OpenAI at the Dota 2 World Championships • Hacker News Board

Re: OpenAI at the Dota 2 World Championships

#31

Post by edran » Sat, Aug 12, 2017, 12:04 AM UTC

justicezyx wrote: post #20

To be clear 1v1 mid lane is harder than 1 unit v 1 unit in sc. But it is at least 3 levels behind 5v5 human play. So first it's laining, which is showing in the game. Then is the ganking basically 2-3 player working together. Above that that's 5 men team fight. Then you have the strategy planning in the band pick phase, and in game movement coordination. There are other things like item choices, in game communication…

The micromanagement task in StarCraft Broodwar (with TorchCraft) that has been used by a few people already in the past year to do some RL research has been tackled basically only with #units > 1, because getting a balanced 1v1 is possible only with certain combination of units, and it's indeed pointless otherwise.

Kiting is maybe a notable exception, but it is also restricted to a relatively small group of unit matchups.

See for instance the following papers:

- https://arxiv.org/abs/1609.02993

- https://arxiv.org/abs/1702.08887

- https://arxiv.org/abs/1703.10069

- https://arxiv.org/abs/1705.08926

Full disclosure: I'm one of such researchers, and I have authored some of the papers on the topic.

Re: OpenAI at the Dota 2 World Championships

#32

Post by cissou » Sat, Aug 12, 2017, 12:05 AM UTC

cissou wrote: post #15

It looks like it's over…? Someone has a replay of the relevant part(s)?

https://www.youtube.com/watch?v=ac1getNs2P8

Re: OpenAI at the Dota 2 World Championships

#33

Post by ronald_raygun » Sat, Aug 12, 2017, 12:05 AM UTC

justicezyx wrote: post #20

To be clear 1v1 mid lane is harder than 1 unit v 1 unit in sc. But it is at least 3 levels behind 5v5 human play. So first it's laining, which is showing in the game. Then is the ganking basically 2-3 player working together. Above that that's 5 men team fight. Then you have the strategy planning in the band pick phase, and in game movement coordination. There are other things like item choices, in game communication…

I'm not sure about DOTA, but in LOL there is also champion select which happens before the game starts, where you try to pick/ban/steal champions to get a good team comp and get the other team to have a bad team comp. It's lots of know overall strategy, what your enemy likes, and some game theory.

You can lose the game before your summoner even hits the field.

Re: OpenAI at the Dota 2 World Championships

#34

Post by mAritz » Sat, Aug 12, 2017, 12:05 AM UTC

kmnc wrote: post #27

How do things like reaction time, and actions per second work with something like this? Is it just an assumed advantage the ai gets, or does it simulate the limitations of a human? If it doesn't, how big of an advantage is it in an ai versus human competition?

That was pretty much the entire reason it won. It did some standard high-level 1v1 laning techniques, but if human "conditions"[1] were implemented it would look a lot different.

Just the blocking of the lane creeps at the start is already super-human and gives the bot a huge advantage.

[1] like lower actions per minute, latency, occasionally missclicking and not being 100% certain about distances

Re: OpenAI at the Dota 2 World Championships

#35

Post by hhmc » Sat, Aug 12, 2017, 12:08 AM UTC

mIREdeRMedeFLaO wrote: post #7

_xhok wrote: post #2

Is it taking in raw pixels or reading from game memory? How did they gather data to train for this? (Did Valve give them an API?)

https://developer.valvesoftware.com/wiki/Dota_Bot_Scripting > Bot scripting in Dota is done via lua scripting. This is done at the server level, so there's no need to do things like examine screen pixels or simulate mouse clicks; instead scripts can query the game state and issue orders directly to units. Scripts have full access to all the entity locations, cooldowns, mana values, etc that a player on that team woul…

I wouldn't be surprised if they had access to a less restricted AI. Until we see more concrete information I remain sceptical that they are only using the official API.

Re: OpenAI at the Dota 2 World Championships

#36

Post by joefkelley » Sat, Aug 12, 2017, 12:08 AM UTC

This will be HUGE for competitive Dota, even if they never take it further.

In much the same ways top chess players have learned from engines, I have to think top Dota players will practice against and study the hell out of this bot if OpenAI makes it available.

Not everything is replicable by humans... for instance I noticed it constantly animation-canceling razes and only finishing it if it was going to hit; a human will definitely mess this up. But other things can definitely be used. It was positioning somewhat strangely, for example.

Re: OpenAI at the Dota 2 World Championships

#37

Post by criloz2 » Sat, Aug 12, 2017, 12:09 AM UTC

demonshalo wrote: post #26

MEH. I am not impressed. Restricting the number of parameters and variables in order to produce a bot that can do 1 sub-set of tasks really well is nothing special imo. Complexity and simulations of what you have not yet encountered is something humans can do with ease. This is not something a bot can do in a complex environment like DOTA. I'll be the first to admit that I was wrong and that AI is truly a thing once…

is an improvement against the other bots, it can be really good for training players, I would love to test it my self, but yeah dota is biggest that this, not only 5 v5, also 100+ heroes, and professional player can play a variety of heroes on mid with different matchups and its nuisances, the bot still need it a lot to even dominate 1v1.

Re: OpenAI at the Dota 2 World Championships

#38

Post by derimagia » Sat, Aug 12, 2017, 12:09 AM UTC

ronald_raygun wrote: post #33

justicezyx wrote: post #20

To be clear 1v1 mid lane is harder than 1 unit v 1 unit in sc. But it is at least 3 levels behind 5v5 human play. So first it's laining, which is showing in the game. Then is the ganking basically 2-3 player working together. Above that that's 5 men team fight. Then you have the strategy planning in the band pick phase, and in game movement coordination. There are other things like item choices, in game communication…

I'm not sure about DOTA, but in LOL there is also champion select which happens before the game starts, where you try to pick/ban/steal champions to get a good team comp and get the other team to have a bad team comp. It's lots of know overall strategy, what your enemy likes, and some game theory. You can lose the game before your summoner even hits the field.

Yep it's called Drafting

Re: OpenAI at the Dota 2 World Championships

#39

Post by mAritz » Sat, Aug 12, 2017, 12:11 AM UTC

hhmc wrote: post #35

mIREdeRMedeFLaO wrote: post #7

Earlier quoted context omitted.

https://developer.valvesoftware.com/wiki/Dota_Bot_Scripting > Bot scripting in Dota is done via lua scripting. This is done at the server level, so there's no need to do things like examine screen pixels or simulate mouse clicks; instead scripts can query the game state and issue orders directly to units. Scripts have full access to all the entity locations, cooldowns, mana values, etc that a player on that team woul…

I wouldn't be surprised if they had access to a less restricted AI. Until we see more concrete information I remain sceptical that they are only using the official API.

Why? Everything it needed for that 1v1 was available. There was nothing in the FoW that needed to be known, it can keep timers about enemy cooldowns, gold and xp (approximations at least since it's somewhat random).

Re: OpenAI at the Dota 2 World Championships

#40

Post by hhmc » Sat, Aug 12, 2017, 12:14 AM UTC

taion wrote: post #14

Is the "two weeks" of time in the Dota environment, or two weeks of training time running in parallel as with A3C or something?

From what the devs indicated on stream it sounded like processing hours (with a ratio of around 300 in-game hours to 1 processing hour).