Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
DeepRacer for learning RL
Search
貞松政史
April 06, 2019
Technology
0
1.3k
DeepRacer for learning RL
2019.4.6 Developers.IO at OKAYAMA.
貞松政史
April 06, 2019
Tweet
Share
More Decks by 貞松政史
See All by 貞松政史
Amazon Forecast亡き今、我々がマネージドサービスに頼らず時系列予測を実行する方法
sadynitro
0
1k
今日のハイライトをシステマティックに
sadynitro
1
70
はじめてのレコメンド〜Amazon Personalizeを使った推薦システム超超超入門〜
sadynitro
2
2.2k
予知保全利用を目指した外観検査AIの開発 〜画像処理AIを用いた外観画像に対する異常検知〜
sadynitro
0
1k
20230904_GoogleCloudNext23_Recap_AI_ML
sadynitro
0
880
Foundation Model全盛時代を生きるAI/MLエンジニアの生存戦略
sadynitro
0
980
Amazon SageMakerが存在しない世界線 のAWS上で実現する機械学習基盤
sadynitro
0
270
Amazon SageMakerが存在しない世界線のAWS上で実現する機械学習基盤
sadynitro
0
2k
みんな大好き強化学習
sadynitro
0
1.2k
Other Decks in Technology
See All in Technology
オブザーバビリティが育むシステム理解と好奇心
maruloop
2
1.3k
生成AI時代のPythonセキュリティとガバナンス
abenben
0
140
AIでデータ活用を加速させる取り組み / Leveraging AI to accelerate data utilization
okiyuki99
1
470
プレイドのユニークな技術とインターンのリアル
plaidtech
PRO
1
370
ストレージエンジニアの仕事と、近年の計算機について / 第58回 情報科学若手の会
pfn
PRO
3
850
会社を支える Pythonという言語戦略 ~なぜPythonを主要言語にしているのか?~
curekoshimizu
3
770
NLPコロキウム20251022_超効率化への挑戦: LLM 1bit量子化のロードマップ
yumaichikawa
3
500
あなたの知らない Linuxカーネル脆弱性の世界
recruitengineers
PRO
3
160
20251027_findyさん_音声エージェントLT
almondo_event
2
440
「タコピーの原罪」から学ぶ間違った”支援” / the bad support of Takopii
piyonakajima
0
140
ソースを読む時の思考プロセスの例-MkDocs
sat
PRO
1
220
SOTA競争から人間を超える画像認識へ
shinya7y
0
550
Featured
See All Featured
Optimising Largest Contentful Paint
csswizardry
37
3.5k
XXLCSS - How to scale CSS and keep your sanity
sugarenia
249
1.3M
Designing Dashboards & Data Visualisations in Web Apps
destraynor
231
53k
Six Lessons from altMBA
skipperchong
29
4k
The Straight Up "How To Draw Better" Workshop
denniskardys
238
140k
Rails Girls Zürich Keynote
gr2m
95
14k
Code Review Best Practice
trishagee
72
19k
Making Projects Easy
brettharned
120
6.4k
10 Git Anti Patterns You Should be Aware of
lemiorhan
PRO
658
61k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
52
5.7k
Let's Do A Bunch of Simple Stuff to Make Websites Faster
chriscoyier
508
140k
Large-scale JavaScript Application Architecture
addyosmani
514
110k
Transcript
4 D 29 26 1 . 0 I 1
& .-* (2 ,0'/4"# 51 83;7 +)
!&%$9( 6: Attention
3 #cmdevio2019
4 os t m ( L @S g E b
i _d L rMI D @ E ( ( ( ) ( e a n k AWS E
5 DeepRacer
6 D
7 ) (
8 …
9 DeepRacer 4 D 26 9 01 .
10 DeepRacer A A A
11
1 2 3
12 DeepRacer
13
14 DeepRacer 1/18
3D AWS DeepRacer League
15 DeepRacer https://aws.amazon.com/jp/deepracer/
16 DeepRacer ! &%$ +)*2 1 '/*2
(#-, 0. "
17 3D AWS RoboMaker Robot Operating System (ROS) Gazebo rqt
18 AWS DeepRacer League ⁻ 0 1 : 9 A
⁻ 9 2 R ⁻ ⁻ D I ⁻ 1 2 ⁻ https://aws.amazon.com/jp/deepracer/league/
19
20 (Artificial Intelligence, AI) (Machine Learning, ML)
NeuralNetwork DeepLearning
21
22 = 1 (
) ( (
23 L N - ) ( - D Q
24 DeepRacer Cliped PPO PPO (Proximal
Policy Optimization) OpenAI2017
25 ( ( )
)
26 1
27 ) () (
28 DeepRacer
29 DeepRacer + + +
30 DeepRacer
31 DeepRacer …
32 DeepRacer
33 orz
34 DeepRacer + + +
35 DeepRacer D A D
36 $ ' + (# &!
%"
37 ( ) ) https://docs.aws.amazon.com/ja_jp /deepracer/latest/developerguide/ deepracer-train-models-define- reward-function.html
38
39 ⁻ 10 ⁻
:
40 SageMaker RL + RoboMaker
41 SageMeker RLRoboMakerGA
42 SageMaker “RL” ⁻ ⁻ M ⁻ M M
⁻ M ⁻ ⁻ ⁻ J M S
43 DeepRacer ) D ( ) ) ( )
44 SageMaker
https://dev.classmethod.jp/machine -learning/sagemaker-robomaker- deepracer-sample/
45 $# "
! https://github.com/awslabs/amazon-sagemaker-examples
46 Jupyter !
47 ( ( )
48
49 2 1
50 2 2
51 (. ( )(
52 $2# " /1(+ $2#,- ! https://docs.aws.amazon.com/ja_jp/deepracer/latest/developerguide/deepracer -iteratively-enhance-reward-functions.html *
$2%) '.0& )
53 Best Practices when training with PPO
(Unity Technologies) https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Training-PPO.md
54 DeepRacer "% ! "%
$# ! !
55 DeepRacer
56 DeepRacer
57 DeepRacer
58 DeepRacer
59 DeepRacer
60
61 • g • + + • M D c
• R S D a k • b • LL e
62 DeepRacer
None