| Forum Index > General Forum |
Chess discussion continues here | ||
|
calgar
United States1277 Posts
On November 17 2013 04:32 Grumbels wrote: No, he gains between 3 and 4 points per win. His unoffical live is 2872 right now.http://en.chessbase.com/post/chennai-g6-carlsen-wins-second-straight Btw, will Carlsen break the 2900 barrier if he wins the next game? See the 2700 chess live ratings tracker for further details. | ||
|
kusto
Russian Federation823 Posts
On November 17 2013 05:13 Grumbels wrote: Are you by chance a second year Computer Science student with no real experience with building an AI? We're talking about constructing an algorithm for finding a globally optimal decision for EVERY timestamp in the game for it to be ideal. In chess, these are less than 100 in an average game with long breaks in between. In a real-time game (SC or any RTS), you need to perform all these computations at each of these timestamps - i guess that's hardly possible. EDIT: Sure, you can build an AI which executes perfect micro (not the strategic aspect that i'm interested in), therefore beats any human player. ... if that's what you wanted to tell me. | ||
|
gingerfluffmuff
Austria4570 Posts
On November 17 2013 05:54 kusto wrote: We're talking about constructing an algorithm for finding a globally optimal decision for EVERY timestamp in the game for it to be ideal. In chess, these are less than 100 in an average game with long breaks in between. In a real-time game (SC or any RTS), you need to perform all these computations at each of these timestamps - i guess that's hardly possible. Btw. you haven't brought up any arguments yet. Idk, you could scale the intervalls with your computing power. Also imo the used gas approach paired with the seen enemy units should give the AI enough info to build the counters. | ||
|
ghrur
United States3786 Posts
On November 17 2013 05:22 Kentredenite wrote: The fact that SC has fog of war doesn't matter if your goal with the AI isn't to solve the game. If all you want to do is make an AI that beats all human players, you just have to program it to do something that will beat human players, like execute some micro-intensive all-in and execute it perfectly. Compare that to how it took hundreds of millions of dollars' worth of research to get chess engines to the point they are now. The way chess engines are designed doesn't point towards solving chess, either. They rely heavily on human-programmed "evaluation" functions to tell them how good a certain position is because it's still unfeasible to search all possibilities even 10 turns ahead. To solve chess, you'd have to either examine every single possible game or demonstrate somehow that the ones you didn't examine weren't worth examining. But people have tried to make AIs that beat humans. Every year, there's an AI competition for SC:BW with some of the top CS schools competing like Berkely, RPI, UIC, Waterloo, etc. None of these AIs come close to beating a human simply because of imperfect information. Executing 1 build order is fine. In fact, they often do implement 1 build order and abuse units which have high micro ceilings, but having AIs adjust their moves based on inferences because they LACK certain information is harder. There's no AI that's going to have the starsense Flash did in BW, and that's going to make a huge difference. Also: http://xkcd.com/1002/ | ||
|
Skwid1g
United States953 Posts
On November 17 2013 06:01 ghrur wrote: But people have tried to make AIs that beat humans. Every year, there's an AI competition for SC:BW with some of the top CS schools competing like Berkely, RPI, UIC, Waterloo, etc. None of these AIs come close to beating a human simply because of imperfect information. Executing 1 build order is fine. In fact, they often do implement 1 build order and abuse units which have high micro ceilings, but having AIs adjust their moves based on inferences because they LACK certain information is harder. There's no AI that's going to have the starsense Flash did in BW, and that's going to make a huge difference. Also: http://xkcd.com/1002/ They can do risk assessment better than any human, Chess has had 10000x more money/effort put into producing AIs so it's not a very fair comparison. Not to mention the fact that if we're allowing "perfect micro" it'd be incredibly easy to code an all-in that was basically a 100% win against any human. Even your link says top R&D could change it, and I'm quite sure it would. It all depends on what limitations you place on the AI. | ||
|
nosliw
United States2716 Posts
| ||
|
Grumbels
Netherlands7032 Posts
On November 17 2013 05:54 kusto wrote: We're talking about constructing an algorithm for finding a globally optimal decision for EVERY timestamp in the game for it to be ideal. In chess, these are less than 100 in an average game with long breaks in between. In a real-time game (SC or any RTS), you need to perform all these computations at each of these timestamps - i guess that's hardly possible. EDIT: Sure, you can build an AI which executes perfect micro (not the strategic aspect that i'm interested in), therefore beats any human player. ... if that's what you wanted to tell me. If you look at people that are knowledgeable versus people that are ignorant about AI's: you are probably some evil hybrid where you know just enough to throw out some terms but not enough to make sense. Minimax is soooooo irrelevant to Starcraft 2 AI research that it hurts my brain whenever someone brings it up. | ||
|
ghrur
United States3786 Posts
On November 17 2013 06:06 Skwid1g wrote: They can do risk assessment better than any human, Chess has had 10000x more money/effort put into producing AIs so it's not a very fair comparison. Not to mention the fact that if we're allowing "perfect micro" it'd be incredibly easy to code an all-in that was basically a 100% win against any human. Even your link says top R&D could change it, and I'm quite sure it would. It all depends on what limitations you place on the AI. Except it wouldn't work because you can counter the all-in if a computer just does it deterministically. Suppose the AI does BBS, I build a barracks earlier, build up a ramp, and live through it. What now? I mean, you all say it's incredibly easy to do it, so why haven't people done it yet? There are graduate research in CS project at one of the world's leading CS schools being done on this, and yet it's incredibly easy to do? I find that hard to believe... And risk assessment provided they have perfect information... How does an AI determine whether or not a person is going 2port wraith, drops, hidden expo, etc. if they don't have perfect information? What heuristics do you use to determine this? What happens if the computer isn't able to scout the heuristic? How do you adapt it? There's a lot of unanswered questions. In fact, why don't I just challenge one of you to code that incredibly easy all-in and see for yourself. Alternatively, you can get an idea of how good these AIs are here: http://www.twitch.tv/certicky Also, that link wasn't meant to be proof... it was meant to lighten the mood, lol. | ||
|
wingpawn
Poland1342 Posts
Sample Huge study The main difference is that out of remaining 40-30%, almost every single "not the best" move requires your opponent to play the most precise sequence that can be very long and include many possible tempting deviations or really dangerous counters that only chess program can calculate to the end as harmless. | ||
|
The_Red_Viper
19533 Posts
On November 17 2013 06:17 Grumbels wrote: If you look at people that are knowledgeable versus people that are ignorant about AI's: you are probably some evil hybrid where you know just enough to throw out some terms but not enough to make sense. Minimax is soooooo irrelevant to Starcraft 2 AI research that it hurts my brain whenever someone brings it up. All you do is say people have no idea, i am not impressed. So pls enlighten us who think like kusto, cause clearly you know it better... I think Kusto explained it quite well, if you don't agree pls use some arguments.. | ||
|
Aerisky
United States12129 Posts
| ||
|
heartlxp
United States1258 Posts
you can almost feel through the movements alone that Anand is feeling impatient and eventually becomes mentally crushed sad to see but crazy stuff by Carlsen hopefully Anand can recover for some more exciting games! | ||
|
undyinglight
United States611 Posts
| ||
|
heartlxp
United States1258 Posts
On November 17 2013 07:38 undyinglight wrote: I am rooting for Anand as he and Morphy are likely my two favorite players of all time (Unless you count Deep Blue as a player). Though he is not a GM Josh Waitzkin is a player I love as I learned so much from him in Chessmaster 4400 back in the day. Though I am still holding out hope for Anand to take a victory I have been very impressed by Carlsen's play going up two games over Anand. If Anand pulls it back this is going to be a mighty comeback! man Waitzkin is the best, i learned so much from his lessons too! too bad there's no chessmasters updates anymore  | ||
|
nFo
Canada56 Posts
| ||
|
Grumbels
Netherlands7032 Posts
On November 17 2013 07:24 The_Red_Viper wrote: All you do is say people have no idea, i am not impressed. So pls enlighten us who think like kusto, cause clearly you know it better... I think Kusto explained it quite well, if you don't agree pls use some arguments.. If you look at the performance of the strongest human chess player versus the strongest computer chess player then they are still somewhat close with the human player having chances every other game. On the other hand, a minimally competent hypothetical Starcraft 2 AI would be way beyond the capabilities of any human player. It would be more along the lines of matching up Ursain Bolt with a Ferrari and see which one will win the race. This is based on the strength of a computer to execute micro and macro perfectly which is beyond the level of human players. All you need to do is find some sort of safe build that lets you get to mid-game at which point you can send some medivac drops or whatever around and immediately win. kusto being mystified at people finding enjoyment in chess when computers can find the best moves is actually much like being mystified at people enjoying starcraft because while spectating a game you can see the mistakes of the players. It's an ignorant perspective that only a non-chess player could have. For actual chess players it's irrelevant that you can check the best move with a computer much like how it's irrelevant that you can check the replay later on to see what you did wrong while playing starcraft. (of course there are some cons to computers and I'm personally not a devoted fan) And well, bringing up the concept of game tree search because technically starcraft 2 is a zero-sum two person turn based game is so beyond irrelevant that I don't even know where to start. Nobody was talking about "perfect play based on theoretical optimal outcomes based on game tree search" because it's totally not relevant when building an AI that can beat human players. | ||
|
Deleted User 173346
16169 Posts
| ||
|
bananafone
68 Posts
 On November 17 2013 07:24 The_Red_Viper wrote: All you do is say people have no idea, i am not impressed. So pls enlighten us who think like kusto, cause clearly you know it better... I think Kusto explained it quite well, if you don't agree pls use some arguments.. Alright, hes original post implied chess to be inferior to Starcraft on the basis that computers were simply better at it than humans. A bold claim, even if it was only implied. Several arguments was then made that if someone wanted to create an AI (and used the same amount of effort as that used to build the chess AIs) it would be rather easy to create one witch had the same properties as the chess ones. That being that it would be able to beat all human players. We're talking about constructing an algorithm for finding a globally optimal decision for EVERY timestamp in the game for it to be ideal. In chess, these are less than 100 in an average game with long breaks in between. In a real-time game (SC or any RTS), you need to perform all these computations at each of these timestamps - i guess that's hardly possible. This part is where it goes completely wrong. If the point from this is to say that constructing a perfect AI is hard, then yes you're completely right, it is. It is also infinitely harder than constructing one for chess. However the original argument was an implication that starcraft was superior to chess on the basis that computers are better at chess than they are at starcraft if the performance parameter is success vs tophumans. These two arguments have absolutely nothing to do with each other, unfortunately. | ||
|
Cel.erity
United States4890 Posts
| ||
|
Sprouter
United States1724 Posts
| ||
| ||
StarCraft 2 StarCraft: Brood War League of Legends Super Smash Bros Heroes of the Storm Other Games summit1g5716 Liquid`RaSZi2263 FrodaN548 shahzam528 WinterStarcraft466 PiGStarcraft281 ToD256 C9.Mang0203 Livibee157 JimRising  131ArmadaUGS40 ZombieGrub36 |
|
The PondCast
WardiTV Qualifier
Online Event
Korean StarCraft League
Afreeca Starleague
Rush vs Soulkey
CranKy Ducklings
AI Arena Tournament
BSL Open Qualifier
Sparkling Tuna Cup
IPSL
[ Show More ] |
|
|
EPT
