Jump to content
IGNORED

Microsoft's AI Just Shattered the Ms. Pac-Man High Score


Flojomojo

Recommended Posts

http://sploid.gizmodo.com/microsofts-ai-just-shattered-the-ms-pac-man-high-score-1796091352

 

The researchers chose Ms. Pac-Man as it was written to be far less predictable than the original version of the game, which will help the AI they developed better deal with real-world situations that are equally unpredictable. Because as impressive as this feat is, few companies are willing to pay for something that just sits around and plays video games all day.

Update, 2:14 p.m. EST/EDT: Microsoft has reached out to us to clarify that the high score for the Atari 2600 version of Ms. Pac-Man is 266,360 points.

 

 

  • Like 5
Link to comment
Share on other sites

Kind of cool, I guess, but I doubt the all time human high score on 4 ghost 2600 Ms. Pac-man is a mere 266K. The game is not that hard; I'm pretty sure I broke 300K on it back in high school. Rolling it would be tough with only 4 lives, but I bet someone has done it before Microsoft's AI.

 

Edit: Yup, I just checked the 2600 high score club results and the high score on Ms. Pac-man is a million...

Edited by hepcat
Link to comment
Share on other sites

Gizmodo botched the story, was fundamentally confused about the differences between arcade Pac-Man, Atari 2600 versions, and Ms. Pac-Man.

 

Here's a better story

https://www.wired.com/story/mircosoft-ai-ms-pac-man/

 

and here's the source.

https://blogs.microsoft.com/next/2017/06/14/divide-conquer-microsoft-researchers-used-ai-master-ms-pac-man/#sm.0001y11attlemfpzrfk1b5haiqep0

 

If the maximum score is a million, all they can do is hit it. Shattering won't happen.

  • Like 1
Link to comment
Share on other sites

Lol I don't know if this is the case but it'd be pretty sad if you don't conceptually already understand that arcade ms pac and our beloved VCS port are pretty much entirely different games. It's almost like they grew up in a different era or something!! :lol:

 

Regardless though both games get stingy with the scoring once you hit the 300k mark or so.. you can't get any points from ghosts, but also unlike Pac-man you don't get 5000 pt bonuses consistently. Although I don't think the vcs ever gets to a point where the ghosts don't turn blue... do they?

Edited by NE146
  • Like 2
Link to comment
Share on other sites

Lol I don't know if this is the case but it'd be pretty sad if you don't conceptually already understand that arcade ms pac and our beloved VCS port are pretty much entirely different games. It's almost like they grew up in a different era or something!! :lol:

 

Regardless though both games get stingy with the scoring once you hit the 300k mark or so.. you can't get any points from ghosts, but also unlike Pac-man you don't get 5000 pt bonuses consistently. Although I don't think the vcs ever gets to a point where the ghosts don't turn blue... do they?

 

If I remember correctly, on the 2600 version you get a one second blue time on every board 8 and higher. It never drops to zero. Also, the 2600 Ms. Pac seems less stingy with bananas on the random fruit levels.

Link to comment
Share on other sites

This is really cool, thanks for staring. Perhaps someone invents a "Johny 5" type approach where the bot interfaces the game capturing light and sound from a CRT telivision and servo actuated CX-40 joystick performs the in-game inputs, using the same feedback loops a human would. That would be awesome!

Link to comment
Share on other sites

Am going to guess before reading, and say they used a fluid diffusion model (model the maze as a three dimensional model of fluid going through trenches, with certain parts of the maze weighted. Similar to what Norbert Kehrer did when creating an algorithm to automatically place the direction arrows for his automatic mazes...am guessing that a markov model is then built from this fluid model and traversed, as a base-line)

 

-Thom

Link to comment
Share on other sites

Ah close, Definitely a parallel Markov traversal like I thought, but with a significant amount of high level domain specific knowledge in each Markov thread, with extreme care to make each thread take care of a specific aspect of a given reward... The algorithm seems to have a high vector factor, and therefore scales quite nicely.

 

The domain specific knowledge describing the reward is the key here, the rest is the grind-work to approach the reward.

 

-Thom

 

(p.s. forget about the fucking articles, they're written by morons, AND ACTUALLY READ THE PAPER!, where the actual data is...it's far more insightful)

 

https://arxiv.org/pdf/1706.04208.pdf

  • Like 2
Link to comment
Share on other sites

jr pac man would also be better test, but the thing is, since they've got a solution for ms pac man,

 

it should theoretically also work for all pac-man type games.

 

and most maze games in general.

 

later

-1

I'd like to see any human or AI crack one million on Jr Pacman 2600. The ghosts intelligence and speed are merciless on that game. Even on teddy bear mode (one ghost) I still can't unlock all the mazes... :dunce:
Link to comment
Share on other sites

I'd like to see them apply similar research to arcade Donkey Kong which IMHO would be much harder to make AI for since the goal isn't grinding through board after board but maximizing points until the kill screen.

 

I wonder if an AI can beat the 1.2 million point human record.

 

A

  • Like 1
Link to comment
Share on other sites

Join the conversation

You can post now and register later. If you have an account, sign in now to post with your account.
Note: Your post will require moderator approval before it will be visible.

Guest
Reply to this topic...

×   Pasted as rich text.   Paste as plain text instead

  Only 75 emoji are allowed.

×   Your link has been automatically embedded.   Display as a link instead

×   Your previous content has been restored.   Clear editor

×   You cannot paste images directly. Upload or insert images from URL.

Loading...
  • Recently Browsing   0 members

    • No registered users viewing this page.
×
×
  • Create New...