Upvote Upvoted 8 Downvote Downvoted
DeepMind StarCraft II Demonstration
posted in Esports
1
#1
0 Frags +

Colleague just linked me this:
https://www.youtube.com/watch?v=cUTMhmVh1qs

Pretty impressive tbh

Colleague just linked me this:
https://www.youtube.com/watch?v=cUTMhmVh1qs

Pretty impressive tbh
2
#2
11 Frags +

haha were all gonna fucking die

haha were all gonna fucking die
3
#3
24 Frags +

deepmind TF2 team to finally dethrone FROYO

deepmind TF2 team to finally dethrone FROYO
4
#4
2 Frags +
lucrativedeepmind TF2 team to finally dethrone FROYO

God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.

[quote=lucrative]deepmind TF2 team to finally dethrone FROYO[/quote]

God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.
5
#5
1 Frags +
Comanglialucrativedeepmind TF2 team to finally dethrone FROYO
God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.

I mean you're right, but I was just fuckin' around man lol

[quote=Comanglia][quote=lucrative]deepmind TF2 team to finally dethrone FROYO[/quote]

God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.[/quote]

I mean you're right, but I was just fuckin' around man lol
6
#6
7 Frags +

@Comanglia The AI has an APM cap to make it more aligned with human limitations. If they were to take the same approach with TF2 I'm sure they'll cap the wep accuracy and reaction time for the same reason.

https://i.imgur.com/3YEKdu9.png

@Comanglia The AI has an APM cap to make it more aligned with human limitations. If they were to take the same approach with TF2 I'm sure they'll cap the wep accuracy and reaction time for the same reason.

[img]https://i.imgur.com/3YEKdu9.png[/img]
7
#7
3 Frags +

oriole really needs a glass of water lol

on a serious note this is really amazing, I can't wait to see even more advances in this field

if you don't want to watch the whole thing just watch this part, its like 3 minutes and gets the important parts across

https://www.youtube.com/watch?v=cUTMhmVh1qs&t=8485s

oriole really needs a glass of water lol

on a serious note this is really amazing, I can't wait to see even more advances in this field

if you don't want to watch the whole thing just watch this part, its like 3 minutes and gets the important parts across

https://www.youtube.com/watch?v=cUTMhmVh1qs&t=8485s
8
#8
1 Frags +

did any of the players try to cheese deepmind? I didn't scroll through every match but it looked like deepmind was way less effective at defending

did any of the players try to cheese deepmind? I didn't scroll through every match but it looked like deepmind was way less effective at defending
9
#9
eXtelevision
0 Frags +

TLO talked about trying to drop a hidden pylon in base for one of the un-broadcasted games of his. Apparently he had the placement barely off so it was revealed early, and the AI was actually doing a similar strat so it survived and won. I'd like to go through and cast the ones that weren't shown, but don't have time in the neXt few days unfortunately.

Def the thing with stacking workers is really crazy, I want to try that. Would be very curious to see what sort of Zerg shenanigans it would come up with.

TLO talked about trying to drop a hidden pylon in base for one of the un-broadcasted games of his. Apparently he had the placement barely off so it was revealed early, and the AI was actually doing a similar strat so it survived and won. I'd like to go through and cast the ones that weren't shown, but don't have time in the neXt few days unfortunately.

Def the thing with stacking workers is really crazy, I want to try that. Would be very curious to see what sort of Zerg shenanigans it would come up with.
10
#10
2 Frags +
hooli@Comanglia The AI has an APM cap to make it more aligned with human limitations. If they were to take the same approach with TF2 I'm sure they'll cap the wep accuracy and reaction time for the same reason.

https://i.imgur.com/3YEKdu9.png

where was it written in the article that they capped reaction time and apm?

articleIn its games against TLO and MaNa, AlphaStar had an average APM of around 280, significantly lower than the professional players, although its actions may be more precise. This lower APM is, in part, because AlphaStar starts its training using replays and thus mimics the way humans play the game. Additionally, AlphaStar reacts with a delay between observation and action of 350ms on average.

-> to me it's a byproduct of how it learned initially.

Also as mentioned on teamliquid, if they did implement a cap, they should have capped EPM. If the bot has 280APM and 280 EPM, it's way faster than Flash who has 400 apm and like 220 EPM at broodwar.

@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/

[quote=hooli]@Comanglia The AI has an APM cap to make it more aligned with human limitations. If they were to take the same approach with TF2 I'm sure they'll cap the wep accuracy and reaction time for the same reason.

[img]https://i.imgur.com/3YEKdu9.png[/img][/quote]
where was it written in the article that they capped reaction time and apm?

[quote=article]In its games against TLO and MaNa, AlphaStar had an average APM of around 280, significantly lower than the professional players, although its actions may be more precise. This lower APM is, in part, because AlphaStar starts its training using replays and thus mimics the way humans play the game. Additionally, AlphaStar reacts with a delay between observation and action of 350ms on average.[/quote]

-> to me it's a byproduct of how it learned initially.

Also as mentioned on teamliquid, if they did implement a cap, they should have capped EPM. If the bot has 280APM and 280 EPM, it's way faster than Flash who has 400 apm and like 220 EPM at broodwar.

@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/
11
#11
eXtelevision
2 Frags +
Twiggy@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/

What I think would be more impressive than how they currently have the Dota2 OpenAI setup, is rather than having 1 computer AI control 5 heroes, have 5 computer AIs all controlling 1 hero and having to independently assess+anticipate what their teammates are going to do. Definitely would be fun to see that sort of approach for TF2, 6 separate bots whose actions are motivated by their own individual desire to win.

[quote=Twiggy]@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/[/quote]

What I think would be more impressive than how they currently have the Dota2 OpenAI setup, is rather than having 1 computer AI control 5 heroes, have 5 computer AIs all controlling 1 hero and having to independently assess+anticipate what their teammates are going to do. Definitely would be fun to see that sort of approach for TF2, 6 separate bots whose actions are motivated by their own individual desire to win.
12
#12
0 Frags +
eXtineTwiggy@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/

What I think would be more impressive than how they currently have the Dota2 OpenAI setup, is rather than having 1 computer AI control 5 heroes, have 5 computer AIs all controlling 1 hero and having to independently assess+anticipate what their teammates are going to do. Definitely would be fun to see that sort of approach for TF2, 6 separate bots whose actions are motivated by their own individual desire to win.

wait i thought this was what was done already, similar to the quake ctf setup. Isn't it the case?

[quote=eXtine][quote=Twiggy]@Comanglia bots can already play team based games. It has been done on dota 2 and quake :
https://deepmind.com/blog/capture-the-flag/[/quote]

What I think would be more impressive than how they currently have the Dota2 OpenAI setup, is rather than having 1 computer AI control 5 heroes, have 5 computer AIs all controlling 1 hero and having to independently assess+anticipate what their teammates are going to do. Definitely would be fun to see that sort of approach for TF2, 6 separate bots whose actions are motivated by their own individual desire to win.[/quote]
wait i thought this was what was done already, similar to the quake ctf setup. Isn't it the case?
13
#13
0 Frags +
lucrativeComanglialucrativedeepmind TF2 team to finally dethrone FROYO
God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.

I mean you're right, but I was just fuckin' around man lol

do you think this is a fucking game?

[quote=lucrative][quote=Comanglia][quote=lucrative]deepmind TF2 team to finally dethrone FROYO[/quote]

God it would be terrible to play against a DeepMind AI in any FPS. Especially with the way it learns. It could literally have the dumbest most predictable movement ever and it still wouldn't matter cause it would hit 90+% of its shots on hitscan (limiting factor being computational resources and Source hit registration). For DeepMind AI the most efficient strategies would likely be picking whatever classes do the best hitscan Deathmatch at medium to long range so Sniper, Spy and Scout. If multiple AI could learn to play as a team we might actually see a medic. Demo and Soldier? probably not tbh.

The AI would have to be kneecapped with smoothed out aim and delayed reactions for players to even begin to stand a chance.[/quote]

I mean you're right, but I was just fuckin' around man lol[/quote]
do you think this is a fucking game?
14
#14
0 Frags +

@Twiggy see https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/

@Twiggy see https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/
15
#15
0 Frags +
hooli@Twiggy see https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/

yeah? what am i supposed to find there? In my first post i already quoted this very article which is why i'm confused.

[quote=hooli]@Twiggy see https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/[/quote]
yeah? what am i supposed to find there? In my first post i already quoted this very article which is why i'm confused.
16
#16
0 Frags +

Sorry m8, looks like they capped the mean APM and not the the APM itself. the APM went up too 1500 iirc

Sorry m8, looks like they capped the mean APM and not the the APM itself. the APM went up too 1500 iirc
17
#17
0 Frags +
hooliSorry m8, looks like they capped the mean APM and not the the APM itself. the APM went up too 1500 iirc

the pros can hit 1500 also

[quote=hooli]Sorry m8, looks like they capped the mean APM and not the the APM itself. the APM went up too 1500 iirc[/quote]
the pros can hit 1500 also
18
#18
0 Frags +

https://www.youtube.com/watch?v=8tq1C8spV_g

DeepMind was also the company behind the first AI to beat a top level human in Go, and I'd highly recommend watching the AlphaGo documentary if something like this interests you even a little bit.

It's a great watch despite how boring the idea of computer vs human might be in a film. Last I checked it was still on Netflix

https://www.youtube.com/watch?v=8tq1C8spV_g

DeepMind was also the company behind the first AI to beat a top level human in Go, and I'd highly recommend watching the AlphaGo documentary if something like this interests you even a little bit.

It's a great watch despite how boring the idea of computer vs human might be in a film. Last I checked it was still on Netflix
19
#19
0 Frags +

I'd be interested in watching deepmind play tf2, but have them limited to standard 6s meta classes. Sorta like how alphazero has played games in chess where it's sometimes limited to opening theory 10 moves in, like the french defense and see how it plays.

I'd be interested in watching deepmind play tf2, but have them limited to standard 6s meta classes. Sorta like how alphazero has played games in chess where it's sometimes limited to opening theory 10 moves in, like the french defense and see how it plays.
20
#20
0 Frags +

Comparing APM is pretty useless because SC players spam a lot of keys to keep their fingers warm whereas the AI is only useful APM, so the AI has a major advantage in that regard.

Also they say in the vid that the AI's reaction time is only 350 ms (whereas pro gamers probably average <200) which is an unfair comparison because the AI sees an action, thinks, and reacts to it in 350 ms which I don't believe a human could maintain. When you take a reaction speed test, you are just waiting for one consistent trigger and responding to it. You don't have to think. Just look at the insane stalker micro of the AI, blinking individual stalkers on low hp to the back of the pack immediately. No human could do that.

There's definitely a large mechanical advantage to the AI here, but it's still p cool.

Comparing APM is pretty useless because SC players spam a lot of keys to keep their fingers warm whereas the AI is only useful APM, so the AI has a major advantage in that regard.

Also they say in the vid that the AI's reaction time is only 350 ms (whereas pro gamers probably average <200) which is an unfair comparison because the AI sees an action, thinks, and reacts to it in 350 ms which I don't believe a human could maintain. When you take a reaction speed test, you are just waiting for one consistent trigger and responding to it. You don't have to think. Just look at the insane stalker micro of the AI, blinking individual stalkers on low hp to the back of the pack immediately. No human could do that.

There's definitely a large mechanical advantage to the AI here, but it's still p cool.
Please sign in through STEAM to post a comment.