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.

(Geek.com)   PhD creates AI to play Super Mario Bros. So how'd it do? Exploits glitches like a boss   (geek.com) divider line 93
    More: Interesting, Ph.D.  
•       •       •

10165 clicks; posted to Geek » on 14 Apr 2013 at 5:51 PM (1 year ago)   |  Favorite    |   share:  Share on Twitter share via Email Share on Facebook   more»



93 Comments   (+0 »)
   
View Voting Results: Smartest and Funniest

Archived thread

First | « | 1 | 2 | » | Last | Show all
 
2013-04-14 02:25:54 PM  
This sounded really interesting until the guy started talking. I couldn't stand it.
 
2013-04-14 03:13:57 PM  

marius2: This sounded really interesting until the guy started talking. I couldn't stand it.


This. Skip to 6:00 or so and it starts to get really good. Very impressive stuff.
 
2013-04-14 03:57:09 PM  

Msol: marius2: This sounded really interesting until the guy started talking. I couldn't stand it.

This. Skip to 6:00 or so and it starts to get really good. Very impressive stuff.


THESE. Wish I'd read the first two comments before sitting through that.

That said, I'll repeat the "skip to the 6 minute mark" thing cos it is worth a watch.
 
2013-04-14 03:59:17 PM  
Now if only they can do this with Dark Souls.
 
2013-04-14 04:05:08 PM  
Let's see the computer beat Battletoads now
 
2013-04-14 04:56:25 PM  
That was pretty cool. I wonder how well it would play Top Gun?
 
2013-04-14 05:21:39 PM  
Socially awkward hipstergenius meme - nailed.

Other that than, interesting stuff. I'd love to see it play TF2.
 
2013-04-14 06:05:23 PM  

Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now


2 player.
 
2013-04-14 06:07:49 PM  
In terms of side-scrollers I'd like to see that AI take a crack at Ghosts and Goblins and/or Bionic Commando.
 
2013-04-14 06:10:19 PM  
Hearing that music from Tetris gave me some serious flashbacks.
 
2013-04-14 06:10:50 PM  

Msol: marius2: This sounded really interesting until the guy started talking. I couldn't stand it.

This. Skip to 6:00 or so and it starts to get really good. Very impressive stuff.


Oh, did he finally get to the goddamn point?

Jesus, I bet his 22 page paper could be boiled down to three sentences and a picture.
 
2013-04-14 06:12:25 PM  

quatchi: In terms of side-scrollers I'd like to see that AI take a crack at Ghosts and Goblins and/or Bionic Commando.


It's probably not especially bad at those types of games, where reaction time is at a premium.  If you read the paper, the AI is fantastic at finding the best "right now" move.  If it figures out the the best move is to jump over or shoot a monster, it won't miss the timing.

It's a complete disaster at games like Tetris where the best "right now" move is "put a piece down quickly, which gives me a few points", as it only tracks success by score, and doesn't "know" the rules of the game other than that.
 
2013-04-14 06:15:43 PM  
Damn, that's really impressive.
 
2013-04-14 06:20:54 PM  
ok, the intro to the video was boring, but the guy's narration of it playing the game is awesome.  "The only winning move is not to play"
 
2013-04-14 06:37:19 PM  
Let me see if I can get to the point a little faster:

* First, play the game for a bit and let the computer 'watch' the results, so that you can teach the computer a simple definition of what it means to be 'winning' at the game -- like "If my score goes up, that's good; if I'm moving to the right, that's somewhat less good but still good; if I die, that's really bad."

* Use that to create a very simple function that can look at any two game states and say "Which of these two states is better?"

* Now, hook your program up to an emulator, so it can push the buttons on the controller, save the game at a given position, and restore it to that position. Don't teach it to actually watch the screen; just teach it to look at memory and use the "How am I doing?"" and "Which of these two states is better?" function.

* Now, you basically do what chess programs do. To decide what to do, a chess program basically looks at all the plausible moves it can make now, and what will likely happen in the next few moves after each of those possibilities, and says "Which of these moves will lead to the most 'winning' outcome?" This videogame-playing program does the same thing; using save states, it tries out various combinations of button presses from a given position, and says "Which of these button-press combinations looks like it will lead to the most 'winning' outcome?"

* Then it makes that move, waits a fraction of a second, and does it all again -- "Okay, now which button-press combinations will lead to the most 'winning' algorithm?"

* Repeat until you win or screw up.

The fact that it's only looking about a second ahead shows why it's good at Super Mario Bros and bad at Pac-Man. In Super Mario Bros, there's always something you can do in the short term to make your position better; even if there's nothing else interesting to do, just find a way to move to the right. But in Pac-Man, if you can only look a second ahead, and you're just trying to improve your score, and you're more than a second away from any dots, then you're screwed -- you don't see anything that will improve your position other than "don't die", so you dodge ghosts but otherwise just move around randomly until you happen to stumble upon some dots.

And as he said, the interesting thing about this approach is that it's an incredibly simple form of "AI", yet still works well on some games. With a slightly more complicated sort of "Am I winning?" function -- one that could look at more than just 'Which bytes in memory are going up and down?' -- this could work on a lot more games. For example, if you just added 'Being closer to dots and fruit is good' to the Pac-Man "Am I winning?" function, I have a feeling it could probably beat the game.
 
2013-04-14 06:43:58 PM  

KickahaOta: I have a feeling it could probably beat the game.


Can you beat Pac-Man? I always sucked at that game.
 
2013-04-14 06:53:09 PM  

Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.


I'm having a flashback...so much blood...
 
2013-04-14 07:00:05 PM  
and his thesis is applicable in the real world... how?

/ i thought it was tres neato
 
2013-04-14 07:03:15 PM  
This research will not lead to very special missile defense systems or deep space satellites. Fun, but the product is more likely the result of OCD untreated than the motivation for immediately practical critical applications.
 
2013-04-14 07:05:29 PM  

thisiszombocom: and his thesis is applicable in the real world... how?

/ i thought it was tres neato


Someone in the next realm up tried this with our world, and the result is Kim Kardashian.

Only way to explain her "success."
 
2013-04-14 07:08:01 PM  
Let's see the computer play Kaizo Mario.
 
2013-04-14 07:11:47 PM  

xkillyourfacex: This research will not lead to very special missile defense systems or deep space satellites.


I think it's unfair to judge it against those kinds of problems.  This type of problem solving is about automating common tasks moreso than perfecting extraordinary ones.  It's a line of research that could lead to more efficient factories, faster algorithms for common data processing tasks, that sort of thing.  Its strength is that it finds exploitable bugs that human intuition wouldn't consider.

That said, this doesn't seem like thesis-level stuff to me.  He took somebody else's algorithms and applied them to a new field with a few tweaks.  He didn't really contribute to any reasonable scientific advancement.
 
2013-04-14 07:12:07 PM  

thisiszombocom: and his thesis is applicable in the real world... how?

/ i thought it was tres neato


Research doesn't have to have immediate practical results. Most of the time, the only thing it's used for is as a base for some other research project, which may have an immediate practical result, or just add to the sum of human knowledge again to be used as another base for more research.

I'm sure it'll result in a robot apocalypse somewhere down the line, though.
 
2013-04-14 07:14:28 PM  

zero7717: I'm sure it'll result in a robot apocalypse somewhere down the line, though.


Probably because of this:

KickahaOta: * Use that to create a very simple function that can look at any two game states and say "Which of these two states is better?"


I'll be in my bunk...er.
 
2013-04-14 07:16:14 PM  

xkillyourfacex: This research will not lead to very special missile defense systems


Even if they teach it to play this?

theinterrobang.com
 
2013-04-14 07:18:18 PM  

fusillade762: xkillyourfacex: This research will not lead to very special missile defense systems

Even if they teach it to play this?

[theinterrobang.com image 480x320]


Or this:

encrypted-tbn2.gstatic.com
 
2013-04-14 07:18:43 PM  

Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.


That stupid game. Anyone who claims to have beat it is lying.
even with the game genie it couldnt be done. Think it was the last level where you had to go through a tunnel of spikes.
every time you spawn you would die.
it was a broken game
 
2013-04-14 07:20:20 PM  

teenage mutant ninja rapist: Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.

That stupid game. Anyone who claims to have beat it is lying.
even with the game genie it couldnt be done. Think it was the last level where you had to go through a tunnel of spikes.
every time you spawn you would die.
it was a broken game


I did it single player with save states, but that's it.  I did, however, beat the Game Boy version, which was also incredibly difficult.
 
2013-04-14 07:20:37 PM  

antidisestablishmentarianism: KickahaOta: I have a feeling it could probably beat the game.

Can you beat Pac-Man? I always sucked at that game.


You can, yes. If you beat level 255, the level counter rolls over and you get this screen. It's not "beating the game" in the sense that there's an ending, but the game can't be played past that point.

upload.wikimedia.org
 
2013-04-14 07:23:07 PM  

Sword and Shield: antidisestablishmentarianism: KickahaOta: I have a feeling it could probably beat the game.

Can you beat Pac-Man? I always sucked at that game.

You can, yes. If you beat level 255, the level counter rolls over and you get this screen. It's not "beating the game" in the sense that there's an ending, but the game can't be played past that point.

[upload.wikimedia.org image 224x288]


Kill screen.  Back when memory space was incredibly limited.  I wonder if any of the modern ports have that...
 
2013-04-14 07:24:58 PM  

teenage mutant ninja rapist: Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.

That stupid game. Anyone who claims to have beat it is lying.
even with the game genie it couldnt be done. Think it was the last level where you had to go through a tunnel of spikes.
every time you spawn you would die.
it was a broken game


I never played the original game but I do remember the cartoon.
 
2013-04-14 07:26:14 PM  

serial_crusher: That said, this doesn't seem like thesis-level stuff to me.


Not sure where you got the idea that a guy who already has a PhD is writing a thesis.
 
2013-04-14 07:34:22 PM  

serial_crusher: That said, this doesn't seem like thesis-level stuff to me.


You can write a thesis on anything.
 
2013-04-14 07:37:49 PM  

quatchi: In terms of side-scrollers I'd like to see that AI take a crack at Ghosts and Goblins and/or Bionic Commando.


Bionic Commando?  Umm... really love the game, but it's easy.  Super easy.

/Almost beat Super Gouls 'n Ghosts once
//Damn game locked up just as I got to the last guy the second time
/// :(
 
2013-04-14 07:44:34 PM  

Telos: Bionic Commando? Umm... really love the game, but it's easy. Super easy.


I wouldn't say super easy, but not bad once you got full control of the physics and got enough health to take a few hits.

The XBLA remake is awesome.
 
2013-04-14 07:45:07 PM  

teenage mutant ninja rapist: Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.

That stupid game. Anyone who claims to have beat it is lying.
even with the game genie it couldnt be done. Think it was the last level where you had to go through a tunnel of spikes.
every time you spawn you would die.
it was a broken game


Presumably you have not even played I Wanna Be The Guy.
 
2013-04-14 07:46:28 PM  

LazarusLong42: teenage mutant ninja rapist: Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.

That stupid game. Anyone who claims to have beat it is lying.
even with the game genie it couldnt be done. Think it was the last level where you had to go through a tunnel of spikes.
every time you spawn you would die.
it was a broken game

Presumably you have not even played I Wanna Be The Guy.


I need to play this game soon.
 
2013-04-14 07:48:16 PM  

Mentat: serial_crusher: That said, this doesn't seem like thesis-level stuff to me.

You can write a thesis on anything.


You can major in Game Boy if you know how to bullshiat.

/Shouldn't be obscure
 
2013-04-14 08:05:24 PM  
the tetris one was neat.  the computer knew it was going to lose so it paused the game as its last move.  kinda like the kid who flips the game board pissed that he can't win
 
2013-04-14 08:12:44 PM  

MrEricSir: serial_crusher: That said, this doesn't seem like thesis-level stuff to me.

Not sure where you got the idea that a guy who already has a PhD is writing a thesis.


You're right, I misinterpreted.  I'd like to change my answer to "either the university is wasting money or this guy has some awesome hobbies"
 
2013-04-14 08:15:57 PM  

fusillade762: xkillyourfacex: This research will not lead to very special missile defense systems

Even if they teach it to play this?

[theinterrobang.com image 480x320]


That would only be effective against North Korea's long range missile guidance systems.
 
2013-04-14 08:30:09 PM  

eyeq360: Let's see the computer play Kaizo Mario.


I'm watching the YouTube of that... And I'm wondering how the hell I wound up watching that. Forgot where I started.

And I bet the computer would do fine, eventually, It's got patience.
 
2013-04-14 08:55:47 PM  

Flappyhead: Precious Roy's Horse Dividers: Let's see the computer beat Battletoads now

2 player.


That is, quite literally, impossible.  At least without player 2 stealing lives from player 1 after a game over.
 
2013-04-14 09:03:45 PM  
I've never seen anyone put their degree on a computer science paper, or call themselves doctor, and he did both. I wonder what backwater he graduated from?

<Google search>

Carnegie Mellon, hmm....
 
2013-04-14 09:08:55 PM  

ltdanman44: the tetris one was neat.  the computer knew it was going to lose so it paused the game as its last move.  kinda like the kid who flips the game board pissed that he can't win


Yep. A learning algorithm met its existential end.
 
2013-04-14 09:13:07 PM  

marius2: This sounded really interesting until the guy started talking. I couldn't stand it.


Yeah, he reminded me why my school made us do biannual talks on our work. I hated doing it at first but then I got really good at it now I can present my work like a pro.  It's clear he didn't get any of that training.  You can have the coolest work in the world but if you can't effectively communicate it then no one cares.
 
2013-04-14 09:19:26 PM  

RedPhoenix122: Sword and Shield: antidisestablishmentarianism: KickahaOta: I have a feeling it could probably beat the game.

Can you beat Pac-Man? I always sucked at that game.

You can, yes. If you beat level 255, the level counter rolls over and you get this screen. It's not "beating the game" in the sense that there's an ending, but the game can't be played past that point.

[upload.wikimedia.org image 224x288]

Kill screen.  Back when memory space was incredibly limited.  I wonder if any of the modern ports have that...


The Java-based port used for Google's splash image on Pac-Man's anniversary included the kill-screen, coded in intentionally.
 
2013-04-14 09:34:36 PM  

wildcardjack: eyeq360: Let's see the computer play Kaizo Mario.

I'm watching the YouTube of that... And I'm wondering how the hell I wound up watching that. Forgot where I started.

And I bet the computer would do fine, eventually, It's got patience.


I never heard of Kaizo Mario until just now.
 
2013-04-14 09:50:22 PM  
That was pretty interesting.
I love how it dealt with tetris, even tho it was epic failure.

/Stacks pieces to make the wrong kinds of points.
/Pauses a split second before the end.
/The only winning move is not to play.
 
2013-04-14 09:56:24 PM  

KickahaOta: The fact that it's only looking about a second ahead shows why it's good at Super Mario Bros and bad at Pac-Man


Course making an AI that can beat Pac-Man would be trivial since there's a known pattern to go through to get through each level.
 
Displayed 50 of 93 comments

First | « | 1 | 2 | » | Last | Show all

View Voting Results: Smartest and Funniest


This thread is archived, and closed to new comments.

Continue Farking
Submit a Link »
On Twitter





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