EPTBoxeR: "AlphaGo won't beat humans in StarCraft" - Page 10
| Forum Index > SC2 General |
|
loppy2345
39 Posts
| ||
|
Big J
Austria16289 Posts
There are tools in SC1 and SC2 that players can only dream to abuse mechanically and information that players know about but which we are simply not capable of processing fast enough, which would be easily handled by an advanced AI. Maybe SC1 offers less of them - I don't know as many specifics about the engine of that game - in SC2 however just certain rushes with bot-control would be probably unbeatable/only beatable by mirroring them in certain matchups. | ||
|
Superbanana
2369 Posts
Its about being able to outplay the human strategically. Not with a good micro system, perfect macro, multitask and attention. Sure the AI could be good at those things and its an achievement in itself. But they will make no point if it looks like a dumb bot winning with "speed". The limited APM idea might be a good way to go. This way the AI must pick good decisions and distribute attention, instead of doing evrything at the same time like a super archon. I don't know how much thought they put on this project for now. But winning at an RTS using no strategy and exploiting the real time part won't display how awesome the AI is. the AI should be limited by hotkeys and control groups, screen vision (not tracking all information outside the fog of war, except for what is provided by the minimap), clicking stuff... at the very least. edited | ||
|
trulojucreathrma.com
United States327 Posts
Just a while ago, when as a chess player you talked about AI with a go player, they were glad to point out that in go, amateur humans wreck the best computers. Go was this elegant game that computers couldn't phantom and wouldn't for a long long time. Also, AlphaGo plays go and only go. And the DeepMind project doesn't have Starcraft as a target yet. I can beat AlphaGo at tic tac toe, which is trivially solvable. Also, RTS games can be set up in a modular fashion. You can define problems, like build order and micro, and solve them independently of each other. This makes it much easier. Also, RTS games are convergent, not divergent. Even in chess the endgame was solved. You could just use a table and the outcome was forced. In go you get more possiblities, not less, like in chess or RTS. In the strategic sense, every ending has a certain theme. | ||
Poopi
France12904 Posts
On March 13 2016 21:33 trulojucreathrma.com wrote: Are Flash and Boxer stupid? What arrogance to claim your game can't be player better by AI than by humans. What do they know about AI? They didn't even go to school. Just a while ago, when as a chess player you talked about AI with a go player, they were glad to point out that in go, amateur humans wreck the best computers. Go was this elegant game that computers couldn't phantom and wouldn't for a long long time. Also, AlphaGo plays go and only go. And the DeepMind project doesn't have Starcraft as a target yet. I can beat AlphaGo at tic tac toe, which is trivially solvable. Also, RTS games can be set up in a modular fashion. You can define problems, like build order and micro, and solve them independently of each other. This makes it much easier. Also, RTS games are convergent, not divergent. Even in chess the endgame was solved. You could just use a table and the outcome was forced. In go you get more possiblities, not less, like in chess or RTS. In the strategic sense, every ending has a certain theme. They are not stupid, it's for PR, plus since you can't have a fair match anyways (because humans are limited by their mechanical abilities) before the real AI issues are solved, a match most likely will never happen. | ||
|
Deleted User 26513
2376 Posts
On March 13 2016 21:33 trulojucreathrma.com wrote: Are Flash and Boxer stupid? What arrogance to claim your game can't be player better by AI than by humans. What do they know about AI? They didn't even go to school. Just a while ago, when as a chess player you talked about AI with a go player, they were glad to point out that in go, amateur humans wreck the best computers. Go was this elegant game that computers couldn't phantom and wouldn't for a long long time. Also, AlphaGo plays go and only go. And the DeepMind project doesn't have Starcraft as a target yet. I can beat AlphaGo at tic tac toe, which is trivially solvable. Also, RTS games can be set up in a modular fashion. You can define problems, like build order and micro, and solve them independently of each other. This makes it much easier. Also, RTS games are convergent, not divergent. Even in chess the endgame was solved. You could just use a table and the outcome was forced. In go you get more possiblities, not less, like in chess or RTS. In the strategic sense, every ending has a certain theme. I think they are saying this, not because they think that it's true, but because "the community" expects them to say this. In reality the human doesn't stand a chance against a well made AI. | ||
|
Superbanana
2369 Posts
The fans don't want to hear something like "meh... its going to be a super AI, no way to win ggwp" from Boxer. I don't think he believes the AI will never win, its a statement that he will do his best when the time comes and also a "come at me bro". | ||
|
Big J
Austria16289 Posts
On March 13 2016 21:29 Superbanana wrote: The real challenge i not to win. Its about being able to outplay the human strategically. Not with a good micro system, perfect macro, multitask and attention. Sure the AI could be good at those things and its an achievement in itself. But they will make no point if it looks like a dumb bot winning with "speed". The limited APM idea might be a good way to go. This way the AI must pick good decisions and distribute attention, instead of doing evrything at the same time like a super archon. I don't know how much thought they put on this project for now. But winning at an RTS using no strategy and exploiting the real time part won't display how awesome the AI is. the AI should be limited by hotkeys and control groups, screen vision (not tracking all information outside the fog of war, except for what is provided by the minimap), clicking stuff... at the very least. edited Yeah, I think limited APM and similar restrictions - like limited click precision, actual cursor movement across the screen - should be invoked, or it is going to be a very uninteresting stomp. The big problem with this however is that there are no such restrictions for humans and thus any player will always seek to exectue as well as he/she can. I think you would have make the AI behave at similar speed/precision to a human, which in itself makes the challenge a bit nonsensical. | ||
|
trulojucreathrma.com
United States327 Posts
Not even AI can outperform humans at being human. You can never make a computer that is more human than a human is. | ||
|
redviper
Pakistan2333 Posts
On March 13 2016 02:38 Axieoqu wrote: I would assume Starcraft would be even easier for the AI because mechanics are so important. Just consider how well the simple blink/micro bots work. In a real matchup we would have an AI that moves a mouse and presses keys. Its not hard to build and it will be very fast, but not infinitely faster than humans. It will also need a way to analyse the data from the screen so will need a capture mechanism. Bots like Automaton are different since they inject commands without physical constraints and (I guess) interpret game data as already processed structure. Also teaching DeepMind to play SC2 (or Broodwar which is even worse!) will be very difficult. A real learning challenge. Unlike Go and Chess, the areas where it must excel are greater (moving mouse, tapping keys), the breadth of options is greater (strategic choices, tactical choices, reactive choices) and there is an inherent lack of full knowledge. RTS would be interesting for deepmind like systems because it needs to work with partial, incomplete and stale knowledge. It may have to deal with feints and misdirections also (like nexus cancel). Boxer may be right for the present. But this kind of learning would be useful for AI, and the ceiling for AI is much higher than humans. | ||
|
redviper
Pakistan2333 Posts
On March 13 2016 21:58 trulojucreathrma.com wrote: If AI wins at RTS against a human, obviously it has to get the game in the realms where AI outperforms humans and get the advantage there. Not even AI can outperform humans at being human. You can never make a computer that is more human than a human is. There is nothing really special about human. We could potentially (in 50 years or so) simulate a whole human brain, the nerve connections, the chemical impulses and the external inputs. About 5 or 6 years ago at SC (supercomputing conference, not starcraft) IBM showed a first simulation of a rat brain. In 50 years you could simulate a whole person. But AI doesn't have to simulate every facet of human to be more human than human (whatever that means). It can learn to have reactions that are indistinguishable from humans (hence the turing test). As I said previously, there is no ceiling on how smart an AI can be, there is for people (the size of the brain and the energy/heat requirements). | ||
|
shadymmj
1906 Posts
| ||
|
DinosaurPoop
687 Posts
On March 13 2016 05:34 Clbull wrote: Bots have already surpassed humans in StarCraft. If you ever saw any of the AI competitons held at the University of California, you'd see bots with superior APM that are able to pull off absurd strategies. https://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp/report2015.shtml#mvm | ||
NonY
8751 Posts
Even if this is a challenge of pure strategy for the AI, it must do it with these hardware limitations to get exposed to all of the elements of strategy of an SC player. A huge part of the strategy of SC is having to adjust it for the level of execution that you can expect to have. Not only that, but figuring out which actions are currently the highest priority is a huge challenge as well. If all actions can be done so quickly as to almost be simultaneous, this whole aspect of strategy can be ignored. | ||
|
nepeta
1872 Posts
On March 13 2016 22:41 DinosaurPoop wrote: https://webdocs.cs.ualberta.ca/~cdavid/starcraftaicomp/report2015.shtml#mvm To be fair, those bots are developed and trained to beat other bots, the most successful ones abuse general or specific flaws of bots, instead of emulating a human game. Deepmind does the oposite, it is trained vs humans. | ||
|
nepeta
1872 Posts
On March 13 2016 23:11 NonY wrote: I don't get everyone who is assuming that the AI is allowed to cheat. The hardware rules are pretty clear for SC. Anything that is supposed to require one button press in the game must take one button press in the real world. Any "macro" that turns a combo of presses into one press is banned. And certainly any system that completely bypasses having to press anything is not going to be allowed. It also cannot gather any info except by looking at a monitor and listening to headphones. Any other method is clearly a hack. These are basic things for anyone who is played competitive SC to understand. Even if this is a challenge of pure strategy for the AI, it must do it with these hardware limitations to get exposed to all of the elements of strategy of an SC player. A huge part of the strategy of SC is having to adjust it for the level of execution that you can expect to have. Not only that, but figuring out which actions are currently the highest priority is a huge challenge as well. If all actions can be done so quickly as to almost be simultaneous, this whole aspect of strategy can be ignored. Some of the best AIs for broodwar can do 45k apm easy, and still be beaten by a d player. The whole question of strategy, ie decision making, has got very little to do with that. A d player with 1000 apm would still be a bad player compared to jd on 200. | ||
|
thezanursic
5497 Posts
Well that's not true, eventually the AI would be good enough. | ||
|
NonY
8751 Posts
On March 13 2016 23:16 nepeta wrote: Some of the best AIs for broodwar can do 45k apm easy, and still be beaten by a d player. The whole question of strategy, ie decision making, has got very little to do with that. A d player with 1000 apm would still be a bad player compared to jd on 200. I happen to have some experience playing BW so I'm familiar with this truth. I'm not sure what it has to do with my post though. | ||
|
Korakys
New Zealand272 Posts
*An external robot interfaces mechanically with a separate computer that is running the game (it can bring its own mouse and keyboard though). *Robot has built in signal delays if required (I think humans peak reaction speed is about 80ms). *Robot is APM limited to a comparable human level (estimate 300APM over a whole game). *Best of 9 against the best human player of a recent tournament where the robot can't be specialised for that match-up. I still think that the AI would still win, but that it could take it years of training to do so. One thing to keep in mind though is that the pace of AI development is not linear - it's going to sneak up on us. | ||
|
nepeta
1872 Posts
On March 13 2016 23:35 NonY wrote: I happen to have some experience playing BW so I'm familiar with this truth. I'm not sure what it has to do with my post though. I know you have played a little broodwar, don't worry on that part  I was trying to say that limiting apm or not has very little bearing on the cognitive capabilities of AIs at this time. The human mind is doing so much parallel stuff in the thinking department that it sometimes is limited by the fingers/apm. AIs are basically at the level of... nematodes and trying to make up for being this by using a few crude heuristics and huge amounts of apm. | ||
| ||
