SLIDE 1
General Atari 2600 Game Playing
Michael Bowling Work with: Joel Veness, Marc Bellemare, Anna Koop, Mostafa Vafadoost
http://www.arcadelearningenvironment.org
Friday, September 14, 2012
Friday, September 14, 2012
General Atari 2600 Game Playing Michael Bowling Work with: Joel - - PowerPoint PPT Presentation
General Atari 2600 Game Playing Michael Bowling Work with: Joel - - PowerPoint PPT Presentation
General Atari 2600 Game Playing Michael Bowling Work with: Joel Veness, Marc Bellemare, Anna Koop, Mostafa Vafadoost http://www.arcadelearningenvironment.org Friday, September 14, 2012 Friday, September 14, 2012
SLIDE 2
SLIDE 3
http://www.arcadelearningenvironment.org
Friday, September 14, 2012
SLIDE 4
Varied Many Independent Interesting
Friday, September 14, 2012
SLIDE 5
Friday, September 14, 2012
SLIDE 6
Friday, September 14, 2012
SLIDE 7
A.I.
0100010101...00001110101
Reinforcement Learning
Friday, September 14, 2012
SLIDE 8
A.I.
0100010101...00001110101
Planning
Friday, September 14, 2012
SLIDE 9
A.I.
Model
Model Learning
Friday, September 14, 2012
SLIDE 10
A.I.
Expert Imitation/Apprenticeship Learning
Friday, September 14, 2012
SLIDE 11
A.I.
. . .
Transfer Learning
Pitfall! Pitfall II
Friday, September 14, 2012
SLIDE 12
A.I.
Intrinsic Motivation
Friday, September 14, 2012
SLIDE 13
Friday, September 14, 2012
SLIDE 14
Training Games Testing Games
Friday, September 14, 2012
SLIDE 15
Friday, September 14, 2012
SLIDE 16
Contingency Awareness: knowing what you control
(Bellemare et al., AAAI 2012)
Friday, September 14, 2012
SLIDE 17
(Bellemare et al., AAAI 2012)
Contingency Awareness: knowing what you control Unaware Contingency Aware
Friday, September 14, 2012
SLIDE 18
(Bellemare et al., AAAI 2012)
Contingency Awareness: knowing what you control
1.0 0.5 0.0 0.2 0.4 0.6 0.8 1.0
Inter-Algorithm Score Fraction of Games
Inter-Algorithm Score Distribution
MaxCol Extended MaxCol Basic Extended
Friday, September 14, 2012
SLIDE 19
Sketch-Based Hashing: tug-of-war vs. standard hashing
(Bellemare et al., NIPS 2012)
1.0 0.5 0.0 0.2 0.4 0.6 0.8 1.0
Fraction of games Inter-algorithm score Tug-of-War Standard
Hash Table Size: 1000
1.0 0.5 0.0 0.2 0.4 0.6 0.8 1.0
Fraction of games Inter-algorithm score Tug-of-War Standard
Hash Table Size: 5000
1.0 0.5 0.0 0.2 0.4 0.6 0.8 1.0
Fraction of games Inter-algorithm score Tug-of-War Standard
Hash Table Size: 20,000
55 Testing Games
Friday, September 14, 2012
SLIDE 20
Model Learning: pixels, probabilities, and priors
(Bellemare et al., In Prep)
Friday, September 14, 2012
SLIDE 21
Model Learning: pixels, probabilities, and priors
(Bellemare et al., In Prep)
Friday, September 14, 2012
SLIDE 22
http://www.arcadelearningenvironment.org
Questions?
Source code for ALE and all agents available!
Friday, September 14, 2012
SLIDE 23
Friday, September 14, 2012
SLIDE 24
Will there be a competition? No vs.
Friday, September 14, 2012