Dota 2 AI beats pro players in 5v5

[Jaeger]

1 or 2 years ago OpenAI made news when it beat the top Dota 2 players in 1v1. However, many people said that 1v1 is nothing compared to 5v5 in complexity, and AI will not bet humans in 5v5 anytime soon.

Now AI on won in 5v5.

Can someone who knows Dota 2 shed some light on this topic? Has Dota 2 been beaten by computers?


https://www.youtube.com/watch?v=eaBYhLttETw

[Gendarme]

Lmao

[gibson]

It makes sense, a computer processor can process much more information much quicker than our brain can. It's just a matter of giving it the correct imputs to learn the game properly

[Gendarme]

But can AI feel love?

[deleted_user0]

the question is would an AI open 3 kav proxy stable?!

[supahons]

yes, like deepblue in chess or alpha go in go

How do you win against an ai with 3400+ elo? (like current chess engines)

[Jaeger]

yes, like deepblue in chess or alpha go in go

How do you win against an ai with 3400+ elo? (like current chess engines)

Well it took a lot of work to get to this point as you know. They only beat GO with AI a couple years ago, couldn't do it before. And they couldn't beat Dota 2 in 5v5 until this year.

[momuuu]

I recall watching some human vs AI showmatches in the last TI (which was last year or something, possibly where these showmatches are from? I dont follow the dota scene much). From what I recall this match is from semi-pros against the bots, and I think in the end the semi pros won that series 2-1 or something (this is all vague memory though, I could be wrong). Later they had a real top team play against OpenAI and that top team actually did win somewhat easily (although it wasn't a stomp).

Dota is a very intuitive game however, where experience and mechanics beat true strategical understanding. From what I recall, the AI actually had some very apperant strengths and weaknesses. I believe mechanically the AI players were superb, and knew almost exactly how to push their abilities to the limit. The calls they made in close skirmishes were amazing. However, I believe the AI was not coded to have perfect information of other AI members, so it really was a team of 5 different AIs and not one big overmind. This resulted in mediocre performance in team fights, where especially the actual pro team did much better iirc. Strategically the AI also behaved a bit weirdly, where they'd not actually choose to go for a true support/carry/mid/offlane hierarchy as much as is common in top level dota, and they seemed to overvalue pushing out the lanes to a point where it seemed like feeding.

In the end it seemed like the AI for dota is strong at AI things; they have better intuition because they probably internalized a huge amount of data and played a huge amount of games. However, coordination was somewhat subpar and strategically the AI players played poorly. The strategic performance is what would be interesting, and I found that that part was disappointing. The AI overperforms at management of abilities, as we would expect from the AI, but it underperformed at the strategic side. It would be like an aoe3 bot with amazing micro and macro but questionable shipment/build order/composition decisions. So in the end it only showed how limited humans are in controlling the game and also showed how humans make up for that by having a better strategic understanding. It'd have been more exciting to see the AI actually outstrategize the humans and come up with things we'd have never considered to be good.

[HUMMAN]

As jerom said ai was not that good in 5v5 also they played pre determined teams (no picking) so still a way to go. But they showed some unstandart play that is possible (often carrying healing salves in mid game) which was interesting.

[Jaeger]

I recall watching some human vs AI showmatches in the last TI (which was last year or something, possibly where these showmatches are from? I dont follow the dota scene much). From what I recall this match is from semi-pros against the bots, and I think in the end the semi pros won that series 2-1 or something (this is all vague memory though, I could be wrong). Later they had a real top team play against OpenAI and that top team actually did win somewhat easily (although it wasn't a stomp).

Dota is a very intuitive game however, where experience and mechanics beat true strategical understanding. From what I recall, the AI actually had some very apperant strengths and weaknesses. I believe mechanically the AI players were superb, and knew almost exactly how to push their abilities to the limit. The calls they made in close skirmishes were amazing. However, I believe the AI was not coded to have perfect information of other AI members, so it really was a team of 5 different AIs and not one big overmind. This resulted in mediocre performance in team fights, where especially the actual pro team did much better iirc. Strategically the AI also behaved a bit weirdly, where they'd not actually choose to go for a true support/carry/mid/offlane hierarchy as much as is common in top level dota, and they seemed to overvalue pushing out the lanes to a point where it seemed like feeding.

In the end it seemed like the AI for dota is strong at AI things; they have better intuition because they probably internalized a huge amount of data and played a huge amount of games. However, coordination was somewhat subpar and strategically the AI players played poorly. The strategic performance is what would be interesting, and I found that that part was disappointing. The AI overperforms at management of abilities, as we would expect from the AI, but it underperformed at the strategic side. It would be like an aoe3 bot with amazing micro and macro but questionable shipment/build order/composition decisions. So in the end it only showed how limited humans are in controlling the game and also showed how humans make up for that by having a better strategic understanding. It'd have been more exciting to see the AI actually outstrategize the humans and come up with things we'd have never considered to be good.

Thank you for the comment. But did you watch the video that I posted from this year?

[t3rror1sta]

yes, like deepblue in chess or alpha go in go

How do you win against an ai with 3400+ elo? (like current chess engines)

you walls + water

[momuuu]

I recall watching some human vs AI showmatches in the last TI (which was last year or something, possibly where these showmatches are from? I dont follow the dota scene much). From what I recall this match is from semi-pros against the bots, and I think in the end the semi pros won that series 2-1 or something (this is all vague memory though, I could be wrong). Later they had a real top team play against OpenAI and that top team actually did win somewhat easily (although it wasn't a stomp).

Dota is a very intuitive game however, where experience and mechanics beat true strategical understanding. From what I recall, the AI actually had some very apperant strengths and weaknesses. I believe mechanically the AI players were superb, and knew almost exactly how to push their abilities to the limit. The calls they made in close skirmishes were amazing. However, I believe the AI was not coded to have perfect information of other AI members, so it really was a team of 5 different AIs and not one big overmind. This resulted in mediocre performance in team fights, where especially the actual pro team did much better iirc. Strategically the AI also behaved a bit weirdly, where they'd not actually choose to go for a true support/carry/mid/offlane hierarchy as much as is common in top level dota, and they seemed to overvalue pushing out the lanes to a point where it seemed like feeding.

In the end it seemed like the AI for dota is strong at AI things; they have better intuition because they probably internalized a huge amount of data and played a huge amount of games. However, coordination was somewhat subpar and strategically the AI players played poorly. The strategic performance is what would be interesting, and I found that that part was disappointing. The AI overperforms at management of abilities, as we would expect from the AI, but it underperformed at the strategic side. It would be like an aoe3 bot with amazing micro and macro but questionable shipment/build order/composition decisions. So in the end it only showed how limited humans are in controlling the game and also showed how humans make up for that by having a better strategic understanding. It'd have been more exciting to see the AI actually outstrategize the humans and come up with things we'd have never considered to be good.

Thank you for the comment. But did you watch the video that I posted from this year?

All I wrote was about this year too. Didnt watch this video as Ive already seen a lot of this stuff.