| Forum Index > SC2 General |
|
perturbaitor
2 Posts
On February 03 2019 03:43 Fecalfeast wrote: cool haiku but what does it mean? 1 post to be snarky in this thread are you polypoetes? robot intellect some unimpressed, flaws, unfair disregard future | ||
|
Grumbels
Netherlands7031 Posts
https://cloud.google.com/tpu/docs/pricing They said that each agent requires 16 TPU's to train, but I don't know how many agents compete in the league. But let's assume it's about a thousand. That means 16,000 TPU's running in parallel (for reference, AlphaZero used 5,000). The maximum pricing given is $4.50 / h. That comes down to a cost of ~1.7 million dollar a day, and about 25 million dollar for the full two weeks of training, not including the initial phase of learning from replays. But I picked the number of agents at random, maybe I'm off by an order of magnitude. And obviously Google's pricing of TPU services to outside clients is not the same as their internal assessment of its cost for in-house research teams. Someone on Reddit said that it would cost them 25 million a day to train AlphaStar, but I don't know how that person came up with those numbers. Possibly many people within the SC2 AI or ML community have access to inside information. But if these numbers are roughly correct, and if you assume that a functioning AI that works for all match-ups and all maps and all patches would require ten times the training, that's hundreds of millions of dollars for training an AI. And that's disregarding the labor costs of the dozens of people working on this project for the last couple of years. | ||
|
shadymmj
1906 Posts
if you want an AI that uses optimal micro to beat a human player, then I'm sure the hardest blizz AI with a marine splitting/stalker blinking/roach burrow cheat would have beaten humans about 5 years ago. | ||
|
pvsnp
7676 Posts
On February 03 2019 13:03 shadymmj wrote: some really moronic people in here, the point of google AI is to outthink a human, not outclick a human. no one would be impressed with an AI that can outmicro pros. if you want an AI that uses optimal micro to beat a human player, then I'm sure the hardest blizz AI with a marine splitting/stalker blinking/roach burrow cheat would have beaten humans about 5 years ago. And believe it or not, the Deepmind engineers are well aware of that. Which is exactly why they've designed AlphaStar as they have. | ||
|
Jampackedeon
United States2053 Posts
I am much less impressed with the matchscore results (10-1) once I learned that the computer could see the whole screen at once in the first 10 games meaning that it didn't have to use its APM to adjust the screen and manipulate the amazing blink stalker micro. But, the presentation was strong and they were very transparent which I appreciate. Probably the most interesting thing overall was that it seems to have proved a case for oversaturating mineral lines in the beginning. I was Zerg player so for me the possibilities for some of the more complex Zerg ability interactions in the future really fascinate me! | ||
|
Rukis1
2 Posts
1. It's an artifiticial intelligent 2. It's programmed to play in the most optimal way. 3. It's training is also inhuman, thus playing other AI will enforce inhuman tactics and most efficient inhuman am. 4. The agents playing all have different strategies, while most of them used stalkers as a majority it vastly differed in how they built around them. Thus, when criticising Alphastar I really wish people would consider the constraints and positives it has. You cannot say definitively that X player trained to play against Y bot will destroy it. As far as I can tell out of all 11 games, Alphastars openings were nearly flawless. It was aggressive, paid attention to buildings, units and position. Let's consider the proxy gate game. We talk about boldly moving up the ramp but I also think many pros would do the same, considering the pressure needed in order to win that particular game. If it had not moved up that ramp, it would of lost that game. Especially since the only units it had was stalkers vs Robo, sentry and few stalkers. Not to mention like I stated before is what Alphastar looks at in considering future probabilities and necessary actions to win. By all means Alphastar isn't unbeatable but you are very undermining it's achievements. That's just down right ignorant and a greedy opinionated thinking. P.S. I would make a more refined post but I'm not up for the challenge at 4am in China. I'm not arguing for or against but I really think people need to analyze this situation with more than "I think" or "it thinks". No one should be winning reviews, the point of a review is to show mistakes and weaknesses. I also really dislike people saying if.... "If this happened then X would win." When in fact we hardly know nothing about how AlphaStar would react to the situation. "But in the last game..." Yeah the last game was one agent, with more human like camera control. We actually don't know if the other agents would respond the same way. Our sample pool is just to small to be critizing and arguing over optimal apm. | ||
|
Eteoneus
20 Posts
| ||
|
Cyro
United Kingdom20295 Posts
Since the non-handicapped versions went 10-0, I think there is no basis to say that they played with flaws that could be abused. A very important note that most people are missing: The version of Alphastar that dropped a game had only been trained for 7 days, vs the 14 days of the by far the most impressive one. Training for half of the time caused an MMR drop of around 1000 while using the camera interface only caused an MMR drop of around 200. | ||
Poopi
France12886 Posts
 It’s really annoying that we didn’t get to see more games without camera, we can only guess what happened had they played 9 more games | ||
|
Cyro
United Kingdom20295 Posts
| ||
|
agsub
Singapore368 Posts
https://www.linkedin.com/pulse/ai-beats-human-players-starcraft-heres-why-matters-kai-xin-thia/ | ||
|
rockslave
Brazil318 Posts
On February 09 2019 06:21 Eteoneus wrote: The back-propagation algorithm will not go uphill in the parameter space. Back propagation is not an optimization algorithm. It is a technique for calculating derivatives. What you mean is gradient descent. | ||
|
Eteoneus
20 Posts
| ||
|
rockslave
Brazil318 Posts
I'm just correcting your wrong terminology and you're being belligerent without reason. | ||
|
Eteoneus
20 Posts
| ||
|
rockslave
Brazil318 Posts
| ||
|
Eteoneus
20 Posts
| ||
|
rockslave
Brazil318 Posts
Back propagation is not an optimization algorithm. It's also not "NN or ML". It's an algorithm created to differentiate functions expressed as the composition of functions whose derivatives are known. It's older than both of us and it's basically the chain rule, known for hundreds of years. The optimization algorithm used in most neural networks is stochastic gradient descent (btw, it goes uphill sometimes because it's stochastic). It is not the only one. L-BFGS was also used with some success. It is not purely gradient based (uses second derivative information), so it also goes uphill. If someday you decide to learn something instead of fighting people who are helping you, I suggest you read Nocedal's book, it's a very good introduction. | ||
|
Incognoto
France10239 Posts
On February 03 2019 05:51 Grumbels wrote: Can anyone help calculate the cost of training AlphaStar? https://cloud.google.com/tpu/docs/pricing They said that each agent requires 16 TPU's to train, but I don't know how many agents compete in the league. But let's assume it's about a thousand. That means 16,000 TPU's running in parallel (for reference, AlphaZero used 5,000). The maximum pricing given is $4.50 / h. That comes down to a cost of ~1.7 million dollar a day, and about 25 million dollar for the full two weeks of training, not including the initial phase of learning from replays. But I picked the number of agents at random, maybe I'm off by an order of magnitude. And obviously Google's pricing of TPU services to outside clients is not the same as their internal assessment of its cost for in-house research teams. Someone on Reddit said that it would cost them 25 million a day to train AlphaStar, but I don't know how that person came up with those numbers. Possibly many people within the SC2 AI or ML community have access to inside information. But if these numbers are roughly correct, and if you assume that a functioning AI that works for all match-ups and all maps and all patches would require ten times the training, that's hundreds of millions of dollars for training an AI. And that's disregarding the labor costs of the dozens of people working on this project for the last couple of years. wait for real? | ||
|
Dangermousecatdog
United Kingdom7084 Posts
| ||
| ||
|
|
|
EPT
