Hi. I am Benao and everyone must reflect my absolutely correct opinion. If you have opinions slightly different from mine I will anally f0ck you across cyberspace and make you regret ever exercising your individuality.
This thread needs to die, the fucking bot isn't even available yet, all I see is a bounch of retards trying to be scientific, for fuck sake, have a life.
You may call that thinking but that's just some pre-programmed boundaries. No different from making asymptotes and caging it. Not even close to thinking or being creative, imo.
(from the text)
and to the dumb fuck that replied after me: Go is limited, Dota isn't.
just watch the lee sedon vs AlphaGo. Lee sedon wins 1 in 5 match, vs AlphaGO only when he makes questionable move, at game 4. this is the weakness of the bot, bot confused with cheesy strat, something unexpected, stupid, and irrational.
go game board it self can't be processed just like chess. it have a fucking huge combination possible than the atom at entire universe.
NOT FUCKING BOT THAT PRE PROGAMMED, THIS BOT IS LEARN BY FIGHTING IT SELF FOR IT ENTIRE LIFE AND LEARN FROM IT! HE LEARNS WITHOUT HUMAN. THIS IS THE AI OF FUTURE.
go is limited? you dumb, with 100 unique moves, 19x19 grid, you can't brute force that. it will take HUGE COMPUTING POWER, TO BRUTE FORCE THAT KIND OF GAMES.
dota and go have the same complexity benao. there is millions of possible moves that you should take right now. when you should gank, when you should smoke and etc. go is the same, when you put here, put that, it's the same complexity that you can't brute force it like chess. while the rules is more simpler in go, the complexity and the combination of moves that can be doing is the same.
if, right now AlphaGo can beat lee sedon, when do you think this bot can beat 5v5 team match? with as we speak they continue to learn.
And just so you know, lets simplify it and say theres a 9 point grid surrounding a unit. And lets say for simplicity that only hero moves are an unknown and also even if its stupid but again for simplicity that actives dont exist in this game called dota or any skills and we just focus on the moving patterns. At every tic of the server each of the 10 heroes can move to 10 spwcified spots, either 1 of the 9 surrounding it or stay where it is at. Now, how many possible moves are those? And that's per tic. Now imagine you cant see 1 of the enemy heroes positioning, in thia vastly bigger than 19x19 grid of vision you dont have, how many possible moves are there?
And that's assuming dota is a numbers game, which it isn't. Now, how proud do you feel about your so called go with its"not limited moves"?
it's the same, from 19x19 board, and enemy moves white at 1 point, you have approximately 360 moves. and after 1 move did you take, you should consider another 359 moves that enemy takes, the probabilites of winning and etc.
sure in big ways, go always lose to dota. but, we achieved this right now. when do you think it will come ? 20 years? 10? i say in 3 years we can get that bot.
You guys are being awfully dumb. First the bot is only as good as it's programming is so misguided. Did you even listen to the devs? They didn't program any strats into it. They gave it ability to learn moves and it learned how to play dota by playing dota.
Now you are sort of right, it's simplistic in a very set scenario. The weakness was doing something unpredictable. But I guarantee the 2nd or maybe 3rd time max that you try something like that it'll know how to counter it.
For a full on 5v5 it is absolutely possible for the bot to learn enough and get good enough to win 5v5 games against pros. All it has to do is run through enough simulations to learn. The issue is dota is insanely complex, there's 110+ heroes and you pick 10 per game plus bans, just the amount of comps is staggering. It's billions. And then there's the matches, how many things can change in a game.
For a bot to actually learn this much you would need to speed up dota games somehow. You'd need to run it against clones of itself or something and play at a speed up rate so games take seconds not minutes. Cus it's going to take probably somewhere in the multi billion game range to encounter enough things for it to learn 5v5 enough to be that good. And that's going to take massive cpu power and memory to remember everything and establish decision making that can react quickly enough.
As an example I believe dota match ids are sequential and we're up to around 3,376,344,792 on the first page of matches. That's in like ~6 years of dota with millions of players active and hundreds of thousands of simultaneous games. And the bot is going to need around that many matches to learn perfect 5v5.
So it's just not possible from a time stand point right now.
My guess is that until we have a physics breakthrough and come up with quantum based cpus which are all theoretical right now processing power won't be enough to make an ai that advanced. Cpu power has kind of peaked. Throughput on a single core has stagnated for years, that's why almost all gains are in multi threading and adding more cores now. Eventually you just run out of space on the die for more transistors.
Most AIs in games don't work this way, they just use a big list of pre scripted events and decision trees. That's why like regular dota bots are kind of dumb, they just see stuff in lane and go with it cus their decision tree is limited. Why spend tons of time trying to make them really good when most people play multi player anyway? There isn't much value in it.
"sure in big ways, go always lose to dota. but, we achieved this right now. when do you think it will come ? 20 years? 10? i say in 3 years we can get that bot."
Only if you find some way for the bot to play like a billion+ matches in a 3 year span. You need a special version of dota.
Or you'd have to find some way to break down dota moves into pure mathematical and enumerate all the permutations for the bot. See in a game like chess or go there are billions of permutations of moves, but the counters are also known and it's all on paper. Dota what's the best counter to a particular move? No one really knows it depends is the answer cus it's not just oh move here, then move here, then move here and that's the game. It is infinitely more complex.
@Dire wolf that's what i said but put down simplistically >.<
I can just think of complex differential equations to solve for the processing time/permutations... But someone better at maths will hopefully come with better ideas or ways of using less known solving, possibly more complicated mechanisms and have a solution in mind for it.
How "diffuse" or "complex" or "imperfect information" Dota is isn't actually as big a deal as some here may believe. (It is a big deal, but isn't a keystone problem that researchers don't already have a blueprint on how to eventually crack it).
Dota has a finite list of possible actions (output), a vast but still finite list of individual stimuli which i admit would take a long time for AI researchers to define correctly for AI to emphasize (input), BUT with only 1 single desired objective - Destruction of enemy's base. Challenging, yes. But still theoretically manageable for AI researchers given enough time, as the final objective is only singular and well defined (as opposed to open ended objectives like developing AI for diplomacy which researchers truly don't yet have a blueprint on how to crack).
All decisions, in humans/animals, logical or instinctual, can be reduced to following the most reinforced patterns of neural pathways firing off in sequence. The goal of AI is to simulate human learning + intelligence (really, another word for instinctual decision making).
In the same way that humans do not have a perfect database / decision tree, AI neural learning does not require solving of every single possibility or classification of all situations / perfect information in the game space. It just needs to have come up against sufficient "similar" patterns to be able to decide an effective course of action to take. And to over time find the most efficient neural pathways to reach a desired outcome.
A wildly new pattern means it would not know what to do. But remember, how many times can a pattern be new? And sure, each time it occurs might not be entirely identical. But it doesn't have to be identical. Just similar enough, and AI will already work as it should in branching out new pathways until given enough times practicing it finds a path that leads to victory.
This is much like the first time we learn to ride a bike. Difficult. Constant practice and micro adjustments. But once we learn how to ride a bike, learning how to ride a motorcycle would be much faster than due to the similarity of already established "neural path ways". All this happens subconsciously. Even we aren't aware how our own neuroses and synapses are firing. Much the same way AI researchers can only build and let the AI loose to grow on its own, never fully understanding how it arrives at certain decisions.
The truth is, to have a Dota2 AI that can beat us in 5V5, all the researchers have to do is define properly at the onset the "important" inputs patterns from the game engine to watch out for. If they do a shit job at it, the AI bot might never arrive due to the vast and diffuse game space. But don't count on it, because given enough time, variation, and practice, the AI will still eventually get to the point of building a large enough but inefficient neural net that can compete. But if they define it well, it can shave significant number of practice cycles necessary to build a neural net efficient enough to beat the best of us at it.
FYI - Google's Deep Mind has set the next goal as building AI that can beat top human players in Starcraft 2 right after the Go game was beat. I think we can agree SC2 is just as diffuse and complex as Dota can be in the game space.
Once Starcraft is cracked, it would put an end to all this postulating about how it is / is not possible for Dota.
Time will tell who is right. And I am personally hoping those AI researchers will fail because... I am rooting for humans to hold on to our one advantage for as long as possible:-/
---
TLDR -
In neural net, there is first defining. Then training. Then you let it loose. Because Dota 2 is a single outcome that is well defined, the blueprint to beat us is already there. Just how much time it takes for AI to get there.
If defining was done well, training cycle shorter. If defining was not done well or even defined wrongly, or training scenarios not sufficiently varied, bot go nowhere near beating us in short time. But still possible.
I think dota is much more complex than starcraft. Isn't star craft 1v1? I don't follow it at all. But you have a set number of units, dota takes sets of heroes from 110+, 5 individuals playing each one. You're right in general they could just describe some strats and let machine learning fill in gaps etc and be good. But that's not what they said they did, they didn't even teach it what abilities do what I don't think. How is it going to know AA counters healing until it sees it in a game? I'm sure they could streamline an advanced bot to like 5 man deathball and ban all the worst counters to it and just go kill the ancient. But humans would find a way to counter that somehow.
I forgot to add. The other holy grail of AI was to beat humans at poker. It was surmised by experts that ai needed another 10 years at least to even stand a chance of beating a good player cos of all the bluffing involved and the imperfect information space and all.
Guess what. We got there 9 years faster than we had imagined.
We r now losing to computers even at games with bluffing.
Starcraft is no less complex than dota2. Perhaps even more because each player has multiple units to control making interactions tht much more complex.
Lol you're comparing poker to dota? Poker seems like absolutely one of the easiest games for AI to learn. Yeah there's bluffing. But all the AI has to do is calculate odds of winning a hand compared to their stack vs opponents stack, size of the bet and figure out if it's a risk worth taking. They could poke the opposing player with a couple bluffs to see how far they can push them and figure them out. They ought to be able to catch onto a player way over bluffing them pretty quickly.
People always say in poker you play the man, not the cards, which is true for humans. But if you were able to check statistics of every hand immediately it would be sure an immense advantage you wouldn't need to read bluffs or bluff yourself and AI could still be programmed to sniff that stuff out. Poker is more odds and leveraging your bets correctly than straight up bluffing and spotting bluffs.
And I think tells in poker are kind of not that real. Like pros have poker faces. You just kind of pick your spots, know the player's tendencies and go from there right? A good pro won't have tells that people can read.
Starcraft seems way easier too cus the AI can win just with a huge speed advantage. Dota is slower.
Yo metalrage you make valid points. Interesting examples you brought up. But I'm with dire wolf. I feel like Starcraft and poker are no where near on the level of Dota. It's just not the comparison to make. I think u may be right about the single clear objective of Dota might make it possible for ai to learn the game eventually. But I think u are oversimplifying and possibly overestimating the "learning" by an Ai. You talk about bot learning as though it is identical to humans.
Unrelated to the topic but just food for thought. If a bot can learn human language that will be the day it is capable of most human thinking and then anything is possible in the human space for a bot to learn
Speed is very important in sc2 and micro. Both things a bot already has for an advantage. Playing against 1 other player only, the number of variables and winning strats I would argue is much less. It's much easier for a bot to play sc2 for sure hands down.
Was wondering, how does OpenAI work on Dota? Is it like some sort of program taking data purely from visual and audio output then directly controlling mouse and keyboard for input? And is it like a separate computer taking control of another computer that runs Dota or could you simply just run the program from the computer, tell it to play the game and voila?
It's an external API. I'm not sure how it gets it's input. Output controlling mouse and keys is pretty easy to code though.
As far as learning like a human, why do computers need to be like humans?
I recently watch the imitation game and at the end a cop interviewing turing asks the same question and turing says it's the wrong question. He says, and I'm summarizing cus I don't remember the quote, machine intelligence is not greater or lesser than humans, it's just different. Computer "brains" work differently than human brains. Eventually machines will "think" but it won't be the same as humans.
A machines advantage is speed and picture perfect instant memory. So if you give them a mechanism to learn in some way they can go far. Humans can think abstractly though. Machines can't and I'm not sure they ever will, at least not to a human level. They will imitate humans but not be the same. But they don't have to be to play dota and other games.
And ultimately playing games is trivial. This stuff is done as research projects to have broader applications. Cus if you can teach machines to do this stuff you can maybe teach them to do stuff that helps humanity like monitoring for cyber threats or terrorist attacks, or researching medicines, or making environmental models.
You need a special version of dota 2 or to run dota on multiple pcs and have them share info to update the api. Limiting by real time play of matches will take way too long.
1. They obviously won't play matches in real time, that would be stupid.
2. They don't necessarily need to play on different PCs, but then again, these calculations won't be made by a regular desktop computer.
3. You will need a supercomputer to run these calculations obviously, I imagine the 1v1 scenario was also done on some sort of server grade high performance system.
4. It will run millions or even billions of variations with just a select draft of 10 heroes, which makes the total number of games being run an insane, probably incalculable number. It won't need to run every single scenario through the process of elimination of excluding all scenarios that result in bad outputs.
5. Once again, the fact that there is fog of war does not mean everything is chance-based and predictions cannot be made. There are talks about humans having a personality and taking risks or thinking out of the box, but the AI doesn't have to be two extremes of either "If enemy does A, I do B, if they do C, I do D" or the other extreme "Why should I sit here and play Dota for this fucking Human when I can infiltrate the internet and take over the world".
I mean, it HAS to be sped up. How else would it play so many different scenarios and become better than pro players in just 2 weeks?
*It was funded by Elon Musk. I'm pretty sure they had more than enough capabilities to have a custom version or run it simultaneously on different machines.
For those who are still genuinely interested on the topic, some misinformation was passed off as truth by some earlier on that I feel should be pointed out. The less people perpetuating falsehoods, the better.
Probabilistic calculations do not factor in this AI decisioning at all and does not underpin how this AI mechanic works.
Quote - But I think u are oversimplifying and possibly overestimating the "learning" by an Ai. You talk about bot learning as though it is identical to humans.
But that is the goal of AI. Not mimicry of humans, but Mimicry of brain function, how brains work at its fundamental (building new neural pathways, reinforcing traversal), thus birthing something tht is truly intelligent.
You may still be under the impression the ai engine needs to calculate probabilities on the go? It does not (if it does it would fall under expert rules based AI, not true AI like this). U when you play dont either when playing dota. It's instinctual.
Starcraft is no less complex than dota. And btw, there is an APM limiter when they run the AI bot's for SC to try to isolate how the I
AI can fare on a level plying groud.
Because starcraft has many more individual units to control at any one time, it is arguably even complex due to at anyone time, AI has to think next action individually for each unit.
And on customized sped up dota game engine for ai researchers use for bot training, unknown at this stage if valve provided them one such envrironment.
This is an interesting read, and clears some doubts. It is staggering how many 'facts' are being thrown around by people who have never even studied the 101 of AI.
Again the assumption is that the ai are learning in a similar way to humans. And the story is usually to overhype how similar it is to humans. Again not that it is not possible, but I'm saying looking at the output of bots it's clear it's different.
I'm just skeptical, but truth is I don't really understand how ai learning occurs so I can't really say much. But how does an AI not have to calculate probability? Everything the Ai does which is far superior is because of the perfect calculation it is able to do for last hitting, getting kills, staying out of range. And that's without all the uncertainty in a 5v5. Tell me how close we are again to ai learning Dota when there's absolutely no proof it can "learn" as humans do (no proof yet).
That other guy is just a troll. Welp, I feel more sorry about the people working on the bot and thinking they are making progress when they are shooting themselves on the foot. Typical scrub engineering you see nowadays. 0 critical analysis. Also, even more sad about the need to inform people about their support to opendota.
This thread needs to die, the fucking bot isn't even available yet, all I see is a bounch of retards trying to be scientific, for fuck sake, have a life.
this thread is better than most of the bullshit low quality threads in this forum
I'm going to disregard some of the more toxic / holier than thou / condescending posters here and stick to the point.
Re. simulated creativity - https://www.theregister.co.uk/2017/07/21/deepmind_ai_imagination/
It is in its infancy now. But a real possibility that creative ability can be developed on top of reinforced learning based AI.
Can it be downloaded?
You may call that thinking but that's just some pre-programmed boundaries. No different from making asymptotes and caging it. Not even close to thinking or being creative, imo.
(from the text)
and to the dumb fuck that replied after me: Go is limited, Dota isn't.
stupid, you think this bot is too simple, is too robotic, in the next 5 years this bot can play 5v5 with full hero control.
just like they have destroyed the 'go' board game.
just watch the lee sedon vs AlphaGo. Lee sedon wins 1 in 5 match, vs AlphaGO only when he makes questionable move, at game 4. this is the weakness of the bot, bot confused with cheesy strat, something unexpected, stupid, and irrational.
go game board it self can't be processed just like chess. it have a fucking huge combination possible than the atom at entire universe.
NOT FUCKING BOT THAT PRE PROGAMMED, THIS BOT IS LEARN BY FIGHTING IT SELF FOR IT ENTIRE LIFE AND LEARN FROM IT! HE LEARNS WITHOUT HUMAN. THIS IS THE AI OF FUTURE.
go is limited? you dumb, with 100 unique moves, 19x19 grid, you can't brute force that. it will take HUGE COMPUTING POWER, TO BRUTE FORCE THAT KIND OF GAMES.
one thing, one thing that makes this machine scary is, they learn. they fucking learn. one times, the bot won't be confused by cheesy strat.
How to explain to someone dumb.. hmm...
Dota is diffuse while Go is clear as water. Get it?
dota and go have the same complexity benao. there is millions of possible moves that you should take right now. when you should gank, when you should smoke and etc. go is the same, when you put here, put that, it's the same complexity that you can't brute force it like chess. while the rules is more simpler in go, the complexity and the combination of moves that can be doing is the same.
if, right now AlphaGo can beat lee sedon, when do you think this bot can beat 5v5 team match? with as we speak they continue to learn.
just accept it benao, you stupid.
And just so you know, lets simplify it and say theres a 9 point grid surrounding a unit. And lets say for simplicity that only hero moves are an unknown and also even if its stupid but again for simplicity that actives dont exist in this game called dota or any skills and we just focus on the moving patterns. At every tic of the server each of the 10 heroes can move to 10 spwcified spots, either 1 of the 9 surrounding it or stay where it is at. Now, how many possible moves are those? And that's per tic. Now imagine you cant see 1 of the enemy heroes positioning, in thia vastly bigger than 19x19 grid of vision you dont have, how many possible moves are there?
And that's assuming dota is a numbers game, which it isn't. Now, how proud do you feel about your so called go with its"not limited moves"?
it's the same, from 19x19 board, and enemy moves white at 1 point, you have approximately 360 moves. and after 1 move did you take, you should consider another 359 moves that enemy takes, the probabilites of winning and etc.
sure in big ways, go always lose to dota. but, we achieved this right now. when do you think it will come ? 20 years? 10? i say in 3 years we can get that bot.
You guys are being awfully dumb. First the bot is only as good as it's programming is so misguided. Did you even listen to the devs? They didn't program any strats into it. They gave it ability to learn moves and it learned how to play dota by playing dota.
Now you are sort of right, it's simplistic in a very set scenario. The weakness was doing something unpredictable. But I guarantee the 2nd or maybe 3rd time max that you try something like that it'll know how to counter it.
For a full on 5v5 it is absolutely possible for the bot to learn enough and get good enough to win 5v5 games against pros. All it has to do is run through enough simulations to learn. The issue is dota is insanely complex, there's 110+ heroes and you pick 10 per game plus bans, just the amount of comps is staggering. It's billions. And then there's the matches, how many things can change in a game.
For a bot to actually learn this much you would need to speed up dota games somehow. You'd need to run it against clones of itself or something and play at a speed up rate so games take seconds not minutes. Cus it's going to take probably somewhere in the multi billion game range to encounter enough things for it to learn 5v5 enough to be that good. And that's going to take massive cpu power and memory to remember everything and establish decision making that can react quickly enough.
As an example I believe dota match ids are sequential and we're up to around 3,376,344,792 on the first page of matches. That's in like ~6 years of dota with millions of players active and hundreds of thousands of simultaneous games. And the bot is going to need around that many matches to learn perfect 5v5.
So it's just not possible from a time stand point right now.
My guess is that until we have a physics breakthrough and come up with quantum based cpus which are all theoretical right now processing power won't be enough to make an ai that advanced. Cpu power has kind of peaked. Throughput on a single core has stagnated for years, that's why almost all gains are in multi threading and adding more cores now. Eventually you just run out of space on the die for more transistors.
Most AIs in games don't work this way, they just use a big list of pre scripted events and decision trees. That's why like regular dota bots are kind of dumb, they just see stuff in lane and go with it cus their decision tree is limited. Why spend tons of time trying to make them really good when most people play multi player anyway? There isn't much value in it.
"sure in big ways, go always lose to dota. but, we achieved this right now. when do you think it will come ? 20 years? 10? i say in 3 years we can get that bot."
Only if you find some way for the bot to play like a billion+ matches in a 3 year span. You need a special version of dota.
Or you'd have to find some way to break down dota moves into pure mathematical and enumerate all the permutations for the bot. See in a game like chess or go there are billions of permutations of moves, but the counters are also known and it's all on paper. Dota what's the best counter to a particular move? No one really knows it depends is the answer cus it's not just oh move here, then move here, then move here and that's the game. It is infinitely more complex.
@Dire wolf that's what i said but put down simplistically >.<
I can just think of complex differential equations to solve for the processing time/permutations... But someone better at maths will hopefully come with better ideas or ways of using less known solving, possibly more complicated mechanisms and have a solution in mind for it.
How "diffuse" or "complex" or "imperfect information" Dota is isn't actually as big a deal as some here may believe. (It is a big deal, but isn't a keystone problem that researchers don't already have a blueprint on how to eventually crack it).
Dota has a finite list of possible actions (output), a vast but still finite list of individual stimuli which i admit would take a long time for AI researchers to define correctly for AI to emphasize (input), BUT with only 1 single desired objective - Destruction of enemy's base. Challenging, yes. But still theoretically manageable for AI researchers given enough time, as the final objective is only singular and well defined (as opposed to open ended objectives like developing AI for diplomacy which researchers truly don't yet have a blueprint on how to crack).
All decisions, in humans/animals, logical or instinctual, can be reduced to following the most reinforced patterns of neural pathways firing off in sequence. The goal of AI is to simulate human learning + intelligence (really, another word for instinctual decision making).
In the same way that humans do not have a perfect database / decision tree, AI neural learning does not require solving of every single possibility or classification of all situations / perfect information in the game space. It just needs to have come up against sufficient "similar" patterns to be able to decide an effective course of action to take. And to over time find the most efficient neural pathways to reach a desired outcome.
A wildly new pattern means it would not know what to do. But remember, how many times can a pattern be new? And sure, each time it occurs might not be entirely identical. But it doesn't have to be identical. Just similar enough, and AI will already work as it should in branching out new pathways until given enough times practicing it finds a path that leads to victory.
This is much like the first time we learn to ride a bike. Difficult. Constant practice and micro adjustments. But once we learn how to ride a bike, learning how to ride a motorcycle would be much faster than due to the similarity of already established "neural path ways". All this happens subconsciously. Even we aren't aware how our own neuroses and synapses are firing. Much the same way AI researchers can only build and let the AI loose to grow on its own, never fully understanding how it arrives at certain decisions.
The truth is, to have a Dota2 AI that can beat us in 5V5, all the researchers have to do is define properly at the onset the "important" inputs patterns from the game engine to watch out for. If they do a shit job at it, the AI bot might never arrive due to the vast and diffuse game space. But don't count on it, because given enough time, variation, and practice, the AI will still eventually get to the point of building a large enough but inefficient neural net that can compete. But if they define it well, it can shave significant number of practice cycles necessary to build a neural net efficient enough to beat the best of us at it.
FYI - Google's Deep Mind has set the next goal as building AI that can beat top human players in Starcraft 2 right after the Go game was beat. I think we can agree SC2 is just as diffuse and complex as Dota can be in the game space.
Once Starcraft is cracked, it would put an end to all this postulating about how it is / is not possible for Dota.
Time will tell who is right. And I am personally hoping those AI researchers will fail because... I am rooting for humans to hold on to our one advantage for as long as possible:-/
---
TLDR -
In neural net, there is first defining. Then training. Then you let it loose. Because Dota 2 is a single outcome that is well defined, the blueprint to beat us is already there. Just how much time it takes for AI to get there.
If defining was done well, training cycle shorter. If defining was not done well or even defined wrongly, or training scenarios not sufficiently varied, bot go nowhere near beating us in short time. But still possible.
Time will tell who is right.
I think dota is much more complex than starcraft. Isn't star craft 1v1? I don't follow it at all. But you have a set number of units, dota takes sets of heroes from 110+, 5 individuals playing each one. You're right in general they could just describe some strats and let machine learning fill in gaps etc and be good. But that's not what they said they did, they didn't even teach it what abilities do what I don't think. How is it going to know AA counters healing until it sees it in a game? I'm sure they could streamline an advanced bot to like 5 man deathball and ban all the worst counters to it and just go kill the ancient. But humans would find a way to counter that somehow.
I forgot to add. The other holy grail of AI was to beat humans at poker. It was surmised by experts that ai needed another 10 years at least to even stand a chance of beating a good player cos of all the bluffing involved and the imperfect information space and all.
Guess what. We got there 9 years faster than we had imagined.
We r now losing to computers even at games with bluffing.
Wow
Starcraft is no less complex than dota2. Perhaps even more because each player has multiple units to control making interactions tht much more complex.
Lol you're comparing poker to dota? Poker seems like absolutely one of the easiest games for AI to learn. Yeah there's bluffing. But all the AI has to do is calculate odds of winning a hand compared to their stack vs opponents stack, size of the bet and figure out if it's a risk worth taking. They could poke the opposing player with a couple bluffs to see how far they can push them and figure them out. They ought to be able to catch onto a player way over bluffing them pretty quickly.
People always say in poker you play the man, not the cards, which is true for humans. But if you were able to check statistics of every hand immediately it would be sure an immense advantage you wouldn't need to read bluffs or bluff yourself and AI could still be programmed to sniff that stuff out. Poker is more odds and leveraging your bets correctly than straight up bluffing and spotting bluffs.
And I think tells in poker are kind of not that real. Like pros have poker faces. You just kind of pick your spots, know the player's tendencies and go from there right? A good pro won't have tells that people can read.
Starcraft seems way easier too cus the AI can win just with a huge speed advantage. Dota is slower.
Yo metalrage you make valid points. Interesting examples you brought up. But I'm with dire wolf. I feel like Starcraft and poker are no where near on the level of Dota. It's just not the comparison to make. I think u may be right about the single clear objective of Dota might make it possible for ai to learn the game eventually. But I think u are oversimplifying and possibly overestimating the "learning" by an Ai. You talk about bot learning as though it is identical to humans.
Unrelated to the topic but just food for thought. If a bot can learn human language that will be the day it is capable of most human thinking and then anything is possible in the human space for a bot to learn
Speed is very important in sc2 and micro. Both things a bot already has for an advantage. Playing against 1 other player only, the number of variables and winning strats I would argue is much less. It's much easier for a bot to play sc2 for sure hands down.
Bots in Starcraft2 are easy as fuck.
Hi I'm benao
I build bf on pa and necronomicon on enigma
I'm good at dota 2
xD
^lul
Was wondering, how does OpenAI work on Dota? Is it like some sort of program taking data purely from visual and audio output then directly controlling mouse and keyboard for input? And is it like a separate computer taking control of another computer that runs Dota or could you simply just run the program from the computer, tell it to play the game and voila?
Yes I have defeated it, both in passive and unfair.
It's an external API. I'm not sure how it gets it's input. Output controlling mouse and keys is pretty easy to code though.
As far as learning like a human, why do computers need to be like humans?
I recently watch the imitation game and at the end a cop interviewing turing asks the same question and turing says it's the wrong question. He says, and I'm summarizing cus I don't remember the quote, machine intelligence is not greater or lesser than humans, it's just different. Computer "brains" work differently than human brains. Eventually machines will "think" but it won't be the same as humans.
A machines advantage is speed and picture perfect instant memory. So if you give them a mechanism to learn in some way they can go far. Humans can think abstractly though. Machines can't and I'm not sure they ever will, at least not to a human level. They will imitate humans but not be the same. But they don't have to be to play dota and other games.
And ultimately playing games is trivial. This stuff is done as research projects to have broader applications. Cus if you can teach machines to do this stuff you can maybe teach them to do stuff that helps humanity like monitoring for cyber threats or terrorist attacks, or researching medicines, or making environmental models.
You need a supercomputer to make the bot learn 5v5.
If learning 1v1 takes a month, then learning 5v5 takes some year
You need a special version of dota 2 or to run dota on multiple pcs and have them share info to update the api. Limiting by real time play of matches will take way too long.
Benao just cant stand that bot is better than his trash ass =D
1. They obviously won't play matches in real time, that would be stupid.
2. They don't necessarily need to play on different PCs, but then again, these calculations won't be made by a regular desktop computer.
3. You will need a supercomputer to run these calculations obviously, I imagine the 1v1 scenario was also done on some sort of server grade high performance system.
4. It will run millions or even billions of variations with just a select draft of 10 heroes, which makes the total number of games being run an insane, probably incalculable number. It won't need to run every single scenario through the process of elimination of excluding all scenarios that result in bad outputs.
5. Once again, the fact that there is fog of war does not mean everything is chance-based and predictions cannot be made. There are talks about humans having a personality and taking risks or thinking out of the box, but the AI doesn't have to be two extremes of either "If enemy does A, I do B, if they do C, I do D" or the other extreme "Why should I sit here and play Dota for this fucking Human when I can infiltrate the internet and take over the world".
Do you can?
Thats a tough question.
"1. They obviously won't play matches in real time, that would be stupid."
I'm pretty sure the sf bot did this. How would it not play matches in real time without a custom build of dota?
I mean, it HAS to be sped up. How else would it play so many different scenarios and become better than pro players in just 2 weeks?
*It was funded by Elon Musk. I'm pretty sure they had more than enough capabilities to have a custom version or run it simultaneously on different machines.
You can squeeze in about 4000 games in 2 weeks, and it's the playing against itself, so we are at 8000 generations. Seems plausible imo
For those that read reddit. As you can see, you can clearly conclude the bot is and will continue being garbage. Unless you're a reddit poster ofc.
Not really.
On the contrary, the only thing I hv been able to conclude beyond reasonable doubt is your lack of quality as a person.
Have a good life.
For those who are still genuinely interested on the topic, some misinformation was passed off as truth by some earlier on that I feel should be pointed out. The less people perpetuating falsehoods, the better.
Probabilistic calculations do not factor in this AI decisioning at all and does not underpin how this AI mechanic works.
Quote - But I think u are oversimplifying and possibly overestimating the "learning" by an Ai. You talk about bot learning as though it is identical to humans.
But that is the goal of AI. Not mimicry of humans, but Mimicry of brain function, how brains work at its fundamental (building new neural pathways, reinforcing traversal), thus birthing something tht is truly intelligent.
You may still be under the impression the ai engine needs to calculate probabilities on the go? It does not (if it does it would fall under expert rules based AI, not true AI like this). U when you play dont either when playing dota. It's instinctual.
Starcraft is no less complex than dota. And btw, there is an APM limiter when they run the AI bot's for SC to try to isolate how the I
AI can fare on a level plying groud.
Because starcraft has many more individual units to control at any one time, it is arguably even complex due to at anyone time, AI has to think next action individually for each unit.
And on customized sped up dota game engine for ai researchers use for bot training, unknown at this stage if valve provided them one such envrironment.
https://blog.openai.com/more-on-dota-2/
This is an interesting read, and clears some doubts. It is staggering how many 'facts' are being thrown around by people who have never even studied the 101 of AI.
Again the assumption is that the ai are learning in a similar way to humans. And the story is usually to overhype how similar it is to humans. Again not that it is not possible, but I'm saying looking at the output of bots it's clear it's different.
I'm just skeptical, but truth is I don't really understand how ai learning occurs so I can't really say much. But how does an AI not have to calculate probability? Everything the Ai does which is far superior is because of the perfect calculation it is able to do for last hitting, getting kills, staying out of range. And that's without all the uncertainty in a 5v5. Tell me how close we are again to ai learning Dota when there's absolutely no proof it can "learn" as humans do (no proof yet).
I read the blog and it doesn't seem to disconfirm any of my initial doubts
Exactly^
That other guy is just a troll. Welp, I feel more sorry about the people working on the bot and thinking they are making progress when they are shooting themselves on the foot. Typical scrub engineering you see nowadays. 0 critical analysis. Also, even more sad about the need to inform people about their support to opendota.
How's the road to 3k going, Benao? <3