Value Iteration Networks Aviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine and Pieter Abbeel @ UC Berkeley NIPS 2016 チチチチチチチチチ チチチ チチチチチ チチチチチ ＠ 2017/03/09
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Value Iteration Networks
Value Iteration NetworksAviv Tamar, Yi Wu, Garrett Thomas, Sergey Levine and Pieter Abbeel @ UC BerkeleyNIPS 2016 2017/03/09
2
Qhttps://www.slideshare.net/yamaryox/20160421-70945023
3
Q Q
Q
Q4
Qreactive
5
CNN
6
Value Iteration Networks
7
s x, y s (s) a r
( (s), a ) 8
(s)
maxmax...
Value Iteration Value Iteration
Value Iteration Networks10
(s)
maxmax...
(s)
maxmax...
CNNCNN
Value Iteration Networks
(s) CNN
Conv
max& softmax13
13
Value Iteration Networks
CNNBack-Propagation14
15
Grid-World
Mars Rover Navigation
Continuous Control
Grid-World8x8, 16x16, 28x28 3x3CNN conv1=3x3x150, conv2=3x3x1 10, 20, 36 5000 7
CNNFCN(NN)17
Grid-World 18
Grid-World VIN 19
Grid-World 20
Grid-World 21
Mars Rover Navigation128x128108CNN16x16Conv(5x5x6), MaxPool(4x4), Conv(3x3x12), MaxPool(2x2), Conv(3x3x150), Conv(3x3x1) 10,000 7
22
Mars Rover Navigation 23
Mars Rover Navigation VIN84.8%
CNN90.3%
VIN24
Continuous Control(x, y, vx, vy)16x163x3NNCNN Conv1(3x3x150), Conv2(3x3x1)20040
25
Continuous Control
(s) CNN
Conv
max(5x5)& x 326
26
Continuous Control CNN27
Continuous Control
28
29
Value Iteration NetworksEnd-to-End
CNN
30
Web
VIN
31
Related Documents
