Posted: 10/22/2019 7:23 PM
Post subject: AI learning exploits
Quote
DaTeL237
Active player (355)
Joined: 1/16/2008
Posts: 358
Location: The Netherlands
Thought this is worth sharing: btw AI showing unexpected (but goal-aligned) behavior is fairly common. But this video makes it easy to see how this'd work for games and TAS. Link to video
TASes: [URL=http://tasvideos.org/Movies-298up-Obs.html]Mr. Nutz (SNES), Young Merlin 100% (SNES), Animaniacs 100% (SNES)[/URL]
Posted: 10/23/2019 12:37 AM
Quote
BadPotato
Joined: 1/26/2009
Posts: 558
Location: Canada - Québec
Yes, this a very interesting video to show off what's possible. Yet, deep learning require quite a bit of setup and processing power. Since "one episode=one entire game", it would be wise to choose a simple game with precise objectives from state A to state B, or maybe a game which aim for the highest score and define when to end the game. An another issue: the current emulator used for TASing may not be optimized to generate sequences about "high volume of quick games" or use parallelism to exploit the GPU. Maybe one of the easiest way to get started is to start simple. Check project like atari_py and see if the atari binding are good enough to setup a test environment inside a Jupyter Notebook.
Posted: 10/29/2019 2:14 AM
Quote
endrift
Any
Emulator Coder
Joined: 12/14/2014
Posts: 161
I tried working on this; I used to work at that company. I (mostly) made Gym Retro for this, but the results so far have been...mixed.