϧʔϓ (ೖࣾ11೥໨) αʔό։ൃ
 |> ΞϓϦӡ༻
 |> ήʔϜ։ൃ
 |> ػցֶश ࣗݾ঺հ

Slide 3

Slide 3 text

Agenda ήʔϜͷ঺հͱ๊͑Δ໰୊఺ ػցֶशͷར༻ߏ੒ SageMaker RL Λ࢖ͬͨײ૝ͱཁ๬ P3 P9 P14

Slide 4

Slide 4 text

ήʔϜͷ঺հ

Slide 5

Slide 5 text

ϑΝΠτϦʔάͱ͸ LEADER HANDS LEADER LEADER HANDS LEADER HANDS

Slide 6

Slide 6 text

• ໿2िؒͷγʔζϯ੍ • γʔζϯ͝ͱʹΩϟϥΫλʔ͕ ௥Ճ͞ΕΔ ϑΝΠτϦʔάͱ͸

Slide 7

Slide 7 text

• ৽نϑΝΠλʔ௥ՃͰήʔϜόϥϯε่͕ΕΔ͜ͱ͕ൃੜ • ༧૝֎ͷ૊Έ߹Θͤͯγφδʔ͕ൃੜ͢Δέʔε • ૝ఆҎ্ʹڧ͗ͨ͢Χʔυ • ৽͍͠ϧʔϧมߋΛ͢ΔࡍʹશͯͷΧʔυͷӨڹΛௐࠪ͠ͳ͚Ε͹͍͚ͳ͍ ๊͑Δ໰୊ • ਓखͰҡ࣋͢Δͷ͸ίετ͕ߴ͍ • ػցͷྗʹཔΓ͍ͨ • ࠷ڧͷAIΛ࡞ͬͯόϥϯεௐ੔ͷςετϓϨΠΛ͓ئ͍͍ͨ͠

Slide 8

Slide 8 text

ػցֶशͷಋೖ؀ڥ

Slide 9

Slide 9 text

γεςϜߏ੒ Client Application(Ϣʔβ୺຤) APP Server(Socket Layer) APP Server(Logic Layer) RPC AWS • Logic Server͕
 ήʔϜϩδοΫΛ͍࣋ͬͯΔ • Client͸ԋग़෦෼ͷΈ • ϢʔβͷήʔϜதͷߦಈ͸
 ϩάͱͯ͠อଘՄೳ

Slide 10

Slide 10 text

൫໘৘ใͷCNN ωοτϫʔΫ εΩϧ৘ใͷ ϕΫτϧԽ पล৘ใͷDNN બ୒Մೳߦಈͷ֬཰

Slide 11

Slide 11 text

γεςϜߏ੒ Client Application(Ϣʔβ୺຤) APP Server(Socket Layer) APP Server(Logic Layer) RPC AWS • ϓϨΠϠʔϩάΛ࢖ͬͨ
 ڭࢣ͋ΓֶशͷϞσϧ • SageMaker EndpointͰ
 ഑ஔ݁ՌΛਪ࿦ • ϞσϧͷAPIΤϯυϙΠϯτԽ • ඇఀࢭʹΑΔϞσϧ੾Γସ͑ • εέʔϧ΋Մೳ SageMaker Endpoint

Slide 12

Slide 12 text

ˇ πʔϧԽ

Slide 13

Slide 13 text

ڧԽֶश΁ͷऔΓ૊Έ

Slide 14

Slide 14 text

ڧԽֶश • ΑΓΩϟϥΫλʔʹ߹ΘֶͤͨशΛ͢ΔͨΊʹڧԽֶशͷΞϓϩʔν΋ඞཁ • ࣮૷ͨ͠εΩϧΛ౤ೖલʹ͖ͪΜͱݕূ͍ͨ͠ • ৽͍͠ରઓ؀ڥͷӨڹΛڭࢣ͋Γ͚ͩͰ͸ֶशͰ͖ͳ͍ • ຊ౰ʹτοϓϓϨΠϠʔͱಉ͡ڧ͞ʹ͸౸ୡͰ͖͍ͯͳ͍ • ࣗલͰ࣮૷͍͕ͯͨ͠Ͱ͖Δ͜ͱͳΒେ͖ͳྗʹ૬৐Γ͔ͨͬͨ͠ • SageMaker RLͷbetaςετʹࢀՃͤͯ͞΋Β͑ͨͷͰࢼͯ͠Έͨ

Slide 15

Slide 15 text

SageMaker RL beta (ਫฏεέʔϧ) Docker image Training Job Docker image Docker image Docker image (Trainer) Docker image S3 bucket Notebook
 Instance Redis Weights Experience Replay

Slide 16

Slide 16 text

ݕূͨ͠಺༰ • ϑϨʔϜϫʔΫʹ৐Βͣʹࣗ෼ͷ؀ڥΛͦͷ·· • ͜ͷ৔߹͸࿈ܞ͕Ͱ͖ͳ͍ͷͰ1node͔͠ಈ͔ͤͳ͍ Training Job Docker image Notebook
 Instance LogicServer
 (elixir) GameEnv (python) ࣗ࡞RL ϩδοΫ
 (pyton) http S3 bucket model & checkpoint

Slide 17

Slide 17 text

ݕূͨ͠಺༰ • ਪ঑ͷCoach & GymܗࣜͰ࠶࣮૷ • ͜ͷܗͩͱ؆୯ʹਫฏ෼ׂͷԸܙΛड͚ΕΔ Training Job Docker image Notebook
 Instance LogicServer
 (elixir) GymEnv (python) http S3 bucket model & checkpoint Intel Coach

Slide 18

Slide 18 text

ݕূͰͷײ૝ • े෼࣮༻Ͱ͖ΔϨϕϧ • ͨͩ͠ɺ·ͩ໨ʹݟ͑Δൣғ಺ʹటष͍෦෼͕ͨ͘͞Μ࢒͍ͬͯΔ • ग़ྗ͞ΕΔσʔλ͸Tensorflowͷ΋ͷͳͷͰɺ৭ʑऔΓճ͕͠ޮ͘ • CoachͱGymͷྲّྀʹ৐ΔͱΞϧΰϦζϜͷ࣮૷ͳͲΘ͔Βͳ͍Ͱ΋࢖͑Δ • ٯʹGymͳͲͷྲّྀʹ৐Βͳ͍ͱ͍͚ͳ͘ͳΔ(single agentͳͲ) • Ray RLib+ multi agentͳͲ΋ࢼͯ͠Έ͍ͨ • ߋ৽ͷεϐʔυ͕ͱʹ͔͘ૣ͍ • ·ͩ·ͩൃలதͷͨΊɺॻ͖ํ͕ͲΜͲΜ৽͘͠ͳ͍ͬͯ͘ • ·ͩαϯϓϧίʔυಡΈղ͍࣮ͯ૷͍ͯ͘͠ײ͸͋ΔͷͰ
 υΩϡϝϯτྨͷॆ࣮ʹ΋ظ଴

Slide 19

Slide 19 text

Thank you !!