|
On November 13 2012 01:33 marvellosity wrote: So you build in evaluations, like I said with king safety, pawn structure, piece activity and the like, and now computers can appreciate that maybe taking this pawn is not a good idea after all.
If computers only used computational power, top human GMs would easily crush engines time after time.
Basically computers simply cannot tell what is good or bad without being told what is good or bad by authors in the first place. And it's the relative value authors place on various things that distinguish top-end engines.
Edit: but what you say is true, computers do do these many-short-games thing. It's called Monte Carlo analysis.
All the actual evaluation weights are decided by the fine-tuning process involving millions of short games. We're not talking about Monte Carlo analysis, these are just plain engine games - but very many of them.
Before you continue to say that I miss the point, I am actually the author of a very strong chess engine. So while I do decide which evaluation factors go into the engine, it's ultimately the countless ultra-rapid engine games that fine-tune the weights of all the evaluation parameters.
A similar approach can probably be adopted for Starcraft, it would be a very interesting challenge to work out an AI for the game.
|
United Kingdom35817 Posts
On November 13 2012 01:58 RHoudini wrote:Show nested quote +On November 13 2012 01:33 marvellosity wrote: So you build in evaluations, like I said with king safety, pawn structure, piece activity and the like, and now computers can appreciate that maybe taking this pawn is not a good idea after all.
If computers only used computational power, top human GMs would easily crush engines time after time.
Basically computers simply cannot tell what is good or bad without being told what is good or bad by authors in the first place. And it's the relative value authors place on various things that distinguish top-end engines.
Edit: but what you say is true, computers do do these many-short-games thing. It's called Monte Carlo analysis.
All the actual evaluation weights are decided by the fine-tuning process involving millions of short games. We're not talking about Monte Carlo analysis, these are just plain engine games - but very many of them. Before you continue to say that I miss the point, I am actually the author of a very strong chess engine. So while I do decide which evaluation factors go into the engine, it's ultimately the countless ultra-rapid engine games that fine-tune the weights of all the evaluation parameters. A similar approach can probably be adopted for Starcraft, it would be a very interesting challenge to work out an AI for the game.
eh.
Not Houdini, that would be too odd.
And I would have to talk to you at once :d
|
Houdini it is .
|
On November 13 2012 02:14 RHoudini wrote:Houdini it is . So boss.
|
Northern Ireland20797 Posts
I think one of the difficulties given as well is that even the fine strategic minds on Team Liquid frequently disagree with each other's interpretations of where games are won and lost. Thus there becomes additional difficulties in pre-programming responses to different situations, given the lack of consensus on what said responses should be and whatnot.
|
United Kingdom35817 Posts
On November 13 2012 02:14 RHoudini wrote:Houdini it is .
haha amazing.
my patzer chess knowledge bows down to yours
|
On November 13 2012 01:35 Poopi wrote:Show nested quote +On November 12 2012 22:19 See.Blue wrote: You can. What you're going to be using instead is going to be machine learning rather than some sort of pre-scripted set of responses to an information set like people seem to be assuming. It's absolutely possible to program an AI that can beat any top pro. Super smash bros melee use such a system but AI are still easily beatable, despite them learning from your plays. Computers won't be able to beat pros without cheating. ) so naive. Are you assuming that the creators of Super smash bros melee AI put as much effort in it as, for example, the creators of Deep Blue?
|
On November 13 2012 02:36 Cheerio wrote:Show nested quote +On November 13 2012 01:35 Poopi wrote:On November 12 2012 22:19 See.Blue wrote: You can. What you're going to be using instead is going to be machine learning rather than some sort of pre-scripted set of responses to an information set like people seem to be assuming. It's absolutely possible to program an AI that can beat any top pro. Super smash bros melee use such a system but AI are still easily beatable, despite them learning from your plays. Computers won't be able to beat pros without cheating. ) so naive. Are you assuming that the creators of Super smash bros melee AI put as much effort in it as, for example, the creators of Deep Blue? Duh of course
+ Show Spoiler +No I'm not that retarded hahaha
|
On November 13 2012 02:36 Cheerio wrote:Show nested quote +On November 13 2012 01:35 Poopi wrote:On November 12 2012 22:19 See.Blue wrote: You can. What you're going to be using instead is going to be machine learning rather than some sort of pre-scripted set of responses to an information set like people seem to be assuming. It's absolutely possible to program an AI that can beat any top pro. Super smash bros melee use such a system but AI are still easily beatable, despite them learning from your plays. Computers won't be able to beat pros without cheating. ) so naive. Are you assuming that the creators of Super smash bros melee AI put as much effort in it as, for example, the creators of Deep Blue?
In addition to what he's saying, I doubt you have played enough games for it to truly learn and improve anything.
PS. Houdini you rock!
|
On November 13 2012 02:48 JKM wrote:Show nested quote +On November 13 2012 02:36 Cheerio wrote:On November 13 2012 01:35 Poopi wrote:On November 12 2012 22:19 See.Blue wrote: You can. What you're going to be using instead is going to be machine learning rather than some sort of pre-scripted set of responses to an information set like people seem to be assuming. It's absolutely possible to program an AI that can beat any top pro. Super smash bros melee use such a system but AI are still easily beatable, despite them learning from your plays. Computers won't be able to beat pros without cheating. ) so naive. Are you assuming that the creators of Super smash bros melee AI put as much effort in it as, for example, the creators of Deep Blue? In addition to what he's saying, I doubt you have played enough games for it to truly learn and improve anything. PS. Houdini you rock! I doubt the Blizz/Smash AI even is an AI, as opposed to a script lol.
|
On November 11 2012 13:00 EtherealDeath wrote: If there were a major research project for a Starcraft Deep Blue as there was in Chess, we would have a Deep Blue for Starcraft probably within 15-20 years max.
That's a truely frightening idea. Man, that'd be sick actually.
|
Didn't computers kill chess? Or was it the popularity of poker?
There are two problems here, anyway: 1. decision making 2. cheating by abusing the micro abilities. These two have nothing to do with each other, it's kinda silly there is this constant debate of "the computer would be bad at decision making, but it might compensate with good micro", since it's two different problems and they have no relation - outside of some ill-conceived idea of using an AI to cheat in online tourneys, but even then it's mostly problem #1 as #2 is too obvious.
|
Computers killed matches of humans against computers - it's become imba. Current chess engines on an ordinary laptop are stronger than Deep Blue in 1997.
There's no reason to believe that it would be impossible to create a Starcraft 2 AI that would win the GSL - even without abusing the micro abilities.
|
an ai like that would have to be programmed like a connected network, similar but a thousand times more advanced than navy mine sweepers. You could not do it with modern computerlayout (ie one CPU), you would need a few hundred at least.
but from there, training it would only be time consuming, not difficult.
|
Btw, if anyone's wondering:
since the release of version 1.5 on 15 December 2010 Houdini has taken the top spot in every rating list that includes it.[2][3][4][5] Because of this, Houdini is currently used by chess world champion Viswanathan Anand[6] and when GM Peter Svidler was asked which one player he would choose to represent Earth in a hypothetical match against aliens, he answered "Houdini". We do have royalty here. :p How cool that you are interested in Starcraft.
|
i'd love to see some giant think-tank university sponsor an SC2 AI tournament. i know this has happened for SC1.
|
most people here are probably unqualified to answer the question of how difficult making an elite sc2 ai would be. i'm a comp sci major; as i'm sure some of the more thorough answerers in the thread are. imho, it'd be quite possible to make something low-mid masters level; the primary flaw in most game ais is simply lack of work, that is the company didn't put in that much effort to make it good; so there's a lot of room for improvement using some basic techniques. (near) perfect micro does go a long way to helping an ai out, as well as abusing the ai's micro advantage: multi-prong drops and battles are way easier for a computer than a human. and many things in the game are balanced based on human level play; micro-intensive things like creep tumors and single target spells in large battles would be rather strong with a computer's apm. building a basic learning ai that copies build orders isn't that hard; and since it can learn build orders by using existing top level games, it can copy piles of build orders to use.
while building lots of intelligence is hard; making an ai that scouts decently and is moderately adaptive in its build orders is quite feasible; and having build order variety isn't hard by copying piles of pro games for builds.
i concur that the best path for an to win is a perfectly micro-ed all-in; though a macro game with lots of harassment would work decently due to ai's ability to multitask.
|
the biggest problem is for the AI to defend all ins and ajust its positioning with army. It is easy to program the AI to play but hard to program the AI to win.
|
I'm fairly sure that nobody could beat a 6 Rax with PERFECT micro, as in every single Marine stuttersteps and focuses stuff individually.
|
The micro advantage from an AI will definitely out task any human player, so much so that they'll alway win if it comes up with a half decent strategy.
However, SC2 is balanced around humans. The mechanical demands are balanced such that implementation of strategies always have to come from human limits. As someone mentioned earlier, early marine rushes, perfectly and inhumanly microed, might be sufficient in crushing any other strategy in the game we have now.
|
|
|
|