On March 13 2016 17:46 rabidch wrote: it will take a while for google to come up with this though (a few years or more), mostly because of hashing out plans with blizzard (assuming theyll cooperate), designing how theyll train and do decisions, and then getting enough computing power and time to train. even with google's cloud i think it will take a massive amount of computing to train the thing, assuming they choose neutral networks like they did with alphago
Why should they need permission from Blizzard though? Their other AIs would work with just visual input, it is not like they would need access to the game state.
without access to game, with same control inputs given to human players, I think it will take quite a bit of time to optimize even how ai controls the game.
On March 13 2016 17:46 rabidch wrote: it will take a while for google to come up with this though (a few years or more), mostly because of hashing out plans with blizzard (assuming theyll cooperate), designing how theyll train and do decisions, and then getting enough computing power and time to train. even with google's cloud i think it will take a massive amount of computing to train the thing, assuming they choose neutral networks like they did with alphago
Why should they need permission from Blizzard though? Their other AIs would work with just visual input, it is not like they would need access to the game state.
Its technically botting, and I think blizzard could sue a large conpany if they are developing illegal tools - thats why permission.
Im pretty sure an advanced Ai would crush any human player in starcraft.
Just look at those micro bots. Add perfect macro to that and they would need one build order to win 100% of games.
The advantage ai will always have In any game is 100% memory, beeing able to calculate multiple outcomes and perfect multitasking.
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.
I think this point is the biggest issues. Here is an example of what it could accomplish and some.
but isnt that cheating, like we arent supposed to know what the siege tanks are targeting?? why would the AI know which ling is being targeted. its not like a seeker missle where it shows up alerting you. its different then like when a unit is taking damage u then micro it away. Tank shot comes without you really knowing where its gonna hit, maybe the AI could guess, but if your controlling ur tanks u can click on something else no?
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.
but isnt that cheating, like we arent supposed to know what the siege tanks are targeting?? why would the AI know which ling is being targeted. its not like a seeker missle where it shows up alerting you. its different then like when a unit is taking damage u then micro it away. Tank shot comes without you really knowing where its gonna hit, maybe the AI could guess, but if your controlling ur tanks u can click on something else no?
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.
I think you can not make AI vs Human without apply the mechanical constraints to the AI that a Physical player has.
So I guss an AI without any restraint could easily win against any human in SC(l2). 40000 APM? Why not. Perfect macro, perfect micro and millions of games to find out the best reaction possible. First constraint must be APM and Data Input. AI should only see the screen (Pixels), and interact through mouse/kb input. There will be a serious task of "recognition" alone.
Easiest way: Make a robot play the game. And I dont see a robot handel sc(2)
On March 13 2016 19:25 KT_Elwood wrote: I think you can not make AI vs Human without apply the mechanical constraints to the AI that a Physical player has.
So I guss an AI without any restraint could easily win against any human in SC(l2). 40000 APM? Why not. Perfect macro, perfect micro and millions of games to find out the best reaction possible. First constraint must be APM and Data Input. AI should only see the screen (Pixels), and interact through mouse/kb input. There will be a serious task of "recognition" alone.
Easiest way: Make a robot play the game. And I dont see a robot handel sc(2)
The problem is in
Accuracy
Ability to read and predict
No wasted clicks
Always 'having an eye on the screen' aka response time
I'm pretty sure you can make an absolutely crushing AI with relatively low apm (200 or so should be enough). No need to assume this AI is going to need 50k apm for the above. It will help of course it will be able to do even more, but the gains are pretty marginal since speed is one of the least important factors when you can account for the above listed points. Limiting the apm isn't going to do much.
On March 13 2016 10:45 Pwere wrote: Discussing a Starcraft AI is barely better than discussing a CS:GO AI. You first need to limit the huge mechanical advantage before you can discuss what is and isn't possible.
From a purely theoretical point of view, SC is not an interesting challenge for an AI programmer/designer, so I doubt the resources will be dedicated to this for a while, and by then it won't be all that impressive.
If they plan to adapt a general AI to play Starcraft, then that is a different challenge, but the outcome still comes down to where they draw the mechanical/perceptual line.
Totally agree. It doesn't make any sense to beat top pros with limited strategy and perfect execution and then brag about it from an AI perspective. A waste of time in terms of prestige.
It would be kind of awesome to see an AI do a probe rush in BW and win with perfect micro. At least it'd be awesome the first time, and it wouldn't be awesome because of the AI but because it's such an infuriating, unsolvable way to lose.
"unsolvable", you mean no human being can rush a probe tight wall?
Agreed on APM cap in usefullness. I see the problem in access on "game state" = 100% possible data in handy format and "screen/pixeldata only" And AI with access to game state can have possible 100% flawless multitask wich will win you most Sc2 games. If you never miss a beat, really never, then you can afford to send units out to scout everything at any time. Build 99% Depots around the map and always cancel on last sec would give you so much information. Or have 3 Reapers/lings around that give you vision on any angle of attack (on perfect micro those can always escape !). Players might not do that because they need XX minerals for BO but a computer could just scratch those XX mins from perfect Worker-Timings perfect mineral mining and so on. I think of it as a possible 10 player Archon with mind-synchronisation. A 10-Player Archon woul always have 100% of information possible and most of all "awareness", not only minimap-vision
On March 13 2016 17:46 rabidch wrote: it will take a while for google to come up with this though (a few years or more), mostly because of hashing out plans with blizzard (assuming theyll cooperate), designing how theyll train and do decisions, and then getting enough computing power and time to train. even with google's cloud i think it will take a massive amount of computing to train the thing, assuming they choose neutral networks like they did with alphago
Why should they need permission from Blizzard though? Their other AIs would work with just visual input, it is not like they would need access to the game state.
Its technically botting, and I think blizzard could sue a large conpany if they are developing illegal tools - thats why permission.
Im pretty sure an advanced Ai would crush any human player in starcraft.
Just look at those micro bots. Add perfect macro to that and they would need one build order to win 100% of games.
The advantage ai will always have In any game is 100% memory, beeing able to calculate multiple outcomes and perfect multitasking.
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.
but isnt that cheating, like we arent supposed to know what the siege tanks are targeting?? why would the AI know which ling is being targeted. its not like a seeker missle where it shows up alerting you. its different then like when a unit is taking damage u then micro it away. Tank shot comes without you really knowing where its gonna hit, maybe the AI could guess, but if your controlling ur tanks u can click on something else no?
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.
really? always box-clicking the zerglings surrounding the targeted zergling and individually pulling them away and that multiple times at once. I'm pretty sure the APM are insane in that scenario.
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.
but isnt that cheating, like we arent supposed to know what the siege tanks are targeting?? why would the AI know which ling is being targeted. its not like a seeker missle where it shows up alerting you. its different then like when a unit is taking damage u then micro it away. Tank shot comes without you really knowing where its gonna hit, maybe the AI could guess, but if your controlling ur tanks u can click on something else no?
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.
really? always box-clicking the zerglings surrounding the targeted zergling and individually pulling them away and that multiple times at once. I'm pretty sure the APM are insane in that scenario.
I mean sure it isn't a walk in the park, but for a computer with perfect accuracy I don't think it needs to be abnormally high to get a high reward. You can continue to make it better to pretty much unlimited apm, though. Essentially pulling away zerglings and selecting single zerglings with perfect accuracy should be doable while keeping the numbers low.
The Emperor has spoken so Google needs to look up some other things. He gonna bunker rush x3 and score gonna be 3-0. If AlphaGo risks then we gonna get new pimpest play where BoxeR humiliates the machine, hell it's about damn time.
Has anyone posted any links to some of those micro bots? I remember one where lings were micro'd perfectly against tanks, so each tank shot only one died. Made lings look ungodly powerful.
On March 13 2016 20:48 Fighter wrote: Has anyone posted any links to some of those micro bots? I remember one where lings were micro'd perfectly against tanks, so each tank shot only one died. Made lings look ungodly powerful.