|
On January 25 2019 05:55 Nakajin wrote:Show nested quote +On January 25 2019 05:50 Brutaxilos wrote:On January 25 2019 05:48 Charoisaur wrote:
Incredibly impressive. If it can beat one pro it can beat every pro with a bit more time.
The thing that interests me more how it will perform in the other matchups since those are I think more complex and harder to figure out (especially the non-mirrors).
I also think they should limit the max apm since it's not human-like to reach 800 peaks unless its rapidfire/holding down the drone key. Honestly, even if other matchups are like 10-100 times more complex, that's pretty easy for a computer to scale up to. So I doubt it'll have any issues with that. Maybe varieties are gonna be his weakness? Dunno if it need to start up from scratch when you change the maps/races, they probably talk about it but I don't recall. A single engine may have a hard time learning to play all of the match up and maps.
It's not going to make a difference. There will probably be different AIs for different matchups. And in the end you can just conjoin them together in the beginning of a decision tree to make one big AI (if PvZ use AI2342423424 if TvP use AI234324323424 etc.). There's no reason an AI should need to know a PvZ metagame to play TvP for example.
|
I think once all 6 possibles matchups will be properly handled by multiple high-level A.Is, it might be a useful tool to work on balancing each race. If e.g., the best terran A.I always win against the best protoss or zerg A.I with a certain unit composition, balance changes could be applied and tested again with the A.I. Imo, balance should always be done for top pro players, but any skill level should always be balanced which could be done by lower skill level A.I.
|
Started watching. 30 sec into the first replay it shows that the AI isnt worker-pairing. Oh well i wasnt expecting much but im still disappointed - not gonna waste my time watching that garbage
User was warned for this post
|
Maybe Alphastar didn't know basetrade was a solid option. Maybe it hadn't seen/calculated Mana's full armysize.
|
France12795 Posts
The replay vs Mana (exhibition game) shows that the AI is dumb as f*** (and by that I mean not adaptable) on a strategic and tactical level, but very good at maximizing income & army size in its early game.
Super cool to see it able to play vs real players, but each agent seems very narrow minded.
Their article https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ is really detailed if you wanna see the technical details of it
|
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai
|
Czech Republic12129 Posts
MaNa, TLO. I love you guys, but I was cheering for the AI. It looked humanly and I was really amazed. I cannot wait for the next iteration.
|
On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai
Wait what? If this is not an actual SC2 AI, what is?
|
On January 25 2019 06:26 Brutaxilos wrote:Show nested quote +On January 25 2019 05:55 Nakajin wrote:On January 25 2019 05:50 Brutaxilos wrote:On January 25 2019 05:48 Charoisaur wrote:
Incredibly impressive. If it can beat one pro it can beat every pro with a bit more time.
The thing that interests me more how it will perform in the other matchups since those are I think more complex and harder to figure out (especially the non-mirrors).
I also think they should limit the max apm since it's not human-like to reach 800 peaks unless its rapidfire/holding down the drone key. Honestly, even if other matchups are like 10-100 times more complex, that's pretty easy for a computer to scale up to. So I doubt it'll have any issues with that. Maybe varieties are gonna be his weakness? Dunno if it need to start up from scratch when you change the maps/races, they probably talk about it but I don't recall. A single engine may have a hard time learning to play all of the match up and maps. It's not going to make a difference. There will probably be different AIs for different matchups. And in the end you can just conjoin them together in the beginning of a decision tree to make one big AI (if PvZ use AI2342423424 if TvP use AI234324323424 etc.). There's no reason an AI should need to know a PvZ metagame to play TvP for example.
That is not their goal though. Their ultimate goal is to be able to train a model that can make general decisions regardless of the situation. So in the future with just minor tweaks they could take this model and apply it to different domains.
If they end up just having a giant list of various AI's which each specialize in certain things then they aren't making the progress they desire.
|
Czech Republic12129 Posts
On January 25 2019 06:58 cha0 wrote:Show nested quote +On January 25 2019 06:26 Brutaxilos wrote:On January 25 2019 05:55 Nakajin wrote:On January 25 2019 05:50 Brutaxilos wrote:On January 25 2019 05:48 Charoisaur wrote:
Incredibly impressive. If it can beat one pro it can beat every pro with a bit more time.
The thing that interests me more how it will perform in the other matchups since those are I think more complex and harder to figure out (especially the non-mirrors).
I also think they should limit the max apm since it's not human-like to reach 800 peaks unless its rapidfire/holding down the drone key. Honestly, even if other matchups are like 10-100 times more complex, that's pretty easy for a computer to scale up to. So I doubt it'll have any issues with that. Maybe varieties are gonna be his weakness? Dunno if it need to start up from scratch when you change the maps/races, they probably talk about it but I don't recall. A single engine may have a hard time learning to play all of the match up and maps. It's not going to make a difference. There will probably be different AIs for different matchups. And in the end you can just conjoin them together in the beginning of a decision tree to make one big AI (if PvZ use AI2342423424 if TvP use AI234324323424 etc.). There's no reason an AI should need to know a PvZ metagame to play TvP for example. That is not their goal though. Their ultimate goal is to be able to train a model that can make general decisions regardless of the situation. So in the future with just minor tweaks they could take this model and apply it to different domains. If they end up just having a giant list of various AI's which each specialize in certain things then they aren't making the progress they desire.
You will train specific neural networks and then you will create one that covers it all. This actually simulates the human brain(when you play PvZ you load your PvZ persona, when you play ZvP you load your ZvP persona) and it will create one big AI. Then they can actually train this whole new network. If it's feasable - I have no clue, I am way too out of touch with the technology, several years I haven't touched ML which is a shit ton of time 
|
On January 25 2019 06:55 _fool wrote:Show nested quote +On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is?
The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be.
|
Czech Republic12129 Posts
On January 25 2019 07:04 Torvaltz wrote:Show nested quote +On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be.
It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive.
|
On January 25 2019 07:12 deacon.frost wrote:Show nested quote +On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive.
Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on
|
On January 25 2019 07:17 GoloSC2 wrote:Show nested quote +On January 25 2019 07:12 deacon.frost wrote:On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive. Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on
I understand what they are going for but I don't think they're as close to it as you guys seem to, thats pretty much my take. limit it more and then Id be impressed
it works and it makes choices but they arent better than pros nor are they an apocalyptic sign to me
|
Czech Republic12129 Posts
On January 25 2019 07:17 GoloSC2 wrote:Show nested quote +On January 25 2019 07:12 deacon.frost wrote:On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive. Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on
Actually it's hard to imagine/understand. If I didn't write some neural networks at university I would be on the pesimistic side as well. But since I know this I am on the exact other side. I am so fucking amazed. I hope those people can read this, because I am hyped. They did a really good job in a game I love. Well done.
|
On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai
I mean even in that game it played really well, except that it obviously had some kind of glitch or hole when it comes to dealing with drops. It was ahead for most of the game, up until the incident.
It's absolutely a proof of concept.
Although, yeah, it's probably a little unfairly good at micro lol.
|
France12795 Posts
On January 25 2019 07:20 Torvaltz wrote:Show nested quote +On January 25 2019 07:17 GoloSC2 wrote:On January 25 2019 07:12 deacon.frost wrote:On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive. Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on I understand what they are going for but I don't think they're as close to it as you guys seem to, thats pretty much my take. limit it more and then Id be impressed
It's actually extremely impressive.
Basically, in the only version (camera interface) that is relatively close to fair (apm and stuff will have to be examined because it seems the AI exploits it a bit too much), they managed to have an AI not to be dumb for like 9 minutes, before it showed signs of being an AI.
If it were a Starcraft Turing test, most people would have known it was an AI at the warp prism incident. Before that, it looked legit like an human.
The principle of the anonymous tournament from NoRegret & Scarlett (not sure who exactly are the organizers) with AlphaStar would be interesting, with humans and AlphaStar agents anonymously mixed in
|
On January 25 2019 07:04 Torvaltz wrote:Show nested quote +On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be.
A common mistake with AI sceptics is that they opt to use a definition of intelligence that requires one or more properties that they only acknowledge in humans. In that case it's no wonder that they never find intelligence anywhere else. What you're doing is very similar, if not just that.
Of course it made decisions. When and where to expand. What to spend its money on. When to attack. Remember that there are no procedural rules programmed into it. So every click that you saw was a decision.
And making decisions based on observations is what I would call intelligence (definition up for refinement).
You act like the AI was given 20 stalkers and then it micro'd its way out of the game. But it's far more complex than that. It found a way to build those 20 stalkers while harassing. And it found out by itself.
|
United States33360 Posts
On January 25 2019 07:20 Torvaltz wrote:Show nested quote +On January 25 2019 07:17 GoloSC2 wrote:On January 25 2019 07:12 deacon.frost wrote:On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive. Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on I understand what they are going for but I don't think they're as close to it as you guys seem to, thats pretty much my take. limit it more and then Id be impressed it works and it makes choices but they arent better than pros nor are they an apocalyptic sign to me
did you just go "I don't totally understand the facts but this is what I WANT to believe"???????????
weighing a ban ;o
|
On January 25 2019 07:30 Waxangel wrote:Show nested quote +On January 25 2019 07:20 Torvaltz wrote:On January 25 2019 07:17 GoloSC2 wrote:On January 25 2019 07:12 deacon.frost wrote:On January 25 2019 07:04 Torvaltz wrote:On January 25 2019 06:55 _fool wrote:On January 25 2019 06:44 Torvaltz wrote:
so it seems to me like they ended up with a maphacking microbot that got wrecked when they turned that stuff off?
i dont understand the amazement but ok, to me this looks like were still far away from an actual sc2 ai Wait what? If this is not an actual SC2 AI, what is? The type that wins through its intelligence and decision making, not through impossible micro and perfect control through multiple screens. We already know micro bots are great, to me the decision making part is where the real AI breakthrough could be. It had a decision making. Not perfect, but considering this can be called first playable interation it was still impressive. Also it had neither maphack nor was it like the microbots we know. It seems you didn't really try to understand was what going on I understand what they are going for but I don't think they're as close to it as you guys seem to, thats pretty much my take. limit it more and then Id be impressed it works and it makes choices but they arent better than pros nor are they an apocalyptic sign to me did you just go "I don't totally understand the facts but this is what I WANT to believe"??????????? weighing a ban ;o
when did i say that phrase my dude?
|
|
|
|
|
|
EPT
