EPTBut... yeah, that's already happening. Like, those were not small restrictions, that's quite impressive to hope to be able to remove them that fast.
OpenAI's Dota 2 bots vs. 5 top professionals in TI - Page 3
| Forum Index > Dota 2 General |
|
Murlox
France1699 Posts
But... yeah, that's already happening. Like, those were not small restrictions, that's quite impressive to hope to be able to remove them that fast. | ||
|
FreakyDroid
Macedonia2616 Posts
| ||
|
Murlox
France1699 Posts
I'm surprised they're working so fast :-) | ||
|
Nevuk
United States16280 Posts
On June 30 2018 05:48 Murlox wrote: Oh yeah, thank you. I had read it too, guess it went out of my head again after reading some comments here. But... yeah, that's already happening. Like, those were not small restrictions, that's quite impressive to hope to be able to remove them that fast. Yeah, I read it too and had to double check since so many people were complaining about restrictions that won't be there | ||
|
spudde123
4814 Posts
On June 30 2018 05:51 Murlox wrote: Aye but I mean... warding, invisibilty, illusions (bait), scanning, BoT plays... reading the blog, those seem like a big deal, programming-wise. I'm surprised they're working so fast :-) Ideally for most of that they wouldn't need to do that much programming in addition to what their system is now. You expose the bots to that information / possibility of that action, put them to play against themselves for 5000 years, and maybe they will have figured it out. It's hard to evaluate what sort of situations the bots can deal with atm so I don't really know what I should expect in terms of how easily different things can be learned. For example maybe the bots already have some behaviors that somewhat imitate warding, like a hero standing up a hill to give vision to the team. But they might also just be 5 manning relatively brainlessly, in which case I wouldn't have as much faith in them learning warding decently by TI. It's also possible that to some things like scanning they end up hard coding some solutions like they've done with item builds or creep blocking in 1v1. Hopefully they will lift as many restrictions as possible even if the bots turn out to be bad at dealing with something. I agree with WolfInTheSheep in the sense that I would much rather see the game be as close to regular Dota as possible, even if the bots are bad at some parts. That would make it an interesting thing to follow if the game restrictions stayed pretty much the same from benchmark to benchmark and we could see how the bots are improving. But I could see them keeping stuff banned if it turns out that humans can easily beat the bots using certain things. The hero restriction is understandable in the sense that this concept of learning purely through self play seems to run into scalability problems if you start allowing all the different heroes. It seems like you have to figure out how to transfer knowledge across heroes to some extent to make it viable. | ||
|
WolfintheSheep
Canada14127 Posts
On June 30 2018 06:12 Nevuk wrote: Yeah, I read it too and had to double check since so many people were complaining about restrictions that won't be there I mean you've kind of pushed it in the opposite end now. They don't say those restrictions will be gone. They say they're aiming to do so, and say that they don't know if that goal is achievable. | ||
|
Nevuk
United States16280 Posts
On June 30 2018 06:40 WolfintheSheep wrote: I mean you've kind of pushed it in the opposite end now. They don't say those restrictions will be gone. They say they're aiming to do so, and say that they don't know if that goal is achievable. True, true. I took them to be meaning "hope to be able to beat pros without restrictions", but I'm unsure where the emphasis is on that sentence - the "beat pros" or "without restrictions" | ||
|
sertas
Sweden890 Posts
| ||
|
WolfintheSheep
Canada14127 Posts
On June 30 2018 06:56 sertas wrote: Have read some comments but my 2cent is that even the mid 1v1 sf had a bunch of stupid restrictions like no wand and no bottle etc. This clearly means they dont even have a bot that masters 1v1 only mid, only sf. I knew 100% they would never be able to play 5v5 without a ton of restrictions since even their 1v1 is flawed. Their entire methodology needs to change if they want to make a 5v5 bot even close to fighting pro teams. AI learning is iterative. You can shoot for the moon from the start and have nothing for a very long time, or you can start with small goals and have steady progress checkpoints. | ||
|
Murlox
France1699 Posts
On June 30 2018 06:47 Nevuk wrote: I took them to be meaning "hope to be able to beat pros without restrictions", but I'm unsure where the emphasis is on that sentence - the "beat pros" or "without restrictions" Probably both  They're essentially the same, are they not. Getting the stuff advanced enough. | ||
|
intotheheart
Canada33091 Posts
On June 30 2018 06:47 Nevuk wrote: True, true. I took them to be meaning "hope to be able to beat pros without restrictions", but I'm unsure where the emphasis is on that sentence - the "beat pros" or "without restrictions" I think "without restrictions" is what they want to focus on next. My intuition tells me that they're trying to incorporate a wider range of capabilities at a time, and then work on the quality later. After all, if you overtrain an AI to do something like play without wards, that AI may have to unlearn stuff when faced without that restriction. | ||
|
nojok
France15845 Posts
| ||
|
Pontual
Brazil3038 Posts
On July 04 2018 02:26 nojok wrote: Meh, it's just brute forcing and ultra powerful computers in a very limited version of dota. It's disappointing but also very reassuring. I don't want to be in the generation with dematerialized humans, though it's probably our best option to get out of the solar system. I would love to last until those times, living forever in a world where you're not limited by physics but program seems good to me | ||
|
polgas
Canada1770 Posts
| ||
|
broodmann
604 Posts
| ||
|
0fuulm
1 Post
BUT DONT let the humans train against AI. Then call the AI smart. Ver scientific | ||
|
Sn0_Man
Tebellong44238 Posts
On July 07 2018 03:25 0fuulm wrote: Let the AI train AGAINST HUMANS. BUT DONT let the humans train against AI. Then call the AI smart. Ver scientific The AI isn't in "training mode" when playing against humans As in, it doesn't use any results against humans to evaluate it's network and tune anything. Basically, playing against humans does not change the bot in any way (similar to playing against regular bots, which have no learning capacity). In addition, it's actually impossible for a bot to train against humans without those humans in turn "training" against bots. Humans are always learning. Everything we do is indelibly marked on our brains and impacts every future decision in some (potentially infintesimal) way. Only computers can play without in any way changing their knowledge of the game. Lastly, the computer plays 180 years of dota per day against itself. Compared to that, training against humans in real time would be incredibly inefficient. The bot learns SLOWLY from individual games, whereas humans can learn very rapidly from individual games. If a weak bot played weak humans and both were trying to learn from each other, the humans would vastly outpace the bot in learning. | ||
|
codeafin
7 Posts
Machine learning is always interesting because they find ways to do/make sense of things that we would never think of. They'll be coming up with their own metas | ||
|
FreakyDroid
Macedonia2616 Posts
| ||
|
Tanukki
Finland579 Posts
| ||
| ||
