On January 20 2011 05:53 kamikami wrote: Well MamiyaOtaru you can defend the article as you like but I (and the guys arguing with you before) will just hate it because it gives false impression that SC skill level is low and that's it, you cannot change what people feel while reading it.
Why ? Because the title and the first paragraph give the impression that the human opponent is very high level (WCG competitor, 1st in Spain...), it just doesn't state specifically that it was 10 years ago and nowadays he is no one.
So the whole idea is invalidated because someone might not read the whole article?
The entire "argument" in the thread should've gone like this:
"I don't think the thread title is accurate." "yep it isn't." "okay."
And yes they did say that he's no longer competitive at the game.
No, nobody said the whole idea was invalidated. Try reading more thoroughly. First impressions mean everything. Getting things wrong and giving a false pretense are the worst things you can do to an audience. We're pointing out that the writer did just that.
Yes. Right from the very start you've been saying the intro is bad. Not that the article is crap. hahahaha
On January 20 2011 05:36 Marradron wrote: Not all too impressed yet. Did you know that mutas can actually fire backwards if you give the commands at accactly the right frames ? There was briefly a TL BWAPI team I was in and we had a private video of mutas owning infinite scourges just running away and firing backwards while moving. It might still be on youtube but it's most likely still private. All this AI do is mimic smartness. It's still just a simple algorithm changing certain key timings and targeting. I do have to say that it does so very well.
I added the links to some of the videos I made with dropship / tank or goli micro since this is a topic about AI
That muta micro vs the huge pile of templar was at once both terrifying and somewhat expected if the programmers had accomplished their goal of making the AI understand threats properly. Really interesting to read about their development process - and about how the AI embraces the concepts of strong play.
On January 20 2011 05:53 kamikami wrote: Well MamiyaOtaru you can defend the article as you like but I (and the guys arguing with you before) will just hate it because it gives false impression that SC skill level is low and that's it, you cannot change what people feel while reading it.
Why ? Because the title and the first paragraph give the impression that the human opponent is very high level (WCG competitor, 1st in Spain...), it just doesn't state specifically that it was 10 years ago and nowadays he is no one.
So the whole idea is invalidated because someone might not read the whole article?
The entire "argument" in the thread should've gone like this:
"I don't think the thread title is accurate." "yep it isn't." "okay."
And yes they did say that he's no longer competitive at the game.
No, nobody said the whole idea was invalidated. Try reading more thoroughly. First impressions mean everything. Getting things wrong and giving a false pretense are the worst things you can do to an audience. We're pointing out that the writer did just that.
Of course we can go through the article and quote a bunch of stuff too and claim that its the only way to interpret it, or you can accept the fact that the intros carry an air about them that isn't reflected in the rest of the article. Your choice.
So you want me to read more thoroughly into your idea that people don't need to read?
On January 20 2011 02:23 Torte de Lini wrote: No surprise here. A computer does not suffer from strains of the fingers and perfect microing is no problem for it either.
Wasn't there a documentary about this supercomputer and chess that played brilliantly against grandmasters?
SCBW is totally different from Chess, cuz it's in real-time and you have to anticipate MUCH MUCH more possible actions of your opponent.
I guess it's possible for a AI to beat a good player in one out of many matches, but for a PC to constantly beat a top-player just by perfect macro/multitasking is IMO not possible in the near future, just because most top-players already have incredibly strong macro, but they can anticipate, act and not only react etc.
My question is: Did the AI also have to scout through the fog of war, or was it able to "see" stuff without scouting?
Honestly, I don't think it will be very long before an AI is able to beat top BW pros. If the research continues at the rate its going it won't be long. In fact, I'd argue that if they can figure out a way for the AI to survive until it has a reasonable amount of Mutas then it will win every time.
Mass mutas with that AI controlling is basically unstoppable. There's literally no way to outplay micro like that.
On January 20 2011 02:23 Torte de Lini wrote: No surprise here. A computer does not suffer from strains of the fingers and perfect microing is no problem for it either.
Wasn't there a documentary about this supercomputer and chess that played brilliantly against grandmasters?
SCBW is totally different from Chess, cuz it's in real-time and you have to anticipate MUCH MUCH more possible actions of your opponent.
I guess it's possible for a AI to beat a good player in one out of many matches, but for a PC to constantly beat a top-player just by perfect macro/multitasking is IMO not possible in the near future, just because most top-players already have incredibly strong macro, but they can anticipate, act and not only react etc.
My question is: Did the AI also have to scout through the fog of war, or was it able to "see" stuff without scouting?
Read the article. It talks about the development of scouting and how it uses that information.
On January 20 2011 02:23 Torte de Lini wrote: No surprise here. A computer does not suffer from strains of the fingers and perfect microing is no problem for it either.
Wasn't there a documentary about this supercomputer and chess that played brilliantly against grandmasters?
SCBW is totally different from Chess, cuz it's in real-time and you have to anticipate MUCH MUCH more possible actions of your opponent.
I guess it's possible for a AI to beat a good player in one out of many matches, but for a PC to constantly beat a top-player just by perfect macro/multitasking is IMO not possible in the near future, just because most top-players already have incredibly strong macro, but they can anticipate, act and not only react etc.
My question is: Did the AI also have to scout through the fog of war, or was it able to "see" stuff without scouting?
I believe from their discussion near the bottom of page 3 through page 4 of the article that the AI had to scout through the fog of war, along with their entry into the 4th tournament (the full game tournament) implying that, like a regular game, fog of war would be turned on.
On January 20 2011 02:23 Torte de Lini wrote: No surprise here. A computer does not suffer from strains of the fingers and perfect microing is no problem for it either.
Wasn't there a documentary about this supercomputer and chess that played brilliantly against grandmasters?
SCBW is totally different from Chess, cuz it's in real-time and you have to anticipate MUCH MUCH more possible actions of your opponent.
I guess it's possible for a AI to beat a good player in one out of many matches, but for a PC to constantly beat a top-player just by perfect macro/multitasking is IMO not possible in the near future, just because most top-players already have incredibly strong macro, but they can anticipate, act and not only react etc.
My question is: Did the AI also have to scout through the fog of war, or was it able to "see" stuff without scouting?
It could scout and could only react to information that it has scouted. There were categories in the competition for AIs that could only do micro and AIs that didn't scout, but all the super-advanced ones could scout.
The computer that beat Kasparov supposedly cheated, which was covered in another documentary (I forgot the title). The human won easily the first day by exploiting ambiguous positions that the AI couldn't navigate, but the second day the AI changed dramatically (apparently even making mistakes) and won. After that IBM immediately disassembled the computer so nobody could confirm that it played the game un-assisted.
Who cares if one article was stupid and had an awful headline/intro? If you're discouraged or sent away by that, you're not really interested in the AI or the work which is vastly more interesting than simply beating a progamer with ai.
The content is incredible and a little scary, not because of current code/ability, but the potential. Like the arstechnica article headlines - Skynet.
Quite a change from the perfectly micro'ed dragoons and mutas, this overmind bot, I'm impressed! They should provide more videos, with explanations of what is going on, what the code is doing and not doing, it's very interesting what's up so far, but it's not enough.
Hopefully these guys get hired by Blizzard or some other rts studio, this is pure gold compared to BW AI.
On January 20 2011 05:53 kamikami wrote: Well MamiyaOtaru you can defend the article as you like but I (and the guys arguing with you before) will just hate it because it gives false impression that SC skill level is low and that's it, you cannot change what people feel while reading it.
Why ? Because the title and the first paragraph give the impression that the human opponent is very high level (WCG competitor, 1st in Spain...), it just doesn't state specifically that it was 10 years ago and nowadays he is no one.
So the whole idea is invalidated because someone might not read the whole article?
The entire "argument" in the thread should've gone like this:
"I don't think the thread title is accurate." "yep it isn't." "okay."
And yes they did say that he's no longer competitive at the game.
No, nobody said the whole idea was invalidated. Try reading more thoroughly. First impressions mean everything. Getting things wrong and giving a false pretense are the worst things you can do to an audience. We're pointing out that the writer did just that.
Yes. Right from the very start you've been saying the intro is bad. Not that the article is crap. hahahaha
Once again, I would differentiate between the pcgamer article (crap) and the arstechnica (mindblowing, well done)
On January 20 2011 05:11 Beside_kr wrote: WCG has no record of a Oriol Vinyals competing. The only thing close is Oriol Prats Navarro a UT99 player from Spain. The website doesn't have players from the WCGC (2000) but also does not list Spain as one of the countries that competed that year.
Did he mean that he'd played in a preliminary of the Spain qualifiers for Spain? Because that's something significantly different.
=DoGo= participated in the WCG 2001 finals for Spain. Given the amount of time since his peak training, the thread title is definitely an overstatement.
I apologize, WCG and TLPD have DoGo under the name Antonio Crespo Gomez
One problem they will have to look out for is overflowing commands if the game gets to a certain supply amount. As evidenced by map techniques recently, if you flood the game with 10k APM functions of the game stop working correctly or at all. So they will have to streamline it more as well.
Does anyone know if any academic papers in regards to the algorithms used in this have been released yet? I took a scan through the papers released in 2010 by Klein, but couldn't find any. The closest one I could find would be the Hierarchical Bayesian Approach, but that one involved only the PI and only as a 3rd author, and didn't make any mention of possible applications to what was mentioned in the arstechnica article.
and how the developers of the AI could possibly harvest code from some of these hacks to buff up certain aspects of its early game. I'm not quite sure how much this would benefit the AI, since theoretically wouldn't it already have enough APM to do perfect splits / mass unit selects / mass building selects? And also wouldn't this detract from the learning algorithms that the programmers are trying to implement and thus from the advances in AI from a broader perspective?
Also, from reading the article, the team has so far limited itself to a primarily mutalisk based army composition with slight deviations to respond to early aggression (am I reading this right?). As others have mentioned, videos of more games in their entirety would be nice to see how the AI reacts to various things (including cheese early all-ins). I would be interested in reading how the 2nd place AI worked and approached the learning / adaptation problem, since their army composition consisted of more than 1 type of unit. Also it would be fun to see how (and if) the AI could learn to use different builds and unit compositions in response to scouting, since so far it reads like the scouting is mainly used in a defensive manner until a critical mass of mutas is reached. Looking forward to the advances in SC AI in the future!
EDIT: Sorry, stupid question about the academic papers. I forgot about the review period they have to go through with academic journals before they can get published. Hope that finishes soon >_<.
On January 20 2011 05:53 kamikami wrote: Well MamiyaOtaru you can defend the article as you like but I (and the guys arguing with you before) will just hate it because it gives false impression that SC skill level is low and that's it, you cannot change what people feel while reading it.
Why ? Because the title and the first paragraph give the impression that the human opponent is very high level (WCG competitor, 1st in Spain...), it just doesn't state specifically that it was 10 years ago and nowadays he is no one.
So the whole idea is invalidated because someone might not read the whole article?
The entire "argument" in the thread should've gone like this:
"I don't think the thread title is accurate." "yep it isn't." "okay."
And yes they did say that he's no longer competitive at the game.
No, nobody said the whole idea was invalidated. Try reading more thoroughly. First impressions mean everything. Getting things wrong and giving a false pretense are the worst things you can do to an audience. We're pointing out that the writer did just that.
Yes. Right from the very start you've been saying the intro is bad. Not that the article is crap. hahahaha
Once again, I would differentiate between the pcgamer article (crap) and the arstechnica (mindblowing, well done)
I THINK that's what he's trying to say too.
Thank you for this post. That would explain a lot. I am reacting solely to the Ars article, I didn't even notice the underlined "here" in the first line mea culpa. Still , the worst I can say about it is that it has a terrible title, and cherry picks a bit from the Ars article it then links to, where one can get the full story.
The article is definitely flawed. For people reading it who have little or no prior knowledge of starcraft it would most probably lead them to misunderstand the game. The article does mention a level of strategic depth that the AI isn't programmed to address, and it does imply at one point that it cannot currently consistently compete with humans at the highest level of play, however the great significance of that strategic depth as an obstacle to progamer-level competition is never fully aknoweledged, and is badly overshadowed by a bias of pride in the AI's impressive, yet humble accomplishments.
The article does make it clear however just what is involved in those accomplishments, that these methods of AI programming are not what most people would expect. Rather than simply identifying working strategies in every possible situation, essentially telling the computer what to do without any idea of why it's doing it being necessary, the programmers identified many of the underlying parameters behind the types of decisions that need to be made in starcraft, essentially telling the computer the why without the what. This kind of comprehensive understanding of the fundamental elements of gameplay, with practical demonstration of the effectiveness of such understanding, is very interesting from a game theory point of view.
The article's dishonesty or ignorance is pretty reprehensible and definitely deserves pointing out, but it still has merit that makes it very worth reading.
EPT
mea culpa. Still , the worst I can say about it is that it has a terrible title, and cherry picks a bit from the Ars article it then links to, where one can get the full story.
