|
On February 01 2019 01:24 Dangermousecatdog wrote:
Polypoetes, you make an awful lot of assumption that doesn't quite bear out. Pro players generally do want to win over if they win decisively. You get the same ladder points and tournament money no matter how much you think you have won or lost a game by.
But this completely ignores what it means to be human and how humans actually learn. I am making assumptions? Your claim is literally that humans are able to objectively take their human experiences, objectively form their goal, and rewire their brains so it happens more. That is not how humans learn.
Humans learn by re-enforced learning as well. But what is the re-enforcement? You clicking on the screen, trying to kill the enemy army and it either working or failing? Or you looking at the ladder points after the game?
Overmaking workers is an opportunity cost after 16 where you don't get your money back till 2.5 mins after you queued the worker. It makes sense if you are planning to transfer or losing units to harrass. Zerg players for example notably do and does overdrone.
What are you even trying to say?
The point of deepmind PR stunt was not to show it can win against the best human player (mana and tlo aren't even close to the best) but to show that it could outstrategize humans, but in general it just outmuscled them with massive and accurate spikes of APM.
No. Starcraft is a complex game with hidden information. They have an AI that can beat top players. They showned they have the AI architecture and techniques to solve this problem. The solving is in the winning, not in the impressing SC2-playing college kids.
The reason they will play with APM limits is because of the well-known problem of finding minima in very high dimensional phase space. There will be minima in the phase space that are deep, but those cannot be found because they are surrounded by agents that play very badly. Think strategies or play-styles that only work when it is very fine tuned and specifically executed, but then it does really well.
Clearly it finds an easy good minimum using blink stalkers. The question is how good the minima for other unit compositions are. And to cross the phase space from one minimum to the other, it has to travel over a peak, and it needs an incentive to do so. You can do that by rewarding it to build certain units, which they already tried. But you can also do it with an APM limit, because it is clear that it's potential for APM have a lot of synergy with stalkers. So you put an APM limit that changes the phase space and makes stalkers much less optimum, and then find new minima where it uses strategies with other units. And once you are there, you remove the APM limitation and you either find out it stays in that minimum and deepens it further, or it moves back to stalkers.
Yes,a AlphaStar vs AlphaStar match says nothing to us about how we as humans should play or enjoy SC. But that isn't important at all. Look at their Go and Chess projects. They have done very little to 'help' those communities. They do their research into AI using games as the problem to solve. They aren't doing it to help the game in question. They aren't helping the chess community figure out it they should switch to Fischer random, for example.
I really enjoyed seeing AlphaZero's chess games. I don't think I will enjoy watching AlphaZero play SC2 against itself. It says something about the game, not about the AI.
|
United States12224 Posts
On January 31 2019 20:31 Polypoetes wrote:
I don't know how they can place Mana and TLO on that graph. Maybe this is their Blizzard ladder MMR? Anyway, you cannot compare MMRs of two distinct populations. And furthermore, MMR doesn't take into account that there is a rock-paper-scissor effect where certain styles counter others. This is clearly so in SC in general, and Deepmind has made a point out of it several times that their best agent is consistently beaten by an agent that isn't so highly rated. And one reason for them to have this match is to see how play strength in their agent league translates to play strength in the human realm.
I think what they probably did was draw assumptions based off a common data point: where the Elite AI ranks in the skill spectrum among humans. They claimed that the Elite AI is equivalent to "low Gold". Presumably they had a more specific internal MMR value from Blizzard, and were also told what MMR differences equate to what outcome probabilities (as well as probably the exact rating formula itself). DeepMind's primary benchmark was their supervised learning agents which could beat the Elite AI "95% of the time". That translates to a certain rating gap (let's say 500). As the agents train against each other in the AlphaStar League, they can measure win percentages against each other, so if Agent A wins 95% of the time against Agent B, and Agent B wins 95% of the time against Agent C, then you know that Agent A is 1000 MMR higher than Agent C. Eventually you can plot out MMRs for each of the participating agents.
You're of course completely correct that you can't compare MMRs of distinct populations. The environments and the metagames are just too different. We have long known that "6k" on NA is not quite the same as "6k" on EU, which is not quite the same as "6k" on KR, and the best benchmarks we can get are the few players who participate in every region simultaneously, but even then it won't be exact. So "6k" in the AlphaStar League might be TLO level, or it might be higher or lower than that. "7k" in the AlphaStar League might be MaNa level, or it might be higher or lower than that. I guess it sort of serves as an okay point of reference for illustration purposes, but it's by no means scientific fact. Since much of the DeepMind news is marketing-oriented and promotional with laypeople as a target audience, it's important to keep the information they present in perspective.
|
It has nothing to do with the NA ladder or whatever ladder. It is baked into the mathematical definition of Elo. MMR is Elo-based.
|
On January 31 2019 21:55 Polypoetes wrote:
I think you are very wrong to think that making an AI good at SC2 is going to result in an AI that can trick and outmindgame humans. If you think so, you are delusional about what kind of game SC2 is. And the AI will actually show this.
I have asked this before and I still haven't really got an answer from anyone. But how did you think the AI would outsmart the human and win without outmicroing and battle decision making? How would that look?
I think I already answered. Having the AI model the game from the opponent's perspective and factor this inferred information into its decision making process would resemble intelligent behavior much more closely. This is what we call reading the game.
Maybe it is because of my experience and background, but I think I understand that the AI would just always seem 'lucky' and just win. Which is exactly what we saw in these games. People say the AI didn't scout for DTs and would have auto-lost vs DTs, for example. I don't think so. I think it knew. Maybe not the TLO one, but the one Mana played against, pretty sure. Same with the Pylon build in TLO's base and where there AI used all these shield batteries to keep one immortal alive. I think it didn't bug out and place a pylon there. I think it placed it there to be killed, so the opponent's stalkers don't do something else like scout the choke. Same with the Stargate it build at the ramp, then cancelled it as it was scouted, then rebuild it in the back of the base.
Another obvious thing is the AI building more probes while SC2 players think 2 a patch is enough because Blizzard put n/16 over each nexus.
You are giving AlphaStar too much credit. Those could've been brilliant moves, but just as well they might've been blunders that were rendered irrelevant by the AI's far superior mechanics. There's no way of telling.
I think this holds true in general, that an AI that cannot be beaten by humans but that is making decisions that seem to be mistakes are likely not mistakes. This we saw in go where top players said Alphago made mistakes and weak moves. You can only show them to be mistakes by exploiting them and winning them. Of course this was expecially relevant in go because Alphago always knew which move to make to get a position that wins 52% of the time with 1 more point than your opponent, compared to winning only 51% of the time but now with a larger point difference. It was able to ignore at which margin it would win. So if this translates to Starcraft, then the AI rarely wins decisively, but as a result loses way less. Humans don't just want to win. They want to win convincingly. If you win but it feels like you never played any better, you aren't that satisfied.
How are you drawing that conclusion? AlphaStar by default must've underestimated its chances in any engagement. It mostly practiced with opponents that matched it in terms of mechanics. It could very well be the case that AlphaStar's predictions in terms of how favorable an engagement will be are correct only roughly 50% of the time, but MaNa's "subpar" execution made them appear correct most of the time.
I already said I agreed that in the future it will be interesting to make AIs that play weak like a human. But that is certainly not what Deepmind's goal is so far. They want to show they can 'solve' the problem of SC2 by winning against the best player.
If their goal was to make an AI that beats good SC2 players (while cheating, at that), then that's hardly an accomplishment. What would be their next goal, CSGO bots with aimbot? I highly doubt that is their goal...
To all these people that think the bot didn't out-strategize anyone; you will never be satisfied by the bots ability to strategize, because the nature of SC2 is not what you think it is. For you to get what you want you need a different game.
No, we just need the AI to play on roughly equal terms mechanically, or even have it be weaker mechanically Then there will be no way for it to compensate its strategic shortcomings.
On February 01 2019 01:58 Polypoetes wrote:No. Starcraft is a complex game with hidden information. They have an AI that can beat top players. They showned they have the AI architecture and techniques to solve this problem. The solving is in the winning, not in the impressing SC2-playing college kids.
Yes, and they conveniently rigged the games in a way that makes the incomplete information aspect irrelevant.
And no, they do not have an AI that can beat top players. Currently they have an AI that is easily confused and exploitable, and wins vs. skilled human opponents through far superior mechanics and not decision making.
The solving is not in the winning. That's trivial. Solving would be creating an AI that is mechanically worse (or roughly equal) than a skilled human but wins regardless.
|
“ think I already answered. Having the AI model the game from the opponent's perspective and factor this inferred information into its decision making process would resemble intelligent behavior much more closely. This is what we call reading the game.”
This strikes me as overreaching. We don’t know if the AI needs to explicitly model the opponent in order to make good decisions. Humans don’t really play like that either in standard gameplay. It might be that you can come up with strats and builds that are, to a degree, independent of what your opponent is doing, or that force them to react to your play in a predictable fashion.
I mean, at some point you’re demanding that AI’s can on the fly react to a completely novel strategy before they’re considered good, which is a harsher standard than what we apply to humans. Most likely it will never come up because its standard decision making in terms of micro and engagements will be so sharp that SC2 will have no space for something utterly bizarre.
|
Ignorance is one thing, but I cannot believe how arrogant your comments are. Neural networks don't work by modeling the thoughts and intentions of the opponent. This description is also extremely vague. How can you boldly state that you already answered, and then give such a non-answer?
The only conclusion I can draw is that to you an AI that plays SC2 really well doesn't match your definition of 'intelligence'. I suggest you redefine your definition while also reading up on how a neural network plays a game of chess or a game of SC. Simply put, you have a big matrix of game state data. You put that through a very large network of weights, and you get an output from which you compute a game move. You then optimize the weights in that network to get the output you desire.
This simply means that you think that a our current Deeplearning neural network can never be 'intelligent' because it doesn't have 'thoughts'; it doesn't reflect on itself or it's opponent.
As for the 'brilliant moves' made, how am I giving AlphaStar too much credit? It won 0-10! That's all I can go off. Because it is a neural network I cannot judge it as I would judge a human. If it seems only barely stronger than TLO/Mana, but it still wins, then that means nothing. It is optimized to win, not to impress or crush. Yes, those actions may have been blunders. But those kinds of moves are exactly what you would expect from a neural network. That is again why I am so confused because people expected something very different, which they cannot put into words how that would look like.
As for the 'conclusions', they are not conclusions. They are an example. For AlphaGo, Deepmind had an additional system to evaluate the chances at winning. People criticized AlphaGo for making what humans thought to be subpar moves, because it was trying to win more reliably with a small margin while humans would try to win more terrain. It understood what it meant to 'win' a lot better than humans, because it always considers all parameters and evaluates those objectively. And it does so as correctly as the weights have been set. It is just an example literally taken from Go, a more complex game, to show that humans were wrong to evaluate AlphaGo. But you just read that statement and thought 'How did he get that 51%', which is baffling to me. So I doubt you are understanding properly what I wrote just now.
Then you say that their goal of creating an AI that can beat the best players at SC2 is 'hardly an accomplishment'? What does this even mean? That you think they picked the wrong game? This is another puzzling statement. Only a year ago, people here were claiming that it would be a complete theoretical impossibility to create an AI that can beat the best player at SC. Even now, there are people that think humans will be stronger than the best AI Deepmind can come up with, if given the time to practice against it. So how can you say that it is trivial? They have a big team there with a whole bunch of smart people with scientific careers in deep learning/machine learning. They have Google investment and infrastructure. They worked on many other games before they worked on SC; not just go and chess. And you think it is 'hardly an accomplishment'. The people that throw money at Deepmind don't think so.
And yes, that is literally what it means to make a SC2 AI; as if you are making a CSGO aimbot. I think that was intended as an argumentum ad ridiculum, but it isn't. Why don't you go ahead and code some CSGO AI that misses intentionally. CSGO is also a much more straightforward game. You can program an AI to just walk around randomly and it will do pretty well. Furthermore, AI is much better at teamwork than humans are, because it will be a single entity. All teamgames are mostly about teamwork and AI teamworking with AI is a trivial problem.
You also lie about Deepmind making the incomplete information irrelevant. You must mean that it sees the entire map. But I know you know that it still has the fog of war. So it still has perfect incomplete information. So you are not just arrogant, you are also a deliberate liar. Yes, you could have made a point out of the window control, but Deepmind already pre-empted that entire criticism by developing an agent that didn't use that. That agent bugged out and lost. No one denied that this restriction makes the problem more difficult. Which is exactly why they initially didn't put on that restriction. So every single move that AlphaStar did in the 10-0 games, a human player could also have made. Just not all those moves put together.
You also say that the AI is easily confused and is exploitable. Yet if we talk about the complete map version of AlphaStar, it never seemed confused and was never exploited. TLO and Mana on the other hand, they said they were very confused. AlphaStar never played vs humans, TLO and Mana never played vs AlphaStar. And who was the one confused? The humans. And both of them knew they had to try to exploit the AI. Mana only succeeded in doing so by sending forward a single stalker and having the AI's phoenixes lift it in vain. That's about the degree of exploitabilty I saw.
But I guess making an AI that goes 10-0 vs strong human players, that doesn't get confused and is barely exploitable, and simple outplays a human in micro, tactics, build orders, and decision making is trivial.
This really feels like the MBS discussion all over again. People stupidly and against all evidence argue that SC2 is a game of strategy and mindgames. It clearly is not. Any AI beating a top player under any circumstance, you will always have an excuse about why it is 'cheating'. And if you give up on that it is 'simply winning, which is trivial, because it is a machine'.
|
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money.
|
On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money.
Im curious about your "well over a billion in the red". Could you provide a source or sources for that? I am not really doubting it or calling you out, I just legitimately want to educate myself on it. Though I'd point out that it is common business practice to say that things are as unprofitable as possible, if you can get away with it.
Also, I'd like to say that I agree with most of your message, but I do think Deepmind is about more than shareholders. I think that google leadership realizes the level of talent and potential deepmind attracts, and gives them a certain amount of flexibility in what they choose to do with their resources. It wouldn't surprise me if they make some choices (like giving even more "human" type limitations to alphastar), simply for the fun and the challenge.
|
On February 01 2019 05:48 Grumbels wrote:
“ think I already answered. Having the AI model the game from the opponent's perspective and factor this inferred information into its decision making process would resemble intelligent behavior much more closely. This is what we call reading the game.”
This strikes me as overreaching. We don’t know if the AI needs to explicitly model the opponent in order to make good decisions. Humans don’t really play like that either in standard gameplay. It might be that you can come up with strats and builds that are, to a degree, independent of what your opponent is doing, or that force them to react to your play in a predictable fashion.
Humans do play that way, except we only model a fragment of the game we find relevant. In my earlier example I noted how scouting Nexus first and probably a few other things (Probe count, minerals mined, etc.) precludes the possibility of the opponent doing a proxy build (a hypothetical example, I don't know how SC2 PvP works). The way we analyze the information is "he couldn't possibly have had enough minerals for a proxy Gate". The AI module I'm proposing would instead simulate the whole game up to that point, come up with a number of probable game tree paths and rank a proxy opening as highly unlikely.
Sure, what you're describing is also true, but that's even more advanced I'd say. At least making the opponent play a certain way. While it may be the case that modelling the game is not necessary for good decision making, I think it would be an improvement as it would give the decision making module information that is not readily available by "looking" underneath the fog of war. I think that is the essence of what makes a game like SC2 different from a game like chess, and DeepMind are not really addressing that aspect.
@Polypoetes
I will reply tomorrow as it's late now. For now let me just say that it's ironic to accuse me of arrogance seeing as you didn't bother to read what I propose and are constantly making assertions regarding DeepMind's goals regarding SC2 that directly contradict what they say about their AI.
I also never claimed that AlphaStar is maphacking or such. I repeatedly said that it beat TLO and MaNa due to far superior mechanics, which made the incomplete information aspect of the game irrelevant. The same way having a team of aimboters in CS:GO makes tactics irrelevant. You're basically saying that aimboters are tactically superior because when they choose to engage, they dominate. Ridiculous.
|
On January 30 2019 20:38 Acrofales wrote:Show nested quote +On January 30 2019 19:42 Grumbels wrote:On January 30 2019 09:51 maybenexttime wrote:Does anyone know what game speed AlphaStar is playing at during its internal games? Do I remember correctly that they mentioned 200 years of experience in a week? Was it combined playtime across all agents? What I'm wondering is whether they could make an evolutionary algorithm that is trained to reconstruct a replay from one player's perspective. It's very different from simply teaching it to win. Such an approach would teach it how to model the state of the game from incomplete information. The main problem would be quantifying how faithful the reconstruction of a replay is. Then they could turn it into a module and incorporate it into AlphaStar, and make it model the game it is currently playing in real time (assuming it can simulate numerous games of SC2 that quickly). It could come up with realistic scenarios explaining what the AI already knows about the opponent. It could create working hypotheses regarding what has been happening behind the fog of war, and perhaps even verify them via scouting. Is what I'm proposing very far-fetched?  I don't know if I'm understanding you correctly, but you could imagine some sort of implementation where an AI has a belief about the opponent's units and economy, which it acts upon in a game and then verifies via watching the replay. I haven't read the paper they released yet, but from some comments I read I don't think it has these capabilities currently. Also, I don't like spreading misinformation, but I /recall/ having heard that the figure of 200 years is the playtime of the agent which has played the longest time. The week of training probably also includes the initial stage of imitation learning from replays. Depending on how long this lasted, it would mean that if the agent playing vs TLO had 200 years of practice, then the one playing vs Mana, which trained for another week, would have at least 400 years of experience, but possibly much more. But it might be best to read the paper. I mean, the ratio of a week to 200 years is like 1 : 10,000 , and I'm pretty sure you can't speed up SC2 that much even with good hardware and eliminating graphics. So a single agent has to be able to train in parallel with itself. This is a good point. I'm not sure. It would mean that a game of SC2 of normally ~30 minutes would be played in 0.2 seconds. Even having the map and everything loaded into memory in advance, that seems *very* fast to simulate SC2 with 2 quite heavy RL algorithms making the decisions on both sides. On the other hand, they are running it on a rather powerful setup. 16 TPUs can run a pretty hefty NN in very little time. However, the SC2 engine itself is not easily parallelized, and it still needs to compute every unit's actions every step of the simulation.
In theory they could run multiple instances of the same agent simultaneously, then train from datasets of replays, right?
Also starcraft can be sped up a lot. Excluding loading which takes about 5 seconds, I can run a 5 minute game in about 10 seconds on my shitty desktop. So maybe they really can pull it off.
|
On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money.
I don't think we fail to understand the bigger picture. Quite the contrary. If you transition from a complete information game to an incomplete information game specifically due to this difference, circumventing that aspect of the game defeats the purpose.
|
United States12224 Posts
On February 01 2019 09:19 travis wrote:Show nested quote +On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money. Im curious about your "well over a billion in the red". Could you provide a source or sources for that? I am not really doubting it or calling you out, I just legitimately want to educate myself on it. Though I'd point out that it is common business practice to say that things are as unprofitable as possible, if you can get away with it. Also, I'd like to say that I agree with most of your message, but I do think Deepmind is about more than shareholders. I think that google leadership realizes the level of talent and potential deepmind attracts, and gives them a certain amount of flexibility in what they choose to do with their resources. It wouldn't surprise me if they make some choices (like giving even more "human" type limitations to alphastar), simply for the fun and the challenge.
The only figures I can find are that DeepMind's expenses in 2017 were $400 million, more than double its previous year. If that trend continued into 2018, then it's possible. I have a friend who has old academic colleagues in the AI/ML field and they were expressing frustration at the ability to compete with conglomerates like Samsung, Google, and Facebook when it comes to solving problems because these giant companies can just throw endless gobs of money at these research challenges. The figure I heard was that it "cost" DeepMind $25 million per day to train AlphaStar. I say "cost" in quotes because it uses Google's TPUs, and DeepMind is a Google subsidiary, so they can effectively write off the real cost. However, if some third party were to use that same cloud computing power, that's how much Google would have charged them. Obviously, you're not going to get anyone in the academic space to raise that sort of money continuously, so DeepMind's AIs can get trained up much faster and the company as a whole can move much more quickly.
But it is worth noting that while DeepMind does make some sales in the medical field, it's still "in the red" because it's first and foremost a research wing of Google. It's a known cost that will eventually recognize returns in various forms as its AI algorithms become sufficiently advanced (for example, it was responsible for reducing cooling costs at Google data centers by 40%, and improving the longevity of phone batteries on Android 9 by using adaptive brightness).
|
On February 01 2019 09:19 travis wrote:Show nested quote +On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money. Im curious about your "well over a billion in the red". Could you provide a source or sources for that? I am not really doubting it or calling you out, I just legitimately want to educate myself on it. Though I'd point out that it is common business practice to say that things are as unprofitable as possible, if you can get away with it.
Pretty sure that information should be public too, since Alphabet is publicly traded. Suffice to say that Deepmind costs a pretty penny. After some quick googling:
https://qz.com/1095833/how-much-googles-deepmind-ai-research-costs-goog/
https://www.forbes.com/sites/samshead/2018/10/05/deepmind-losses-grew-to-302-million-in-2017/#308def2d490e
Also, I'd like to say that I agree with most of your message, but I do think Deepmind is about more than shareholders. I think that google leadership realizes the level of talent and potential deepmind attracts, and gives them a certain amount of flexibility in what they choose to do with their resources. It wouldn't surprise me if they make some choices (like giving even more "human" type limitations to alphastar), simply for the fun and the challenge.
Oh for sure, Deepmind is basically R&D. Very prestigious R&D. The bottom line is not important here, and Google is more than happy to throw money at it, even more so than usual. But there's still faith in Deepmind returning that investment tenfold, eventually.
On February 01 2019 09:48 maybenexttime wrote:Show nested quote +On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money. I don't think we fail to understand the bigger picture. Quite the contrary.
First you say this.
If you transition from a complete information game to an incomplete information game specifically due to this difference, circumventing that aspect of the game defeats the purpose.
Then you say that? Please explain how exactly superhuman mechanics allow AlphaStar to magically pierce the fog of war.
|
can't wait for the AI to keep improving
|
On February 01 2019 09:19 travis wrote:Show nested quote +On February 01 2019 06:36 pvsnp wrote:
All these people quibbling over "winning" and "fairness" fail to understand the bigger picture. Starcraft is not the end, merely a small stepping stone in the long road of ML, AI, software engineering, and corporate profit. AlphaStar is a precocious toddler playing in the kiddie pool of Starcraft, before its cousins go on to tackle bigger and better things.
AlphaStar is a work in progress, but even when it is completed, how exactly it plays and wins games of Starcraft are little more than PR. If adding APM caps, or camera restrictions, or what have you will significantly contribute to AlphaStar's learning ability then it will be done. If not, it won't. The way it actually plays Starcraft is purely incidental to the real goals.
So long as it improves Deepmind's understanding of AI, so long as it fulfills the technical requirements of Google's engineers, so long as it satisfies the demands of Alphabet's shareholders, then it will be a success.
Deepmind is well over a billion dollars in the red. The budget for AlphaStar alone is orders of magnitude above anything the entire professional Starcraft scene could ever dream of. Follow the money. Im curious about your "well over a billion in the red". Could you provide a source or sources for that? I am not really doubting it or calling you out, I just legitimately want to educate myself on it. Though I'd point out that it is common business practice to say that things are as unprofitable as possible, if you can get away with it. Also, I'd like to say that I agree with most of your message, but I do think Deepmind is about more than shareholders. I think that google leadership realizes the level of talent and potential deepmind attracts, and gives them a certain amount of flexibility in what they choose to do with their resources. It wouldn't surprise me if they make some choices (like giving even more "human" type limitations to alphastar), simply for the fun and the challenge.
Yeah, I think it’s important to keep this sort of anti-capitalist / materialist analysis in mind, but I don’t think it fully governs Deepmind’s actions. They have some degree of independence and ‘artistic license’. Obviously every corporation will eventually be assimilated into the logic of global capitalism, but for now it’s a little above that.
@Excalibur_Z
Btw, I read that Deepmind is responsible for the current voice to text bot google uses.
|
the ai beats a korean or just foreign? also i wanna see it on bw
|
@pvsnp
Circumvent means to go around. AlphaStar makes this aspect of the game relatively irrelevant. The same way having a team of aimboters in CS:GO makes tactics irrelevant. The fact that the aimboters dominate any fight when they choose to engage doesn't make them tactically superior.
The games MaNa lost were rigged in many ways when it comes to the engagements. First of all, there was a vast gap in terms of mechanics between MaNa and AlphaStar - both in terms of battle awareness (not being limited to one screen in case of the AI) and superhuman APM peaks. Secondly, MaNa's experience worked against him. He admitted that he misjudged many engagements due to not being used to playing opponents with such mechanics. Before each battle MaNa overestimated his chances whereas AlphaStar underestimated its chances.
|
I have two main issues with this whole thing.
1) It's pretty clear that control plays a bigger role than decision making in those wins. Continuing to make mass stalkers against 8 immortals is not good decision making. Having your whole army go back to the back of your base 4 times to deal with a warp prism while you could just cross the map and go fucking kill him instead is not good decision making. I think it looks especially bad because it's bad decision making in a way that is somewhat obvious, like, very few humans would make those bad decisions. We are used to "bad decision making" that is much subtler than that.
2) I don't like the PR strategy of DeepMind. It seems like they have to hype the fuck out of the accomplishments that they get, and it makes the whole thing seem really artificial to me. I don't have the exact quotes in mind any more but what they said about this starcraft experience felt overreaching when I read it; what they said about the poker experience was even worse, but the poker experience was somewhat more convincing than the starcraft one (it had issues as well).
edit: my mistake, just realized Libratus wasn't made by the same guys. But the same principle applies to both.
|
On February 01 2019 01:58 Polypoetes wrote:Show nested quote +On February 01 2019 01:24 Dangermousecatdog wrote:
Polypoetes, you make an awful lot of assumption that doesn't quite bear out. Pro players generally do want to win over if they win decisively. You get the same ladder points and tournament money no matter how much you think you have won or lost a game by.
But this completely ignores what it means to be human and how humans actually learn. I am making assumptions? Your claim is literally that humans are able to objectively take their human experiences, objectively form their goal, and rewire their brains so it happens more. That is not how humans learn. Humans learn by re-enforced learning as well. But what is the re-enforcement? You clicking on the screen, trying to kill the enemy army and it either working or failing? Or you looking at the ladder points after the game? What are you even saying? You quote me but don't actually say anything that interacts what I am saying. Your assumptions are still false assumptions. And then you write some nonsense. Do you even play SC2?
|
On February 01 2019 19:24 Nebuchad wrote:
I have two main issues with this whole thing.
1) It's pretty clear that control plays a bigger role than decision making in those wins. Continuing to make mass stalkers against 8 immortals is not good decision making. Having your whole army go back to the back of your base 4 times to deal with a warp prism while you could just cross the map and go fucking kill him instead is not good decision making. I think it looks especially bad because it's bad decision making in a way that is somewhat obvious, like, very few humans would make those bad decisions. We are used to "bad decision making" that is much subtler than that.
2) I don't like the PR strategy of DeepMind. It seems like they have to hype the fuck out of the accomplishments that they get, and it makes the whole thing seem really artificial to me. I don't have the exact quotes in mind any more but what they said about this starcraft experience felt overreaching when I read it; what they said about the poker experience was even worse, but the poker experience was somewhat more convincing than the starcraft one (it had issues as well).
edit: my mistake, just realized Libratus wasn't made by the same guys. But the same principle applies to both.
I don't think you can claim that making stalkers is bad decisionmaking at all. On paper, immortals hard counter stalkers. And in human control they do too. But if you have Alphastar micro capabilities, then suddenly they don't anymore. I think you're mixing cause and effect a bit here. Alphastar learned to make stalkers in most situations *because* it also learned to micro them incredibly well. That seems like a legitimate strategy. It's like when MKP showed that if you split your marines they didn't just get blasted into goo by a couple of banelings, and if you did it well, then suddenly banelings no longer countered marines very well at all. Sure, he microd marines FAR better than his contemporaries, but was his choice to then just make lots of marines a bad choice? Clearly not.
As for (2). They are a commercial enterprise. Of course they're going to hype their accomplishments. What did you think? That said, if you actually watched the video, the guys there are quite honest about their achievements and their aspirations. I don't think they believe they have "solved SC2". Or poker, for that matter, although I suspect poker is pretty close to being solved in all its various forms, whereas SC2 will take a bit longer. Still, Alphastar is quite a remarkable achievement, even with its flaws, and they are justifiably proud of it.
|
|
|
|
|
|
EPT
