Its an automated crossworld puzzle solver. How it works (and my advisor led the project, though I don't work on anything remotely similar) is that it has a large number of solver modules that are each good at a certain kind of clue. One might be really good at looking up famous people based on keywords. Another might be good at... I dunno some other type of crossword clue.
Then each of these modules made lists of possible answers for each clue (subject to length and letter constraints), complete with the confidence they had in various answers.
A central "merger" then collected the candidate answers for each clue from the different modules, and then did lots of tricky search-like algorithms to find a set of answers that seemed the most cohesive.
This system, PROVERB, was at least the best computer system for solving crosswords, at one time, and did fairly well in competitions in which other humans competed too.
With Jeopardy! something similar could approach this issue, as well, except without the added constraint that questions to different answers have to relate to each other on the level of spelling.
Hide something in interstellar space, note its current GalacticPS coordinates and velocity, and come back years later to find it. Probably needs to be fairly accurate.
Easy to implement. First thing that seems plausible that gets taught in a machine learning course. And sometimes, like, say... HERE.... it works pretty well!
1) Playing on lives didn't work because eventually a network would destroy the boss' offensive capabilities and hide in a corner. The game would never progress.
Why on earth would that poor guy want to get to a new level for more hurt? He's a lover, not a fighter! Why can't we all just get along?
They don't need to find the image - they start out knowing where it is, so stuff like template matching is irrelevant.
And since the NN started out knowing where the icons were, figuring out which ones matched which signals is not as impressive. It's just a ho-hum NN application, now.
Except to do something like that you have to analyze the program code of every game you make a trainer for. I've never done that sort of thing, but it seems scary.
An approach like the one they've outlined can probably be moved from game to game with only parameter tweaks.
Template matching might have been faster. If you know exactly what image you're looking for, you can find it in time linear with the number of screen pixels.
I didn't bother reading the article, but it seems unlikely that their algorithm started with the icons it was looking for. I am not an expert on neural nets specifically (though I am a machine learning researcher so I have a passing familiarity with them) but I don't know how you would give it that knowledge.
It's possible that the NN figured out the icons on its own, which would be a *very* cool result.
The lack of parenthesis in the english language make this thing a little more difficult, but if the reporter wanted to indicate that a "violation of rules" was "showing up on time", there would have been a comma after "rules". So while it is confusingly worded and poorly written, it is technically correct.
More than that, if you suddenly try to load a large application that demands RAM, and there isn't enough due to the super pre-fetch or whatever, it's not like they're going to swap the super pre-fetch stuff to disk. They'll just drop it and load eclipse. Computer science is easy, folks!
I refused to learn latex when I was in academia. I am shocked it is still around. But the apps I saw that might have replaced it are probably either too pricey or long dead these days. I remember writing my thesis is Word and I had to reboot the PC after every major format change to free up memory. (Days when 8MB as a lot of memory.)
Or you could have just learned latex and saved yourself the hassle.
Whenever I read slashdot comments, I always think "man these people are insightful", until there is a story I actually know something about (AI). Then I think "man these people only have the barest clue about what they're talking about. this is quite pathetic".
I wonder if it's like that for experts in other topics as well?
Re:Developers need predictable behaviour
on
The State of Game AI
·
· Score: 2, Informative
Yeah... you don't want to use neural networks for game AI.
Reinforcement Learning, on the other hand, is. At its heart the RL problem is the same as the sequential decision making problem. An agent acts in the world, receiving observations and numerical reward signals that it tries to maximize.
The RL community is young (compared to the AI community as a whole) and is building up the theory and experience needed to approach these sorts of problems. All of my work focuses on agents learning to play video games (FPS and platformers, as more or less two separate threads).
It's coming, and we'll be ready to help the AI soon...just leave a couple cycles free from all those fancy graphics so we can do some thinking in the background, ok?
Slashdot Mirror
Its an automated crossworld puzzle solver. How it works (and my advisor led the project, though I don't work on anything remotely similar) is that it has a large number of solver modules that are each good at a certain kind of clue. One might be really good at looking up famous people based on keywords. Another might be good at... I dunno some other type of crossword clue.
Then each of these modules made lists of possible answers for each clue (subject to length and letter constraints), complete with the confidence they had in various answers.
A central "merger" then collected the candidate answers for each clue from the different modules, and then did lots of tricky search-like algorithms to find a set of answers that seemed the most cohesive.
This system, PROVERB, was at least the best computer system for solving crosswords, at one time, and did fairly well in competitions in which other humans competed too.
With Jeopardy! something similar could approach this issue, as well, except without the added constraint that questions to different answers have to relate to each other on the level of spelling.
It doesn't rotate around the sun. It is fixed in both time and space.
Hide something in interstellar space, note its current GalacticPS coordinates and velocity, and come back years later to find it. Probably needs to be fairly accurate.
Except gliders move at about 1 foot per second. They move through small changes in buoyancy combined with orientation changes.
Makes me think of the scene in Austin Powers, where the steamroller is running over the minions.
It's because people who know how to use latex can almost universally use word, but the converse is not remotely true.
Easy to implement. First thing that seems plausible that gets taught in a machine learning course. And sometimes, like, say... HERE.... it works pretty well!
1) Playing on lives didn't work because eventually a network would destroy the boss' offensive capabilities and hide in a corner. The game would never progress.
Why on earth would that poor guy want to get to a new level for more hurt? He's a lover, not a fighter! Why can't we all just get along?
Ok, I skimmed the article.
They don't need to find the image - they start out knowing where it is, so stuff like template matching is irrelevant.
And since the NN started out knowing where the icons were, figuring out which ones matched which signals is not as impressive. It's just a ho-hum NN application, now.
Except to do something like that you have to analyze the program code of every game you make a trainer for. I've never done that sort of thing, but it seems scary.
An approach like the one they've outlined can probably be moved from game to game with only parameter tweaks.
That is the true beauty of machine learning :)
Template matching might have been faster. If you know exactly what image you're looking for, you can find it in time linear with the number of screen pixels. I didn't bother reading the article, but it seems unlikely that their algorithm started with the icons it was looking for. I am not an expert on neural nets specifically (though I am a machine learning researcher so I have a passing familiarity with them) but I don't know how you would give it that knowledge. It's possible that the NN figured out the icons on its own, which would be a *very* cool result.
The lack of parenthesis in the english language make this thing a little more difficult, but if the reporter wanted to indicate that a "violation of rules" was "showing up on time", there would have been a comma after "rules". So while it is confusingly worded and poorly written, it is technically correct.
More than that, if you suddenly try to load a large application that demands RAM, and there isn't enough due to the super pre-fetch or whatever, it's not like they're going to swap the super pre-fetch stuff to disk. They'll just drop it and load eclipse. Computer science is easy, folks!
CS conference submissions I've been involved in (the UAI deadline is tonight!) require latex. So it's dominant at least in the AI community.
I refused to learn latex when I was in academia. I am shocked it is still around. But the apps I saw that might have replaced it are probably either too pricey or long dead these days. I remember writing my thesis is Word and I had to reboot the PC after every major format change to free up memory. (Days when 8MB as a lot of memory.)
Or you could have just learned latex and saved yourself the hassle.
Measuring a particle's usefulness seems like a waste of a particle's usefulness.
Whenever I read slashdot comments, I always think "man these people are insightful", until there is a story I actually know something about (AI). Then I think "man these people only have the barest clue about what they're talking about. this is quite pathetic". I wonder if it's like that for experts in other topics as well?
Yeah... you don't want to use neural networks for game AI. Reinforcement Learning, on the other hand, is. At its heart the RL problem is the same as the sequential decision making problem. An agent acts in the world, receiving observations and numerical reward signals that it tries to maximize. The RL community is young (compared to the AI community as a whole) and is building up the theory and experience needed to approach these sorts of problems. All of my work focuses on agents learning to play video games (FPS and platformers, as more or less two separate threads). It's coming, and we'll be ready to help the AI soon...just leave a couple cycles free from all those fancy graphics so we can do some thinking in the background, ok?