EPT*** http://www.starcraftai.com/wiki/Frame_Rate some resources on this
Go - AlphaGo (Google) vs Lee Sedol (world champ) - Page 6
| Forum Index > General Games |
|
Garrl
Scotland1970 Posts
*** http://www.starcraftai.com/wiki/Frame_Rate some resources on this | ||
|
trulojucreathrma.com
United States327 Posts
That's only fair as with chess or go, moving the piece is immaterial, even for a robot. Unless maybe it is speed chess. The chess or go board would be very very easy to interpret compared to a video game screen. The eye-brain-hand coordination challenge is probably as hard of a problem as the game itself. If you are going to pick a challenge just because it is hard to solve, to show you can do things never done before, why exclude half of the problem. Let's see if they can make a hand that can do 350 apm all across the keyboard, using any key combination. | ||
|
MyLovelyLurker
France756 Posts
On March 12 2016 22:34 Taf the Ghost wrote: The one problem with the DeepMind vs SC pro thought is that DeepMind should be required to be limited to the input speed of the keyboard and mouse. It's incredibly dishonest to allow the computer to perform tasks that a Player simply isn't allowed to because of the interface. That's not really a competition at that point, it's simply allowing the computer to abuse parts of the game engine the human player doesn't have access to. But there'd also be some fairly serious technical issues to work through. It's one thing to have access to the direct API of SC:BW, it's wholly another thing to find a way to be able to play a match instantly. That's what would be required for it to constantly run simulations it would need to "learn" the game. Lastly, on the Go matches, considering they're throwing a parallelized super-computer at the problem, even if Moore's Law holds for the next decade, thus making the processing power available to the home user, there's still the issue of 10s of millions in programming money that went into the code to make this work. That's never going to be common.  Under current approach, Deepmind for 2D video games would very likely be restricted to 60 or 120 APM ( one keyboard and/or one mouse for each rendered frame ). | ||
|
Nakama
Germany584 Posts
On March 12 2016 15:26 andrewlt wrote: And the point you are missing is that it really is pretentious to argue semantics about industry jargon as an outsider. I work in aerospace. We have our own acronyms and jargon and code words like every industry does. There are many terms that have a very specific meaning in the aerospace industry. If I learned anything in reading 5 pages of this thread, it is that the term "brute force" has a specific meaning in the computer industry, a meaning that the people who sound like they work in the industry or follow it closely all use, a meaning that you are pointlessly trying to argue the semantics of. I agree with you . But the problem arises when the people from the industry forget that they use the words in a very specific meaning and then use it in an other enviroment etc. for example if u then try to explain the human brain etc. in analogy to the method AlphaGO uses just look at some posts in here and you will see that it´s not even a rare case | ||
|
Pwere
Canada1556 Posts
On March 12 2016 20:21 Gorsameth wrote: This explanation is wrong. Current AIs do not beat the best poker players in complex variations of poker (NLHE, PLO, etc.). While I'm pretty sure it's more a matter of resources than technology, the reason they're not winning is because poker is a game of balance. Your bluffs have to be balanced with strong hands. Your greedy value bets have to appear in spots you bluff a significant amount. Math for a single hand has almost nothing to do with it.Poker is a game of percentages. It is trivial for a computer to calculate its chance of winning at any single point in the game and react "perfectly" to the information available. Over a large enough sample size to even out the element of chance a computer will win, no doubt about it. If a team with the resources of AlphaGo made a PokerAI, they would get there in a few years, at most, but meanwhile, a few dozen players still beat the best AIs. Same goes for games like Magic or Hearthstone. This is the spot Chess was in ~25 years ago, and now your phones are GM level. | ||
|
Grumbels
Netherlands7028 Posts
| ||
BisuDagger
Bisutopia19152 Posts
| ||
|
Sbrubbles
Brazil5775 Posts
| ||
|
MyLovelyLurker
France756 Posts
On March 13 2016 11:30 BisuDagger wrote: Has anyone considered what a game of go would be like if it was AlphaGo vs AlphaGo? This is how AlphaGo was trained. After an initial phase of learning through playing on an online human server, it has been playing millions of games versus itself in the cloud. | ||
|
JazVM
Germany1196 Posts
| ||
|
CCow
Germany335 Posts
He just won. | ||
|
Makro
France16890 Posts
| ||
|
Buddhist
United States658 Posts
There are other ways to exploit the AI's weaknesses also, for example making seemingly illogical decisions (such as building a CC in the enemy's natural, only to cancel it later); it might trick the AI into believing something is happening that really isn't, causing it to do completely the wrong thing. The main issue will be that any advantage the human gains has to overcome the incredible micro and macro advantage of the AI's perfect mechanics. To make the competition fair, it would be sensible to limit the AI to hardware inputs and outputs, meaning it has to read from the monitor and input through mouse and keyboard. You might then also want to set an APM limit, for example of 300. This will force the AI to work with what a human has available to him, and test the AI's strategic and tactical abilities, rather than raw mechanics. | ||
|
JazVM
Germany1196 Posts
On March 13 2016 18:00 Buddhist wrote: The main advantage a human player would have is exploiting fog of war and strategies that the AI has never seen before. AI will optimize for everything it already knows, it can't optimize for things it doesn't know. There are other ways to exploit the AI's weaknesses also, for example making seemingly illogical decisions (such as building a CC in the enemy's natural, only to cancel it later); it might trick the AI into believing something is happening that really isn't, causing it to do completely the wrong thing. . But isn't the point of AlphaGo exactly the opposite? It would "know" that the cc-cancel is fake and would play accordingly because it learned it from previous games? | ||
|
Salazarz
Korea (South)2590 Posts
On March 13 2016 18:00 Buddhist wrote: The main advantage a human player would have is exploiting fog of war and strategies that the AI has never seen before. AI will optimize for everything it already knows, it can't optimize for things it doesn't know. There are other ways to exploit the AI's weaknesses also, for example making seemingly illogical decisions (such as building a CC in the enemy's natural, only to cancel it later); it might trick the AI into believing something is happening that really isn't, causing it to do completely the wrong thing. The main issue will be that any advantage the human gains has to overcome the incredible micro and macro advantage of the AI's perfect mechanics. To make the competition fair, it would be sensible to limit the AI to hardware inputs and outputs, meaning it has to read from the monitor and input through mouse and keyboard. You might then also want to set an APM limit, for example of 300. This will force the AI to work with what a human has available to him, and test the AI's strategic and tactical abilities, rather than raw mechanics. That's not how it works, though. The AI can optimize for things that could 'potentially' be happening in fog of war just like a human can (or better) because it would have access to all the replays of previously played games fed to it. Tricking AIs with seemingly illogical decisions is what people tried a lot in games like Chess and it doesn't work -- a properly built AI will be able to predict potential outcomes of whatever shenanigans you're doing and react accordingly. | ||
|
CCow
Germany335 Posts
So if it's an obvious fake, I am pretty sure that it would act accordingly. Also since it can literally react within no time, it will be incredibly tough to keep it from having next to perfect map knowledge at all times. Limiting the micro might be forced, but it would set arbitrary limits. You could do that, but it is weird. The thing why I feel it is an amazing thing to have for Go is because it has the potential to show "us humans" the boundaries of what is possible. Putting arbitrary limits on that will completely destroy that point. | ||
|
trulojucreathrma.com
United States327 Posts
If you build a CC and cancel, the only info you give it is that you have spend 400 minerals less than you normally would. This allows it to cut corners. Point is exaclty that a AI plays like a robot. It won't panic or overreact like a human. When a human makes a mistake, they overcompensate it the next game, making again a mistake. An AI can more easily find the optimal path because it has no human bias. I do agree you can adapt to a bad static AI in RTS. You can see through it's pattern very quickly. If the AI is static, you can do it in 3 games. | ||
|
Wegandi
United States2455 Posts
On March 13 2016 04:35 Pwere wrote: This explanation is wrong. Current AIs do not beat the best poker players in complex variations of poker (NLHE, PLO, etc.). While I'm pretty sure it's more a matter of resources than technology, the reason they're not winning is because poker is a game of balance. Your bluffs have to be balanced with strong hands. Your greedy value bets have to appear in spots you bluff a significant amount. Math for a single hand has almost nothing to do with it. If a team with the resources of AlphaGo made a PokerAI, they would get there in a few years, at most, but meanwhile, a few dozen players still beat the best AIs. Same goes for games like Magic or Hearthstone. This is the spot Chess was in ~25 years ago, and now your phones are GM level. Hearthstone is childs play. I would love to watch an AI try to play Vintage (MtG) Grixis control mirror against LSV. Even Caw Blade standard mirror would be quite fun. It would have to be a mirror since that would make it an even playing field, unless you had something close to 50/50 (like say Bgx vs UWx control). You can't have say the AI play Tron against UG Infect. Even if the AI had perfect information and did everything perfect they'd still lose 65%+ of the time. In the more complex formats of Magic there simply is too many branching decisions and asymmetric information for the AI to "dominate" like it does in information perfect games like Chess and Go. There's a reason poker people like David Williams and Efro love playing Magic. | ||
|
Cascade
Australia5405 Posts
Now watching game 5 on https://gogameguru.com/alphago-5/ Is Lee Sedol winning? I don't understand.... Where can I play go online (vs comp, but not alphaGo plz)? edit: Can we live-comment the game? | ||
|
Furikawari
France2522 Posts
| ||
| ||
