|
On May 28 2017 11:08 Poopi wrote:Show nested quote +On May 28 2017 10:31 Ernaine wrote:
Your post starts with 'In chess, [...], but I am not allowed to talk about chess?
Most things you say here seem contradictory or just overall not very well thought out. Ok, so now you say you prefer not to play vs AI. Fine. But earlier you and other suggested that it is pointless to play a game when there is AI that is better than the best human.
You only registered in 2010, so maybe your way of posting is understandable. Especially after you admitted that you play to humiliate other people, the knowledge that they feel worse about themselves because of you, and that is where you find your satisfaction.
You can not tell if I am trolling? Maybe that is what happens when a scientist talks to an illiterate college kid? They don't know if they are hit with knowledge bombs, or being trolled? What? I said firstly to have fun with my opponent. Because that's exactly what is is. If you win you get to boast, while if you lose it's the opponent.
Again, let me point out how an immature, narrow-minded sense of what 'fun' is this is.
Everyone on TL knows what people mean when saying "mechanics", at least most people have a rough idea.
The fact that you seemingly don't know this and rather speak from an outside perspective is really weird on such a forum.
This all came from this claim of yours:
On May 28 2017 08:59 Poopi wrote:
In chess, mechanics don't matter, ...
Which I dispute. Mechanics are important in chess, in the sense I like how the word can be applied. But the issue is, it is poorly defined. And since you seem to know very little about either AI, chess, CS, algorithms, or anything else, I don't see a point in debating that with you.
The fact that you cherry pick what suits you best also increase the probability of trolling.
Sadly, you pick your own words.
|
France12758 Posts
The fact that you attack my knowledge and the previous attack (scientist vs illetrist) confirms that you are trolling, since you are obviously trying to trigger me into being angry, which is what trolls do.
Sadly you tried to bait the worst candidate for this :x.
Hopefully you'll have more luck in your next troll attempts!
|
There is no way humans beat AI. Is he talking about some form of handicap for the AI?
|
On May 28 2017 02:55 loginn wrote:Show nested quote +On May 28 2017 00:16 LetaBot wrote:On May 27 2017 21:55 XenoX101 wrote:
One problem with AI learning a game like SC2 is that it can't speed up the learning process without access to the game code. (I.e. it cannot play 1,000,000,000 games against itself to learn optimal strategies). So it can only play games at the same pace as humans to gather information. The good part is that it can play 24/7 where a normal player can only play 12/7, and the amount of information it can gather per match is much higher than an AI. It could technically make more precise calculations about engagements than any player.
However it may also be possible for the AI to play games in its "head", if it plays enough games to be able to understand the game mechanics well enough. So then even if it can't physically play a lot of games, it can simulate them at a much faster rate than a human can play them. Technically if its mental model is accurate enough it could bypass playing games altogether, and rely solely on its understanding of the game to run "game experiments" in its head. But the flipside is that unlike Go there are too many possible outcomes for each game, such that you would need an incredibly powerful supercomputer to run through even a small fraction of them. So the AI would have to be very well optimized to only look at "worthwhile" simulations, and ignore the billions of not so valuable simulations (e.g. how to deal with mass reapers or a similarly weak startegy) that would only waste processing time.
EDIT: Thinking about this more I see one way that humans can still win, and that is through "sub-optimal" play that the AI would not expect or would be willing to gamble losing to because most players wouldn't do it. This would be something like a DT rush against someone going Robo tech with observers, or a transition into mass Muta after Terran has gone viking. If the AI doesn't scout it, it would not expect this kind of play. On average it will still likely win most games because of the balance of probabilities, but it would not win 100% due to these kind of oddball games where it would have to predict stupidity. Though this is more reflective of the game itself than the AI, where there are always going to be games that lead to inevitable build order losses. So the real test isn't whether AI can always beat human players, or even if it can beat the best players, but whether it can do so with a higher winrate than any existing player, i.e. greater than Flash's 70% in all matchups. Deepmind works closely together with Blizzard, so they will probably have some way to speed up the game. Sub-optimal play won't work either, since even in BW there are bots that can consider the possible build orders and unit compositions based purely on the time of the game (there are only so many minerals you can gather in a certain amount of time). The main issue of bots right now is actually micromanagement. Even with unlimited apm you still have to make tactical decisions, which bots aren't good at yet. Blizzard already confirmed that the API will allow AIs to play the game as slowly/fast as they want and obviously, unless someone is watching the game, there is no rendering necessary so that's a major part of the workload for every tick of the game that's removed. So now the only limit is computer power which we know google has heaps of. Btw the API's expected functionalities have been documented here for anyone caring to take a look : SpecsUpdate 1Update 2From the specs one of the most interesting parts is this : The ability to load a replay and examine the state of the game as it plays. I'm counting on AIs to point mistakes in my play. Actually I'm actively working on that kind of system
Then I think we can classify this as "Assisted AI", since it is allowed to special privileges to the game (fast playback) that human players do not have. This makes it a bit of an easier problem, since it won't need to interpret monitor pixels or rely on ladder match-making to gather info. However this has the downside of making the AI dependent on the source code of SC2, and not transferable to other games, unless those other games also release their APIs to the AI developers.
I guess you can't really blame them since developing AI is hard enough, forcing it to learn with the same limited information that a human has (that is purely visual info) may be outside of reach. But eventually this should be the goal, because if this can be solved, then the AI will be able to learn games that don't readily disclose their source code, or even real-world scenarios that don't have any source code. Evidently this would be a much more powerful, as well as a much more fair AI, since it doesn't need any help from anyone (Blizzard or otherwise) to become good at the game.
|
On May 28 2017 15:44 XenoX101 wrote:Show nested quote +On May 28 2017 02:55 loginn wrote:On May 28 2017 00:16 LetaBot wrote:On May 27 2017 21:55 XenoX101 wrote:
One problem with AI learning a game like SC2 is that it can't speed up the learning process without access to the game code. (I.e. it cannot play 1,000,000,000 games against itself to learn optimal strategies). So it can only play games at the same pace as humans to gather information. The good part is that it can play 24/7 where a normal player can only play 12/7, and the amount of information it can gather per match is much higher than an AI. It could technically make more precise calculations about engagements than any player.
However it may also be possible for the AI to play games in its "head", if it plays enough games to be able to understand the game mechanics well enough. So then even if it can't physically play a lot of games, it can simulate them at a much faster rate than a human can play them. Technically if its mental model is accurate enough it could bypass playing games altogether, and rely solely on its understanding of the game to run "game experiments" in its head. But the flipside is that unlike Go there are too many possible outcomes for each game, such that you would need an incredibly powerful supercomputer to run through even a small fraction of them. So the AI would have to be very well optimized to only look at "worthwhile" simulations, and ignore the billions of not so valuable simulations (e.g. how to deal with mass reapers or a similarly weak startegy) that would only waste processing time.
EDIT: Thinking about this more I see one way that humans can still win, and that is through "sub-optimal" play that the AI would not expect or would be willing to gamble losing to because most players wouldn't do it. This would be something like a DT rush against someone going Robo tech with observers, or a transition into mass Muta after Terran has gone viking. If the AI doesn't scout it, it would not expect this kind of play. On average it will still likely win most games because of the balance of probabilities, but it would not win 100% due to these kind of oddball games where it would have to predict stupidity. Though this is more reflective of the game itself than the AI, where there are always going to be games that lead to inevitable build order losses. So the real test isn't whether AI can always beat human players, or even if it can beat the best players, but whether it can do so with a higher winrate than any existing player, i.e. greater than Flash's 70% in all matchups. Deepmind works closely together with Blizzard, so they will probably have some way to speed up the game. Sub-optimal play won't work either, since even in BW there are bots that can consider the possible build orders and unit compositions based purely on the time of the game (there are only so many minerals you can gather in a certain amount of time). The main issue of bots right now is actually micromanagement. Even with unlimited apm you still have to make tactical decisions, which bots aren't good at yet. Blizzard already confirmed that the API will allow AIs to play the game as slowly/fast as they want and obviously, unless someone is watching the game, there is no rendering necessary so that's a major part of the workload for every tick of the game that's removed. So now the only limit is computer power which we know google has heaps of. Btw the API's expected functionalities have been documented here for anyone caring to take a look : SpecsUpdate 1Update 2From the specs one of the most interesting parts is this : The ability to load a replay and examine the state of the game as it plays. I'm counting on AIs to point mistakes in my play. Actually I'm actively working on that kind of system Then I think we can classify this as "Assisted AI", since it is allowed to special privileges to the game (fast playback) that human players do not have. This makes it a bit of an easier problem, since it won't need to interpret monitor pixels or rely on ladder match-making to gather info. However this has the downside of making the AI dependent on the source code of SC2, and not transferable to other games, unless those other games also release their APIs to the AI developers. I guess you can't really blame them since developing AI is hard enough, forcing it to learn with the same limited information that a human has (that is purely visual info) may be outside of reach. But eventually this should be the goal, because if this can be solved, then the AI will be able to learn games that don't readily disclose their source code, or even real-world scenarios that don't have any source code. Evidently this would be a much more powerful, as well as a much more fair AI, since it doesn't need any help from anyone (Blizzard or otherwise) to become good at the game.
Yeah, I see this kind of assistance kind of necessary. Strictly speaking, even with the API's help, the AI itself, making the strategic decisions, is still realized fairly in my opinion, because it doesn't have access to more information than humans do, it just gathers said information in a different way. Creating an AI that can interpret all the information purely (audio)visually as humans do is not strictly a "strategic" task so to say, and I imagine it would be significantly harder to realize than the already impressive goal set for AlphaGo right now (not sure if it would be possible at all currently). It would be closer to a fully functional human AI than to a StarCraft bot, in my opinion.
|
|
|
This was pretty much the "logical" next step after they created Alphago zero. By going "full reinforcement learning" and nothing else, the rules of each games are only new parameters to pass to the program and are not hardcoded into the program anymore. So making the same program play different game was pretty logical.
It's still damn impressive though. Especially since it (allegedly) runs on lighter hardware than all previous Alphago iterations !
However I don't believe they could just "plug" that same program into sc2. All the previous games are kinda similar in structure (perfect information, turn based system, etc), and sc2 is pretty far from that.
Still, by going with a full reinforcement learning approach, AlphaSC2 would evolve with his own meta and not by affected by our meta at all. Looking at that would be super interesting !
|
I though one of the basic idea was to train in actual matches, not against itself?
|
On December 07 2017 02:37 AbouSV wrote:
I though one of the basic idea was to train in actual matches, not against itself?
Nah it needs to play against itself to play enough games that it can learn effectively. It learned chess in 4 hours, but played 44 million games. If it was restricted to real-time play it would never get good.
|
On December 07 2017 02:37 AbouSV wrote:
I though one of the basic idea was to train in actual matches, not against itself?
My understanding is that they made a new AlphaGo that can learn by solely playing against itself (no external data), and eventually it does massively better than one that has been trained in actual matches (that is, fed previous matches by humans, then allowed to learn through its own games).
I think I heard one commentator say that this new AlphaGo was able to trounce the version that beat Sedol. That's pretty amazing to me, considering how crazy good the normal AlphaGo was.
|
On December 06 2017 23:31 LoneYoShi wrote:It's still damn impressive though. Especially since it (allegedly) runs on lighter hardware than all previous Alphago iterations !
I want to point out that the hardware used for AlphaZero and Stockfish was very different. According to the development forum for Stockfish AlphaZero used 4 TPUs of 45 teraflops each which vastly outmatched the hardware used for Stockfish. If given the same processing power it's not at all clear that AlphaZero would have won.
|
On December 07 2017 03:28 Mendelfist wrote:Show nested quote +On December 06 2017 23:31 LoneYoShi wrote:It's still damn impressive though. Especially since it (allegedly) runs on lighter hardware than all previous Alphago iterations ! I want to point out that the hardware used for AlphaZero and Stockfish was very different. According to the development forum for Stockfish AlphaZero used 4 TPUs of 45 teraflops each which vastly outmatched the hardware used for Stockfish. If given the same processing power it's not at all clear that AlphaZero would have won.
Hmm, in the forum (I assume I found the same one as you) they also mention that the paper states AlphaZero evaluates about 80k positions per turn while Stockfish evaluates 70 million. So I don't think it has much to do with the hardware. I'm not sure if the TPUs are used for learning in which case that would just take longer, but still.
|
On December 07 2017 03:54 GoloSC2 wrote:
Hmm, in the forum (I assume I found the same one as you) they also mention that the paper states AlphaZero evaluates about 80k positions per turn while Stockfish evaluates 70 million. So I don't think it has much to do with the hardware. I'm not sure if the TPUs are used for learning in which case that would just take longer, but still.
The number of evaluated positions per turn or second is not relevant when comparing different engines. AlphaZero gets its strength from a very good and processing heavy evaluation function, hence its slow speed.
|
On December 07 2017 03:54 GoloSC2 wrote:Show nested quote +On December 07 2017 03:28 Mendelfist wrote:On December 06 2017 23:31 LoneYoShi wrote:It's still damn impressive though. Especially since it (allegedly) runs on lighter hardware than all previous Alphago iterations ! I want to point out that the hardware used for AlphaZero and Stockfish was very different. According to the development forum for Stockfish AlphaZero used 4 TPUs of 45 teraflops each which vastly outmatched the hardware used for Stockfish. If given the same processing power it's not at all clear that AlphaZero would have won. Hmm, in the forum (I assume I found the same one as you) they also mention that the paper states AlphaZero evaluates about 80k positions per turn while Stockfish evaluates 70 million. So I don't think it has much to do with the hardware. I'm not sure if the TPUs are used for learning in which case that would just take longer, but still.
Yeah, but they gave Stockfish 64 cores and only a 1 GB hash which is pretty sub-optimal no matter how you cut it.
|
An AI being successful at BW would probably be the same as if passing the Turing Test. There are aspects of BW that I don't think can be learnt by a machine. Remember, AI can never be intuitive, so it can never want or need. Baduk is turn based, so the AI makes it's move accordingly- when will an AI want to or need to attack his opponent in BW? Never.
|
In any case, I'm not trying to downplay Deepminds success here. The mindboggling thing is that Alphazero got it's world class strength by self play in four hours and no human input, not that it beat Stockfish, which is questionable.
|
An AI need a fixed set of rules to be really strong, most if not all boardgames maintain the same ruleset forever, hence AI can get extremly good at them.
Starcraft 2 have patches that in a way changes alot of things, or even add or remove something from the game.
The most logical way as I see it for AI would be to be set to pick random and then teached to execute the most hard-to-stop cheeses with every race. Then maybe it would do top GM performances, because it might to late to scout and if the human throws down a wall of turrets/bunkers the AI would be made so it either cancels the rush and goes economy or find other path such as transport the forces behind the turrets or take map control.
Anyhow, many theorys about this subject. My main point is that Starcraft is a game that has rule-changes and diffrent maps, chess and GO etc is on a fixed set of rules on a fixed map/board.
|
On December 07 2017 04:45 saalih905 wrote:
An AI being successful at BW would probably be the same as if passing the Turing Test. There are aspects of BW that I don't think can be learnt by a machine. Remember, AI can never be intuitive, so it can never want or need. Baduk is turn based, so the AI makes it's move accordingly- when will an AI want to or need to attack his opponent in BW? Never.
But there are also some aspects that the AI has the advantage in. E.g. it has the potential to use its apm more efficiently than any human opponent.
|
On December 07 2017 05:21 GothGirlGames wrote:
An AI need a fixed set of rules to be really strong, most if not all boardgames maintain the same ruleset forever, hence AI can get extremly good at them.
Starcraft 2 have patches that in a way changes alot of things, or even add or remove something from the game.
The most logical way as I see it for AI would be to be set to pick random and then teached to execute the most hard-to-stop cheeses with every race. Then maybe it would do top GM performances, because it might to late to scout and if the human throws down a wall of turrets/bunkers the AI would be made so it either cancels the rush and goes economy or find other path such as transport the forces behind the turrets or take map control.
Anyhow, many theorys about this subject. My main point is that Starcraft is a game that has rule-changes and diffrent maps, chess and GO etc is on a fixed set of rules on a fixed map/board.
While you're right in that the rules of SC2 change with patches, I think that's not that big a problem as you'd just have to run the learning process again. I mean, humans have to adapt to the new patch as well.
Also "teaching" the AI to cheese is probably the worst approach, at least AlphaGo's strength improved with less restrictions from the programming side. AlphaGo changed the Go meta if you will, it's whole point is not to do well what humans came up with but to find strategies/tactics on it's own.
I don't want to necessarily disagree with you, because the as you say Go/Chess and SC2 (or SC:R for that matter) are very different, but after reading a bit into it I find it hard to stay skeptical of the AI's potential.
|
It is just boardgames, if you have a generalish algorithm to win at Go then it makes sense you could use the same algorithm to win at chess. And if you have a really good algorithm for Go, then improving it slightly (for a game like Go where it has no competition other than a previous version of itaelf, and where the better player always wins), will create these results of a seemingly unbeatable engine.
But they clearly can’t just trivially adapt this algorithm to SC2, or else they would have done it by now. AlphagoZero only needed hours or days of training.
I would guess they don’t have to start from scratch, but it might be awhile before they know how to use these techniques for SC2.
|
|
|
|
|
|
EPT
