Do you have adblock enabled?
 
If you can read this, either the style sheet didn't load or you have an older browser that doesn't support style sheets. Try clearing your browser cache and refreshing the page.

(HypeAngel)   Guy creates an artificial intelligence that learns how to play and beat a Super Mario Bros. level   ( hypeangel.com) divider line
    More: Interesting, bros, artificial intelligences  
•       •       •

3465 clicks; posted to Video » on 14 Jun 2015 at 3:04 PM (2 years ago)   |   Favorite    |   share:  Share on Twitter share via Email Share on Facebook   more»



50 Comments     (+0 »)
 
View Voting Results: Smartest and Funniest
 
2015-06-14 01:58:59 PM  
Very very cool.  I thought this was going to be bullshiat at first but he actually did teach a a computer to learn at a very fundamental level.
 
2015-06-14 03:18:51 PM  
Neural networks are positively spooky.
 
2015-06-14 03:20:14 PM  
So which of those networks are for going after Sarah Conner?
 
2015-06-14 03:22:02 PM  
Okay but can we run our own copy? I always wanted to watch a blank slate learn.
 
2015-06-14 03:27:05 PM  
There was something similar to this a few years back... Hm
https://www.youtube.com/watch?v=xOCurBYI_gY  and  http://www.cs.cmu.edu/~tom7/mario/mario.pdf
I believe you can get your own version of that...
 
2015-06-14 03:28:51 PM  

Pichu0102: Okay but can we run our own copy? I always wanted to watch a blank slate learn.


Sure, the source code for the plugin he used is in the video description.
 
2015-06-14 03:31:19 PM  
Is it ironic that it is a human created AI defeating a human's program emulating an AI?.
 
2015-06-14 03:36:40 PM  
AI programmers have been doing this for awhile: algorithms learning how to play video games.
 
2015-06-14 03:37:30 PM  
Auto-play embedded video.

www.drunktiki.com
 
2015-06-14 03:39:38 PM  
I wonder if you could adapt it to, say, play Homeworld? And how long it'd take it to dominate that game.
 
2015-06-14 03:51:25 PM  

RaisingKane: Pichu0102: Okay but can we run our own copy? I always wanted to watch a blank slate learn.

Sure, the source code for the plugin he used is in the video description.


I am on that shiat!
 
2015-06-14 04:15:52 PM  
It played like a computer that learned.  It took what looked like were some really unreasonable unnecessary risks and got out of them through some seriously precise inputs.  Kinda cool to watch.  It's like a very weird mash-up of a terribly stupid player and an expert-level speedrun.
 
2015-06-14 04:21:34 PM  
Teach it to play Portal. Get some hot AI on AI action going. Mmm, yes, thats some science. Naughty, naughty science.
 
2015-06-14 04:22:40 PM  
The game is MASTERBATING
 
2015-06-14 04:28:37 PM  
We are no longer particularly in the business of writing software to perform specific tasks. We now teach the software how to learn, and in the primary bonding process it molds itself around the task to be performed. The feedback loop never really ends, so a tenth year polysentience can be a priceless jewel or a psychotic wreck, but it is the primary bonding--the childhood, if you will--that has the most far-reaching repercussions.
 
2015-06-14 04:28:53 PM  
It seems like the computer memorized how to beat that level more than learned how to play Super Mario World.
 
2015-06-14 04:29:08 PM  
I have always thought neural networks were awesome. Closest I ever came was programming a basic flocking ai into a bunch of triangles back in high school.
 
2015-06-14 04:31:21 PM  
Now set it to Battle Toads.
 
2015-06-14 04:36:59 PM  

dragonchild: It played like a computer that learned.  It took what looked like were some really unreasonable unnecessary risks and got out of them through some seriously precise inputs.  Kinda cool to watch.  It's like a very weird mash-up of a terribly stupid player and an expert-level speedrun.


Yeah, there were quite a few times I was like "where's the farking hitbox on that thing?!!"

I wonder how it would do if you made the level choice random. Could it figure out a new level quicker based on the previous one?
 
2015-06-14 04:50:01 PM  

Smoking GNU: So which of those networks are for going after Sarah Conner?


img.fark.net
 
2015-06-14 05:03:45 PM  

Anenu: It seems like the computer memorized how to beat that level more than learned how to play Super Mario World.


Seems that way, but not really - if you look at the neural network, it's actually just a set of rules: if this block is safe and that block is safe and that other block is not safe and none of these three blocks are safe, then hit the A button, etc. Those rules are then applied to the entire level. It's similar to how you learn to play, knowing that if you're running and press A, Mario will jump along a certain arc to a certain height, so you can do that to jump onto a turtle at a certain position.
It's different than if you recorded a speed run with something like: at t=3, press A; at t=4, press and hold right; at t=7, press A and release right, etc. That would be memorization.
 
2015-06-14 05:06:09 PM  

Peki: I wonder how it would do if you made the level choice random. Could it figure out a new level quicker based on the previous one?


Most likely. Consider that the initial generations start out with "run right until dead". Starting with an advanced network on a new level would likely start better than that, even if it wasn't able to complete the level.
 
2015-06-14 05:11:31 PM  

Theaetetus: Peki: I wonder how it would do if you made the level choice random. Could it figure out a new level quicker based on the previous one?

Most likely. Consider that the initial generations start out with "run right until dead". Starting with an advanced network on a new level would likely start better than that, even if it wasn't able to complete the level.


Well, not even that. It started with standing there until the simulation restarted. Probably ran through some early generations of just pressing A or B and nothing else. Then it pressed right and ooh something happened!

I've been playing Alpha Centauri too much. I've got Zakharov's quote about the feedback principle running through my head now.
 
2015-06-14 05:11:33 PM  
But can it weigh in on matters of ethics in games journalism?
 
2015-06-14 05:32:42 PM  
But is it playing? It looked like a brute force pattern finding system, and matched up "press right" with the right times to press jump.
 
2015-06-14 05:39:59 PM  

wildcardjack: But is it playing? It looked like a brute force pattern finding system, and matched up "press right" with the right times to press jump.


Like most humans would, you mean.
 
2015-06-14 06:02:05 PM  

Anenu: It seems like the computer memorized how to beat that level more than learned how to play Super Mario World.


But isn't that what humans do too?

If I pick up Super Mario World or Super Metroid today, I could breeze through the game and get most of the secrets along the way. I can do that not because I "learned" how to play the game (which is really just a trivial matter of finding out what each button does), but because I memorized the pattern of each level.  And still have those patterns in memory today.
 
2015-06-14 06:04:00 PM  
Stephen Falken seen wringing his hands while grumbling, "Whats wrong with checkers or chess?"
 
2015-06-14 06:45:22 PM  
Meh, the politics tab is full of bots that can do that.
 
2015-06-14 06:45:46 PM  

Smoking GNU: So which of those networks are for going after Sarah Conner?


If I met Sarah Conner I'd off her just to kill this meme.
 
2015-06-14 06:46:25 PM  

Smoking GNU: wildcardjack: But is it playing? It looked like a brute force pattern finding system, and matched up "press right" with the right times to press jump.

Like most humans would, you mean.


Yeah, I was about to say. How do YOU learn how to play a game?

ProfessorTerguson: Stephen Falken seen wringing his hands while grumbling, "Whats wrong with checkers or chess?"


Nah. Need to get my GTNW on. :D

/actually had that game on my computer as a kid
//had a heartattack when I saw the movie
///OMG MOM I BOMBED RUSSIA! LIKE 20 TIMES!
/up for a game of Tic-Tac-Toe?
 
2015-06-14 06:49:11 PM  

Smoking GNU: So which of those networks are for going after Sarah Conner?


4 megabyte animated GIF ensues
 
2015-06-14 06:50:35 PM  

dragonchild: It played like a computer that learned.  It took what looked like were some really unreasonable unnecessary risks and got out of them through some seriously precise inputs.  Kinda cool to watch.  It's like a very weird mash-up of a terribly stupid player and an expert-level speedrun.


So the videogame equivalent of drunken boxing, then.
 
2015-06-14 06:55:52 PM  
I won't be worried until they come up with an AI that learn to beat Battletoads.
 
2015-06-14 06:56:43 PM  

Mister Peejay: dragonchild: It played like a computer that learned.  It took what looked like were some really unreasonable unnecessary risks and got out of them through some seriously precise inputs.  Kinda cool to watch.  It's like a very weird mash-up of a terribly stupid player and an expert-level speedrun.

So the videogame equivalent of drunken boxing, then.


Or drunken sex
 
2015-06-14 07:02:35 PM  

Peki: Smoking GNU: wildcardjack: But is it playing? It looked like a brute force pattern finding system, and matched up "press right" with the right times to press jump.

Like most humans would, you mean.

Yeah, I was about to say. How do YOU learn how to play a game?



I'll let you know when Grey Goo is finished downloading.
 
2015-06-14 08:04:26 PM  
What would happen the enemies appeared at random, instead of exactly at the same time and in the same way each time the NN played the game?
 
2015-06-14 08:05:51 PM  
But can it run through Crysis?
 
2015-06-14 08:21:50 PM  
Does this guy know there's p0rn on the intertube? I would assume not.
 
2015-06-14 09:26:14 PM  

unitednihilists: Does this guy know there's p0rn on the intertube? I would assume not.


Of course he does, why do you think he is creating programs to play his games for him. Leaves both hands free.
 
2015-06-14 10:57:51 PM  
Introduce it to "Asshole Mario" and see if an AI can commit suicide.
 
2015-06-15 12:12:18 AM  

dennysgod: I won't be worried until they come up with an AI that learn to beat Battletoads.


Through this method, they could likely create one that would win faster than most human could.  They'd have to rebuild the fitness score parameters, as it's more complicated than moving right.

The serious complications in Battletoads come from repeated perfections of timing.  On some of those later snake levels (which I've only completed using emulation and repeated fast saving and loading cycles) you have to make something like 100 precise jumps in perfect sequence and timing to complete.  A learning system can keep that timing extremely well.  Given it wouldn't need to rest, and wouldn't rage quit, it could probably use a method such as this to learn how to complete the game in a number of months, where skilled human players take years.  And few have every completed it.

My guess - and mind you, I'm not an expert on artificial neuroscience or even CS - is that a system like this would solve Battletoads faster than 99% of competent video game players could if learning the game from complete scratch.  Most people would never complete the game, even with significant dedication (they just don't have the skill or patience).  Some might, but they'd take years.  I imagine this could learn faster.

Again, just a guess.
 
2015-06-15 12:27:46 AM  
These type of NN have a tough time with non instant reward based games, ones with no score, etc. That's not saying they won't be there soon.
 
2015-06-15 01:42:21 AM  
When the AI can beat Contra with one man then I'll be impressed.
 
2015-06-15 02:30:51 AM  
Big deal, nearly everyone I know has created a two or three of these.
 
2015-06-15 07:02:50 AM  
THAT was kinda cool. Someone I know who had a hand in building Watson, while he was working on his PHD, programmed an AI to beat "Tempest". It was really weird to watch. Sometimes it would just stop, because it had learned when to spot a "no win" situation and just accept a death, since it already had so many extra lives.
 
2015-06-15 12:49:00 PM  
https://blamethelag.files.wordpress.com/2010/03/215551275_9naxe-l-2.jp g
 
2015-06-15 01:08:46 PM  

Wessoman: When the AI can beat Contra with one man then I'll be impressed.


After enough iterations, it wouldn't be a problem at all.

It probably could beat the World Record speed run of just over 9 minutes.
 
2015-06-15 01:11:10 PM  
O RLY?
4.bp.blogspot.com
 
2015-06-15 04:35:34 PM  
I'd like to see one play in a MMORPG against the rest of the world players.  (the rest of us have lives..)
 
Displayed 50 of 50 comments

View Voting Results: Smartest and Funniest

This thread is archived, and closed to new comments.

Continue Farking
Submit a Link »
On Twitter





Top Commented
Javascript is required to view headlines in widget.

In Other Media
  1. Links are submitted by members of the Fark community.

  2. When community members submit a link, they also write a custom headline for the story.

  3. Other Farkers comment on the links. This is the number of comments. Click here to read them.

  4. Click here to submit a link.

Report