EPTBoxeR: "AlphaGo won't beat humans in StarCraft" - Page 11
| Forum Index > SC2 General |
|
Magerius
Canada1 Post
| ||
|
waiting2Bbanned
United States154 Posts
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. I think this might be true for the rudimentary BW bots created to win with a single strat against other bots; this approach wouldn't even work for a chess bot, let alone an artificial neural-network computer with virtually limitless computing power and data resources capable of learning, one which crushed a Go champion. Even limiting its APM to something like 100 it would still roflstomp a human player. | ||
|
Cuce
Turkey1127 Posts
On March 13 2016 23:11 NonY wrote: It also cannot gather any info except by looking at a monitor and listening to headphones. Any other method is clearly a hack. I wonder how ai would react of it will react in time to dt. since there wont be an alert. or how it will scan the screen agains nukes. we as players "feel" something wrong and react to these stuff. well thats just simplified way of saying "players been doing this sutff for so long that neuroplasticity forms a neural reflex akin to muscle memory". The same goes for sublte changes in macroing stuff to fit slight changes in timing due to socuting information, or deviations from meta throught out the game. these stuff are not really conscious desicions, they are more muscle memory. Thats the crux of the issue I thing. yeah sure, it can pull workred rushes everygame to win. But wining is not the result such a development team wants. they want to develop AI, not get gimmicky wins. I wonder how much variables an AI would have to go through to figure out what some to a pro player intuativly. Not sure if we can pull that off in realtime right now. thats another thing, realtime vs turnbased. | ||
|
xQuesian
15 Posts
| ||
|
redviper
Pakistan2333 Posts
On March 14 2016 00:45 Cuce wrote: I wonder how ai would react of it will react in time to dt. since there wont be an alert. or how it will scan the screen agains nukes. we as players "feel" something wrong and react to these stuff. well thats just simplified way of saying "players been doing this sutff for so long that neuroplasticity forms a neural reflex akin to muscle memory". The same goes for sublte changes in macroing stuff to fit slight changes in timing due to socuting information, or deviations from meta throught out the game. these stuff are not really conscious desicions, they are more muscle memory. Thats the crux of the issue I thing. yeah sure, it can pull workred rushes everygame to win. But wining is not the result such a development team wants. they want to develop AI, not get gimmicky wins. I wonder how much variables an AI would have to go through to figure out what some to a pro player intuativly. Not sure if we can pull that off in realtime right now. thats another thing, realtime vs turnbased. You are truly underestimating an advanced AI. AIs build on DNN can certainly create make different decisions based on slightly different scouting information. After a million games the AI would have a much better understanding of the game's strategy than any human being could. | ||
|
TelecoM
United States10682 Posts
| ||
|
LRM)TechnicS
Bulgaria1565 Posts
There are a lot of IFs here. IMO if their purpose is as posed above, in theory if: 1. Google invests enough resources into the project to fulfill their purpose 2. there are no limits for DeepMind to practice and play (for example operating on multiple screens, no APM cap or anything that resembles mouse-keyboard-monitor management in the physical world) and also to gather and analyze information from replays then I think that it will be reasonable to, at least but not limited to, expect DeepMind by operating simultaneously on multiple screens (on preset maps): 1. to execute perfectly his retardedly optimal and or exploitative set of build orders for various matchups vs different actual players/opponents, 2. to have efficient-resource-mining abilities that will far exceed what any human being can physically achieve nowadays 3. to have a micro that will demolish a human player while doing all the above at the same time 4. to know all the possible strategies its opponent is doing just by clicking on a still constructing building of the opponent and checking at what exact % it's from finishing, by how many units and when he has them etc. 5. develop a whole new set of build orders and strategies based on his unlimited capabilities to micro, macro and know what their opponent is doing. Bottom line, if Deep Mind has no limits, it is not unreasonable to expect Deep Mind to stomp all human players. It will feel like you are playing an UMS game titled "IMPO CPU good come play + obs" where everyone is having a laugh out how top players with every race struggle to stay alive mid/late game against DeepMind's ling/queen/scourge/defiler/overlord drops only warcombo. The result will be meaningless as Deep Mind will destroy his opponents that are bound to the physical world. Flash, plays Starcraft with his mind but through tools like mouse, mousepad, keyboard, monitor and not with his mind alone. Limiting APM alone would not do the trick IMO. Google's Deep Mind biggest challenge will be creating something that satisfyingly resembles actual/physical mouse-keyboard-monitor management at least in order to know what kind of stuff are physically possible to be done. This will be needed in order to adequately claim victory in the event of Deep Mind destroying Flash/Jaedong in a bo X series (x>9). | ||
|
Scarlett`
Canada2392 Posts
On March 13 2016 19:20 Liquid`Nazgul wrote: In theory you do know which one they're going to be attacking. As a player you will know which unit will be auto-targeted first, it is just a part of the behavior of units. It isn't random. I don't think this bot knows which one they are targetting by scanning game activity files. This bot knows which one they are attacking, because yes you can determine that based on tank behavior. Professional players play around this (or should) all the time. This video is an even better example (than the marine splitting one) of why AI would totally crush any StarCraft pro given some time to learn. I've seen some suggestions on limiting the APM and amount of clicks to be similar to mouse/keyboard input. I don't think that really matters in the end. I doubt the zergling vs. tank video has a high amount of apm involved. Even if it did you can pretty much get 90% of efficiency with just picking out individual zerglings that are going to be targeted, while the rest of your army is on a-move. It's pretty much just clicking accuracy that you can do with relatively low apms. its 4 actions per ling in the AOE per tank shot (select; move; (wait) select; attackmove); say theres even only 5 tanks and with natural clumping of lings this by itself would require 6000+ APM (assuming theres ~15 lings in the aoe) & dodging with the individually targeted ling also wouldnt work if the units attacking are somewhat clumped up (which is better vs certain armies and worse vs others so deciding how to pre-split is a huge issue in itself) unless the ling is at the edge of the army other than the front ~ one that is indeed very unlikely to be targeted even if the opponent is not controlling their tanks at all it also ignores the fact that the other player can micro the tank -> tell it which unit to target; so the AI would need to have the tank and every possible unit it can be targeting (based on the turret angle) on screen at the time it decides to micro while also knowing whether the tank will have vision of the target long enough to fire because of the delay in the tank shot that makes this micro even theoretically possible, if the terran is controlled by an AI itself it could even re-target the tank after the zerg micros which is much simpler > if i take the previous example thats 2 actions compared to ~60 for a shot all these micro bots built in sc2 are built upon information given directly form the game engine making them trivial to code rather than having to go by the visual and audio output the players are limited to | ||
|
Big J
Austria16289 Posts
On March 14 2016 01:21 Scarlett` wrote: its 4 actions per ling in the AOE per tank shot (select; move; (wait) select; attackmove); say theres even only 5 tanks and with natural clumping of lings this by itself would require 6000+ APM (assuming theres ~15 lings in the aoe) & dodging with the individually targeted ling also wouldnt work if the units attacking are somewhat clumped up (which is better vs certain armies and worse vs others so deciding how to pre-split is a huge issue in itself) unless the ling is at the edge of the army other than the front ~ one that is indeed very unlikely to be targeted even if the opponent is not controlling their tanks at all it also ignores the fact that the other player can micro the tank -> tell it which unit to target; so the AI would need to have the tank and every possible unit it can be targeting (based on the turret angle) on screen at the time it decides to micro while also knowing whether the tank will have vision of the target long enough to fire because of the delay in the tank shot that makes this micro even theoretically possible, if the terran is controlled by an AI itself it could even re-target the tank after the zerg micros which is much simpler > if i take the previous example thats 2 actions compared to ~60 for a shot all these micro bots built in sc2 are built upon information given directly form the game engine making them trivial to code rather than having to go by the visual and audio output the players are limited to You're absolutely right imo, limiting APM greatly alters the challenge for creating the bot. There is a lot of (micro-)management possible with unlimited APM and precision. I think the basic problem of limiting APM is the question "to how little?", and even more "but why even, the point is not to create a human, so why pretend it's a human?". I think the point of such an AI should be that it plays the game to the normal rules, but with maximum efficiency. If that means that it does never box select, it has precise knowledge about unit health from healthbars and gives all orders via the minimap until it needs to identify what enemy dot was actually spotted then that is still fair game. Humans do use the same game mechanics, just not to that efficiency. Which, in my opinion, is the whole point of trying to beat a human with a bot. | ||
|
Liquid`Nazgul
22427 Posts
On March 14 2016 01:21 Scarlett` wrote: its 4 actions per ling in the AOE per tank shot (select; move; (wait) select; attackmove); say theres even only 5 tanks and with natural clumping of lings this by itself would require 6000+ APM (assuming theres ~15 lings in the aoe) & dodging with the individually targeted ling also wouldnt work if the units attacking are somewhat clumped up (which is better vs certain armies and worse vs others so deciding how to pre-split is a huge issue in itself) unless the ling is at the edge of the army other than the front ~ one that is indeed very unlikely to be targeted even if the opponent is not controlling their tanks at all This is accounting for perfection. My view is that AI can be incredibly effective with a lot less than perfection. Pulling out individual units to counter the autotargetting of aoe units is something that can be done without thousands of apm. It won't be perfect, but it will be many times superior to human micro even at low apms. Just imagine your own apm but now with perfect selection onto individual units, perfect ability to predict where automated shots are going to go, and instant reactions and decisions. The difference between that and what you can do now is enormous. it also ignores the fact that the other player can micro the tank -> tell it which unit to target; so the AI would need to have the tank and every possible unit it can be targeting (based on the turret angle) on screen at the time it decides to micro while also knowing whether the tank will have vision of the target long enough to fire because of the delay in the tank shot that makes this micro even theoretically possible, if the terran is controlled by an AI itself it could even re-target the tank after the zerg micros which is much simpler > if i take the previous example thats 2 actions compared to ~60 for a shot all these micro bots built in sc2 are built upon information given directly form the game engine making them trivial to code rather than having to go by the visual and audio output the players are limited to Right, but I think when you introduce an AI into the fold you have to assume its reactions to audio/visual are going to be pretty much instant. What I'm saying though is if audio/visual is instant, the behavior of the tanks can be 'read'. It is all behavior that is standardized and predictable. You can introduce human behavior to battle it of course, that is what will happen when an AI plays a professional, but I don't think it would be close. Just my perspective. 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. Certainly by tournament rules an AI wouldn't be allowed to participate, and neither is it allowed to actually enter Go tournaments. For there to be a fruitful discussion about an AI playing there should be an understanding it wouldn't be playing through headphones and pressing buttons on a keyboard. If we are going to enforce those rules onto an AI we may as well not bother. | ||
|
Tuczniak
1561 Posts
On March 14 2016 00:45 xQuesian wrote: Interesting thing is that with good enough simulation of AI vs AI the game could be balanced at the release date ( and with good design). The patches would only serve to lead the player base in the right way. Although this would require a lot better software than just AI beating humans.If AI gets better than humans, will people start to demand AI vs AI tournaments, since Korean sc won't be the best anymore? | ||
|
Erugua
13 Posts
For me the real question is " can a machine be powerful enough to realise that goal ", and the awnser is obviously yes on SCBW, and maybe not yet on SC2 | ||
|
Oshuy
Netherlands529 Posts
On March 14 2016 02:14 Liquid`Nazgul wrote: Certainly by tournament rules an AI wouldn't be allowed to participate, and neither is it allowed to actually enter Go tournaments. For there to be a fruitful discussion about an AI playing there should be an understanding it wouldn't be playing through headphones and pressing buttons on a keyboard. If we are going to enforce those rules onto an AI we may as well not bother. Would be a fun project. From a robotic point of view, keyboard / mouse handling is not that difficult (trickier if you add the difficulty that the robot must only use 2 arms with 5 fingers each). Identifying elements from screen/sound needs a bit of work, but unless you add random difficulties (lighting issues on the screen, random noise in the room, ...) it is manageable. Knowing the fastest way to execute a list of actions and the ordering is basic, but it would be an interesting way of setting the mechanical limitations on the AI. The main impact for AI optimization is defining how the feedback of the limitation is sent to the action selection... deepmind might just go the "let it learn" route. Not sure they would go for it, but for the looks it could be fun. | ||
nimdil
Poland3751 Posts
| ||
|
beheamoth
44 Posts
If the ai micro manages all of the scouting and just sees everything, it can mathematically work out exactly when builds could have been put down, finished and work out numbers of units out . . they could perfectly respond, eg has y marines, well 8 really well controlled lings beats that so ill pre for this but they could be going tanks seeing that my scout saw tech and 2 gas as x time, 1 tank will be out at x but the cops already started the tech path to what it considers the perfect counter no i think the ai would just crush everything regardless of strategy. That pretty much unimportant if the ai is good enough to hedge bets against everything. this is too difficult to get across in text but no, imagine if you played the perfect game, but lets say you were to macro and the other guy was allining, but you react perfectly on a sub optimal strat . .the computer would make the very best of any situation with the precision micro and timing | ||
|
thePunGun
598 Posts
1. The 'absolute' awareness, which means getting all the information (in exact numbers and the tech) with every scout within less than a second. 2. Having all the info to counter in a huge database and being able to adjust accordingly within seconds. 3. Perfect execution by making every action count(no apm spam, exact commands) which will probably result in superior micro even with apm limited to 200ish. To counter that you would probably have to come up with the weirdest shit ever! Which is why I think that Boxer might actually be able to pull that off, because he is "the emperor" and the very essence of making weird shit work somehow.  | ||
|
necrosexy
451 Posts
On March 14 2016 00:58 redviper wrote: You are truly underestimating an advanced AI. AIs build on DNN can certainly create make different decisions based on slightly different scouting information. After a million games the AI would have a much better understanding of the game's strategy than any human being could. This is like saying if an AI analyzes flipping a coin a million times, it has a better idea of what side the coin will land from this knowledge. The AI cannot reliably determine what course the human will take from the limited scouting input. Moreover, the AI can be duped by the human player with bad info. | ||
|
Pwere
Canada1556 Posts
On March 13 2016 23:11 NonY wrote: It goes further than that. How do you limit the mouse accuracy of the AI? How do you limit its perception so it doesn't analyze every pixel of every frame on the minimap?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. What is, imo, truly a challenge is to develop a live coach that could tell you what to do, and with decent mechanics, its superior strategy and build adaptations would lead a good master player to beat a top Code S contender. You could use one of those programs that finds where your eyes are looking to limit its vision. That's kind of what they're doing in Go... the AI is telling a human what to do to beat another human. That's the only way to make it fair, and would be a huge accomplishment from a theoretical point of view. I also feel like it would be a blast for the player to sometimes not understand what is going on, but still win, and would eventually be a sweet training tool, just like AIs were in Chess (and now Go, apparently). | ||
|
Big J
Austria16289 Posts
On March 14 2016 04:15 necrosexy wrote: This is like saying if an AI analyzes flipping a coin a million times, it has a better idea of what side the coin will land from this knowledge. The AI cannot reliably determine what course the human will take from the limited scouting input. Moreover, the AI can be duped by the human player with bad info. I disagree. One proper scout plus a very precise estimate of how many resources a player could have mined deducted from previous scouting can tell you nearly everything at any point in the game. And an AI may be capable to just force scouts by sheer amounts of micromanagement. Not to mention that such an AI might be aware that it will perform more efficiently in battles, even when outnumbered to some degree and therefore it might just choose to sacrifice more power for information, since it may realize that getting duped is the only way to lose against a mechanically inferior player. There's a lot possible. | ||
|
Befree
695 Posts
The more abstract and less mechanical, the more impressive it would be for a computer obviously. I guess I just wonder where StarCraft falls on this spectrum. The real-time aspect seems to me like it would give a huge advantage to AI which has very little to do with its "intelligence" and more to do with the brute power of computers. I know BW AI tournaments put APM caps to eliminate at least part of this issue, but just the brute power of its awareness seems like it diminishes the meaningfulness of its competitive ability. It's a huge advantage that doesn't come through any sort of creativity or intelligence, just brute strength. | ||
| ||
