The google bot having a mechanical hand controlling the mouse physically is what I proposed when this idea of tackling RTS first came out. It does put a hard limit on what bots can do. Having no AMP ceiling and being able to control 100+ individual units has proven it's worth and effectiveness in SC AI.
But in the end, two things are true.
Whatever humans can do, AI can do. Because humans are just another type of AI.
Any problem can be reduced to a 'data problem', given enough computational power. All the possible game states in any RTS are going to be finite. In principle, you can write code that just exhausts the phase space that is all game states. In the end, all Go and chess AI are just about severely limiting the amount of game states that need to be sampled/checked/evaluated.
Will Google be able to do something impressive? I am not sure. The idea of the bot playing against itself to improve, that may have worked great for Go. But I can see how in RTS that would just push the AI into a way of thinking/playing that would be unique to AI, and easily bypassed and defeated by a human. But maybe with this thought, I fall for the same trap as many have before me, when discussing AI playing chess and go.
Without APM limits, I see a good team write a program that beats Flash in SC within a year. For SC2, I don't know enough about it. I know it is easier than SC BW. But how much easier? And if it is all about mind games/reads/obscure timings, how do you win a mind game vs an AI?
Have the AlphaGo use a mouse cursor and internal buttons like a keyboard, and play that way. Otherwise its completely unfair. If it's not using the same tools as human then there is no point. Its like having a thousand cursors and keys.
On May 27 2017 07:43 Achamian wrote: Have the AlphaGo use a mouse cursor and internal buttons like a keyboard, and play that way. Otherwise its completely unfair. If it's not using the same tools as human then there is no point. Its like having a thousand cursors and keys.
Thats practically what they are doing. It can still reach unthinkable APM levels if not restricted, though. (And I think they're doing that)
On May 27 2017 06:46 mishimaBeef wrote: Yeah if you consider move = click, then it explodes. But usually you think in terms of high level "moves" (tech to vessel, pump marine medic, deflect muta) and use clicks to implement the higher level strategic "moves".
I was talking only about the representation of the current position, as an input to a neural net, which would then classify it as good or bad. This was used (in combination with monte carlo simulation to also evaluate positions) to then determine what the best next move to make was.
All machine learning algorithms have to deal with the curse of dimensionality, where you run into issues the more features you have in the vector representing your training example. In GO, the input is relatively compact. Its just the position of all the pieces on the board. In an RTS, you would have the position of all the units, their hitpoints, their upgrades, the position of all the buildings, their hitpoints, your worker count, minerals, vespene, and so on. Worse, there would be a fog of war, whereas GO you have all the information on the board readily available. In an RTS, the strength of a position is not independent of what the opponent is doing. So your input would have to take into account all the information you've scouted, and how long ago that information was scouted.
I'm not saying its not solvable, but I don't think current alphaGO could. It will be exciting to see how it gets solved, but its a harder problem than GO.
Disclaimer: I'll add *IMO, since I am not an expert by any means
Let's have the bot use all the APM that it can use. Handicapping it defeats the purpose of the challenge of a bot beating a human player. It's like reducing the bot's ability so that humans can stand a chance, which is the same as admitting that the bot already won.
But nobody doubts whether an AI can more flawlessly blink stalkers in a 50 stalker army than humans can. Or whether they can more flawlessly split marines vs banelings or whether they can more flawlessly spread lings vs splash.. Whether they can out-strategize top players though? That's way less of a given, because it's so much more of a fluid game of interactions than the case is for go or chess, where it's more of a mathematical mapping out of possible scenarios.
I mean, I know that go has too many moves for the AI to calculate all the possible moves, but like, there's an insurmountably larger amount of possible bw positions.
On May 27 2017 10:02 polgas wrote: Let's have the bot use all the APM that it can use. Handicapping it defeats the purpose of the challenge of a bot beating a human player. It's like reducing the bot's ability so that humans can stand a chance, which is the same as admitting that the bot already won.
This is a game where mechanics matter, so it's only fair to bring the AI down to humans' physical level. Otherwise, the whole thing is pointless and the bot did basically win already. Look at these videos from 6 years ago (probably already posted in this thread, but still):
I would like to see a longer series (bo7 minimum) between the most sophisticated bot and a top player. I think in macro games bots may be able to just power through with immaculate macro, but I feel like high level players could just upset via finding weak spots with cheese builds. I am just not sure if there will be enough prolonged interest in making bots so sophisticated that they can address any funky cheese or early pressure build.
Just as I would not want to limit the bot's APM, I also would not limit the human to just a standard build. Let the human player bring all the cheese builds he can think of. Fake out the bot or any other tricks. This is my idea of a true test of this challenge.
Yeah, 19x19 go had 2.08168199382×10^170 positions. If you cut a map into 20x20 grid, even with all possible unit combinations moving throughout this 20x20 grid are you gonna even get to that many positions? I mean, it stands to reason some sub-system would either evolve or be designed to handle micro situations in the small scale.
On May 27 2017 11:23 polgas wrote: Just as I would not want to limit the bot's APM, I also would not limit the human to just a standard build. Let the human player bring all the cheese builds he can think of. Fake out the bot or any other tricks. This is my idea of a true test of this challenge.
How do you beat the 100% perfectly executed blink stalker rush?
That humans are not going to be limited to a standard build is a given. Once again, there's no question whether an AI can execute better. With 0 limits to AI execution, it's just a matter of designing a safe build order that lets the computer get a big army and then micro completely flawlessly.
On May 27 2017 11:23 polgas wrote: Just as I would not want to limit the bot's APM, I also would not limit the human to just a standard build. Let the human player bring all the cheese builds he can think of. Fake out the bot or any other tricks. This is my idea of a true test of this challenge.
Not really. It's like challenging superman to a test of speed or strength. We all know superman will win. The question here is if you can beat him in a game of chess and rock-paper-scissors.
If humans can't beat the AI's perfect micro, with any strategy, then my conclusion is AI beats humans in Starcraft. If you want to feed limiting parameters to the AI, then you're just giving humans a crutch for this challenge.
Hmm very interesting. I have a little experience with implementing low-level machine learning algorithms. The thing is right now I don't imagine hardware is capable of gathering and crunching enough data to generate a net that could play Brood War. It's absolutely possible that a net could be constructed given good enough hardware and good enough input data, so it's likely a matter of time until this happens.
But the thing is, who cares? I mean we care right now, but who would really care in the future? People like the human element, people like humans competing. Take for example speed running. All the novelty, the genuine interest, all the hype comes from real life flesh and blood energy. We all love to see a hero succeed, and with speedrunning we all love seeing the runner make a great accomplishment. It's in our nature. The robots have their time and place, but it can never supersede genuine human competition, or replace it, or really even compete with it. As a collective we humans like each other, and that's not going to change.
On May 27 2017 13:39 polgas wrote: If humans can't beat the AI's perfect micro, with any strategy, then my conclusion is AI beats humans in Starcraft. If you want to feed limiting parameters to the AI, then you're just giving humans a crutch for this challenge.
? I'm not sure if you are trolling or genuinely not understanding that Deepmind want to tackle AI problems, and in this case you need to cap the mechanical part to human level for the strategy to even matter. Because you know, Blizzard could program the game so that whenever you play against their AI, the AI wins at the start of the game. See how pointless this is?
edit: @CecilSunkure: the point of this is not for AI to compete with us on a regular basis, but to successfully handle super complex problems such as playing Starcraft at a high level. They hope that if we can make AI that do that, we could use them in real life domains such as medicine, economy or whatever.
WAR BOXER!!! lol tbh though, it would be hilariously one-sided for the AI side. For example, in close and semi close 2p maps, with perfect micro, ai can just rush 2 marines + 10 scvs.
On May 27 2017 13:39 polgas wrote: If humans can't beat the AI's perfect micro, with any strategy, then my conclusion is AI beats humans in Starcraft. If you want to feed limiting parameters to the AI, then you're just giving humans a crutch for this challenge.
Exactly, as long as 15000 apm micro is in the game, the AI is allowed to use it. I'd rather add cooldown for everything, limiting the AI and not that much affecting the human.
It's also interesting to play as a human, by the way. I don't think you can play terran cause everything that is based on multitask is greatly weakened (I suppose the instant a medivac comes into line of sight the bot can see it and react, so you can never really overwhelm him with your multitask). You need to create a set of strategies that cause you to be in a winning position immediately after they're revealed to the opponent.
On May 27 2017 10:33 Liquid`Drone wrote: But nobody doubts whether an AI can more flawlessly blink stalkers in a 50 stalker army than humans can. Or whether they can more flawlessly split marines vs banelings or whether they can more flawlessly spread lings vs splash.. Whether they can out-strategize top players though? That's way less of a given, because it's so much more of a fluid game of interactions than the case is for go or chess, where it's more of a mathematical mapping out of possible scenarios.
I mean, I know that go has too many moves for the AI to calculate all the possible moves, but like, there's an insurmountably larger amount of possible bw positions.
Im really surprised by these kinds of comments. Look, im not a programmer or IT guy, but I have done enough math and programming course in uni to know that in essence it will merely be a series of IF>THENs. It doesnt matter if it takes 100 or 1,000,000 routines and subroutines. People who say this miss the fact that human brain/consciousness/decision making process is nothing more than an elaborate almost infinite number of IF>THENs based on experience and risk taking. AI could do that way faster and with way more calculations. People are trying to romantices consciousness as if it were a magical entity.
As a more rigid example,consider this (let us use bw since Boxer is a bw player):
1. AI (zerg) vs. Boxer (Matchpoint) 2. AI has multiple BOs in database for reference (let us use 3 for example: 12CC, 10Rax, proxy Rax cheese) 3. AI sends scout at normal scout timing 4.1 AI drone in Boxer base - IF worker/building count = 12CC THEN anti-12CC BO 4.2 AI drone in Boxer base - IF worker/building count = 10RAX THEN anti-10RAX BO 4.3 AI drone in Boxer base - IF worker/building count = proxy Rax cheese THEN anti-prc BO. scout ideal proxy rax area.
This is an immense oversimplification, but the point is, if it even remotely possible for humans to imagine and do, the AI can do so with way better efficiency and accuracy.
On May 27 2017 10:33 Liquid`Drone wrote: But nobody doubts whether an AI can more flawlessly blink stalkers in a 50 stalker army than humans can. Or whether they can more flawlessly split marines vs banelings or whether they can more flawlessly spread lings vs splash.. Whether they can out-strategize top players though? That's way less of a given, because it's so much more of a fluid game of interactions than the case is for go or chess, where it's more of a mathematical mapping out of possible scenarios.
I mean, I know that go has too many moves for the AI to calculate all the possible moves, but like, there's an insurmountably larger amount of possible bw positions.
Im really surprised by these kinds of comments. Look, im not a programmer or IT guy, but I have done enough math and programming course in uni to know that in essence it will merely be a series of IF>THENs. It doesnt matter if it takes 100 or 1,000,000 routines and subroutines. People who say this miss the fact that human brain/consciousness/decision making process is nothing more than an elaborate almost infinite number of IF>THENs based on experience and risk taking. AI could do that way faster and with way more calculations. People are trying to romantices consciousness as if it were a magical entity.
As a more rigid example,consider this (let us use bw since Boxer is a bw player):
1. AI (zerg) vs. Boxer (Matchpoint) 2. AI has multiple BOs in database for reference (let us use 3 for example: 12CC, 10Rax, proxy Rax cheese) 3. AI sends scout at normal scout timing 4.1 AI drone in Boxer base - IF worker/building count = 12CC THEN anti-12CC BO 4.2 AI drone in Boxer base - IF worker/building count = 10RAX THEN anti-10RAX BO 4.3 AI drone in Boxer base - IF worker/building count = proxy Rax cheese THEN anti-prc BO. scout ideal proxy rax area.
This is an immense oversimplification, but the point is, if it even remotely possible for humans to imagine and do, the AI can do so with way better efficiency and accuracy.
maybe it can choose adequate build orders based on scouting information but the amount of strategic thinking you have to do in a reactive macro game is much more complicated. Recognizing when it's the best time to attack, where to attack, when it's better to go for harass, finding the best positions for a fight - that are all important decisions that aren't easy for an AI to learn.
EPT
