EPTThank you...
OpenAI's Dota 2 bots vs. 5 top professionals in TI - Page 16
| Forum Index > Dota 2 General |
|
asker71
1 Post
Thank you... | ||
|
ShiaoPi
TAIWAN NUMBAH WAN5956 Posts
On April 20 2019 00:28 asker71 wrote: Hello, everybody! Can someone explain to me how Split Dota tournament is played. Who plays who after the group stage? Thank you... Details for the Split Minor Playoffs have not been announced yet iirc. But this is not really the thread to ask in  | ||
|
saocyn
United States937 Posts
there were 3 Stark differences I've noticed they capitalized on which ideally humans cannot. 1. Perfect Coordination Almost no ganks were uncoordinated, you could always trust there was a follow-up 2. Even Gold Distribution it makes sense that if every hero got farmed, your team as a whole would be stronger than the opponents, even more so if the entire team has perfect synergy. 3. Role Cycling at different points, the AI would allocate different roles to where the tempo of the game was going. If you had more or less gold, they would interchange between even core roles and support roles. This is entirely logical and make sense but due to humans attributing main 1 dimensional roles to most heroes, something like this goes beyond their heads. I'm extremely impressed with where the AI is currently at. The item choices even resemble human thought processes, and it's quite scary that we're almost to the point of getting replaced. I wonder how long it would take for AI to expand their hero pool and find winning patterns in more chaotic/unfavorable situations & drafts. | ||
|
RolleMcKnolle
Germany1054 Posts
On April 23 2019 12:44 saocyn wrote: I finally got around to watching the games and I believe the decision making of Open AI is extremely sensible. there were 3 Stark differences I've noticed they capitalized on which ideally humans cannot. 1. Perfect Coordination Almost no ganks were uncoordinated, you could always trust there was a follow-up 2. Even Gold Distribution it makes sense that if every hero got farmed, your team as a whole would be stronger than the opponents, even more so if the entire team has perfect synergy. 3. Role Cycling at different points, the AI would allocate different roles to where the tempo of the game was going. If you had more or less gold, they would interchange between even core roles and support roles. This is entirely logical and make sense but due to humans attributing main 1 dimensional roles to most heroes, something like this goes beyond their heads. I'm extremely impressed with where the AI is currently at. The item choices even resemble human thought processes, and it's quite scary that we're almost to the point of getting replaced. I wonder how long it would take for AI to expand their hero pool and find winning patterns in more chaotic/unfavorable situations & drafts. There is a shift in farming priorities in human teams as well. I am not sure why you think it wouldnt. IIRC they said at some point that most item-choices including regen are still scripted. | ||
|
Murlox
France1699 Posts
I believe that, in pubs at least, your chances of winning go way up when everyone is happy to play. People are more resilient to lane losses, bad plays, are more willing to communicate, coordinate, etc. Same goes with buying wards, I know I do buy them no matter my position, if I feel we need some at the moment and no one is doing it. I notice a lot of people who want to win do it as well, it's just simpler and faster. All in all, I've never been a fan of the sacrificial 5 or 4. I believe that is a trend that has been developed in competitive dota, where people are willing to forgo their fun in pursuit of something bigger, like actual money, fame, carrier, idk. I never fell this was ok in pubs, en large people seem way more enclined to pick cores than supports. Oddly enough, I think the AI gold distribution is much more human that what we humans are currently doing... | ||
|
Murlox
France1699 Posts
On April 23 2019 14:43 RolleMcKnolle wrote: IIRC they said at some point that most item-choices including regen are still scripted. Yeah, maybe there's a comprehensive video/text somewhere that lists the current limitations, but I wish it was made more clear. Same with their coordination, or planning, at least back in last iteration (TI), there was none of that. If I got this right, they simply did not talk to each other, at all. Yet people (casters, even Bill Gates in his tweet) keep bringing up those points. In the last event, Notail was openly wondering how far can the AI see, how many moves in advance do they have. It is my belief that as a group, there's no planning at all (because they don't talk to each other), and as individuals, there's a very little planning ahead, rather a superhuman assessment of the current situation, all the time. I wish the openAI engineers were a bit more informative / open about that. Maybe dared to correct people, at the price of scratching the magic a little bit. I could be wrong, it's just... what I gathered. | ||
|
spudde123
4814 Posts
Also I think all the behaviors the bots show should be seen in the context of their training process. They don't learn by directly trying to win the game and seeing what works, but rather they have a human crafted reward functions, which includes things like gold, xp, buildings being alive and so on. Originally they just prioritize their own reward which hopefully leads to them learning to gather gold and so on, and later on they start prioritizing the average reward of the team. Some of the "gold for everyone" behavior may just be remnants of the bots learning to gather gold for themselves, rather than it being strictly optimal for winning. But of course everything they end up doing differently is worth considering, and I think especially in their teamfight centric lineups there is merit in distributing resources more evenly. In some other lineup which relies on a last pick AM or some pickoff heavy lineup, I think it would be a bit different. Sometimes some heroes really do benefit more from the resources. Though certainly in pubs people should be a lot more flexible about it. And as someone said above in pro dota I don't think the farm distributions are all that static at all. Sometimes it's all about a hyper carry, sometimes 2 or even 3 heroes are pretty equal in how much room they get. In the past at some point the "support" KotL was farming all the time until aghs, now sometimes it's the Enigma that allows the team to have 4 really farmed heroes. When Furion is played as 5 he often isn't all that poor because he is the one dealing with pushing out other lanes On April 23 2019 15:12 Murlox wrote: Same with their coordination, or planning, at least back in last iteration (TI), there was none of that. If I got this right, they simply did not talk to each other, at all. Yet people (casters, even Bill Gates in his tweet) keep bringing up those points. In the last event, Notail was openly wondering how far can the AI see, how many moves in advance do they have. It is my belief that as a group, there's no planning at all (because they don't talk to each other), and as individuals, there's a very little planning ahead, rather a superhuman assessment of the current situation, all the time. I wish the openAI engineers were a bit more informative / open about that. Maybe dared to correct people, at the price of scratching the magic a little bit. I could be wrong, it's just... what I gathered. I think this depends a bit on what people mean by "planning". From what I understand the bots don't explicitly plan ahead. However, in their learning process they end up trying to optimize longer term rewards, averaged for the entire team. So after a long training time they may exhibit behavior which may look like "they are planning to do something in 30 seconds" because they are foregoing some immediate rewards like killing creeps to move around for a push or to limit the enemy teams farm. I think one aspect where the lack of a "real" planning process sort of shows is when the bots are losing. Humans know that in the end the ancient is the big thing that matters, but for the bots that's not the case at all. They have learnt to care about gold, xp, towers and so on. Their ancient dying is just something extra that gives a negative reward and not something fundamental. From watching quite a bit of the Arena this weekend, it seems to me that when the bots are close to losing they still just try to "minimize their losses", which means not taking fights they are likely to lose and so on. Humans often have these moments where they know just being passive leads to a loss and they go for a last effort smoke play to contest rosh or to try to get a good angle of initiation, but the bots don't really understand such things for now. | ||
| ||
