Post subject: Using lua to build a self learning AI
adelikat
He/Him
Emulator Coder, Site Developer, Site Owner, Expert player (3584)
Joined: 11/3/2004
Posts: 4738
Location: Tennessee
SethBling, the man who brought you realtime SMW ACEing and a new world record is at it again. This time, he built a self learning robot that learned to play Super Mario World. Thought you guys might enjoy this Give it a few eons and this script could obsolete the TASers, lol. If anyone wants to try this out for themselves and is too lazy to get the script and Bizhawk and all that, here's a convenient download
It's hard to look this good. My TAS projects
Editor, Expert player (2459)
Joined: 4/8/2005
Posts: 1573
Location: Gone for a year, just for varietyyyyyyyyy!!
How about using BOINC to make it learn quicker? Who needs protein folding, when you can fold SMW.
Amaraticando
It/Its
Editor, Player (157)
Joined: 1/10/2012
Posts: 673
Location: Brazil
This script is not working for me, not even after renaming "Super Mario World (USA)" to the name of my ROM (inside the script). Worked for somebody here?
Active player (469)
Joined: 2/1/2014
Posts: 928
I watched the video when it was on reddit. I love watching stuff like this ( and coding [sockfolder]). I would love to watch this script transform into something bigger. Is this only set up to be used for mario, could it be set up for games similar? would the evolution of this script in the mario world lead to a faster time each generation?
Editor, Skilled player (1504)
Joined: 7/9/2010
Posts: 1317
That was really interessting! I'm wondering how many generations it would take to overcome a human TASer, probably too much to fit in our life span? Does computer architecture on hardware basis that is like a neural network exist? Maybe that would make it more efficient, since you don't need to simulate the neural network then.
Favorite animal: STOCK Gt(ROSA)26Sortm1.1(rtTA,EGFP)Nagy Grm7Tg(SMN2)89Ahmb Smn1tm1Msd Tg(SMN2*delta7)4299Ahmb Tg(tetO-SMN2,-luc)#aAhmb/J YouTube Twitch
Patashu
He/Him
Joined: 10/2/2005
Posts: 4014
TASeditor wrote:
That was really interessting! I'm wondering how many generations it would take to overcome a human TASer, probably too much to fit in our life span? Does computer architecture on hardware basis that is like a neural network exist? Maybe that would make it more efficient, since you don't need to simulate the neural network then.
The level of information it can process is too coarse to be able to create anything remotely like a TAS. For example: -It currently can't tell different types of ground or different types of enemies apart. -I think it can only change movement once every X frames, to save on processing power.
My Chiptune music, made in Famitracker: http://soundcloud.com/patashu My twitch. I stream mostly shmups & rhythm games http://twitch.tv/patashu My youtube, again shmups and rhythm games and misc stuff: http://youtube.com/user/patashu
Active player (469)
Joined: 2/1/2014
Posts: 928
SethBling in the reddit post had said that the AI didnt fully understand holding the same P-speed as a speedrunner would do, so the AI slows down to avoid enemies and etc.
Player (33)
Joined: 2/16/2012
Posts: 282
Reminds me a bunch of the following, although with different goals and methods. There are a few more videos of his software playing other games too. Link to video
Masterjun
He/Him
Site Developer, Skilled player (1970)
Joined: 10/12/2010
Posts: 1179
Location: Germany
Omnigamer wrote:
Reminds me a bunch of the following, although with different goals and methods. There are a few more videos of his software playing other games too.
Just adding the thread for that.
Warning: Might glitch to credits I will finish this ACE soon as possible (or will I?)