EPTPlay against OpenAI on April 18-21:
https://arena.openai.com/
OpenAI's Dota 2 bots vs. 5 top professionals in TI - Page 15
| Forum Index > Dota 2 General |
|
polgas
Canada1770 Posts
Play against OpenAI on April 18-21: https://arena.openai.com/ | ||
|
Gear 3rd
1244 Posts
| ||
|
haxhax
125 Posts
| ||
|
YBBUB
70 Posts
No illusions or summons = nerf the shit out of n0tail & OG; plus OG couldn't cope last year when DP was in meta...Topson & ana don't really play SF/slark/sniper [but why would they go all out & show any real strats anyway  ]Anyway I liked 4 cores plus 7.21b CM draft [even had more stuns/disables than OG] & I've always thought more even networth distribution has been the way to evolve. Like I've always thought all 5 reaching level 6 with some items is quite strong agst an opponent with more uneven levels & networth. | ||
|
polgas
Canada1770 Posts
| ||
|
polgas
Canada1770 Posts
https://openai.com/blog/how-to-train-your-openai-five/ The OpenAI Final version beat the OpenAI TI8 version 99.9% of the time. And it got good just by scaling it: let it train by itself more. Their graph also says that the OpenAI is not plateauing and is still getting better. Would anyone be able to beat it online this April 18-21? | ||
|
spudde123
4814 Posts
Hopefully some pro teams try to play a number of games against the bot so we see some interesting games. The issue with the benchmark once again was that 2 games isn't enough to see the bot in various different situations. Obviously their teamfighting was again great and from what I saw they moved around the map pretty nicely most of the time. Given that they played against other pro teams it would have been nice to see recordings of those games. The humans don't have any data on the current version, so the bots drafting their own lineup from a 17 hero pool is a big advantage for them. Not sure if it's just a coincidence that the only time the bots lost was when they weren't allowed to pick their best lineup and instead had to deal with something else. Of course seeing them at their best is interesting as well, but for evaluating how well they've learned the game as a whole it would also be important to see them on other types of lineups. | ||
|
polgas
Canada1770 Posts
| ||
|
CuddlyCuteKitten
Sweden2756 Posts
Two things could happen: 1) The AI learns to actually check numbers on things like hero stats/abilities and items and adapts towards that (for example: on this patch sniper has worse stats and I seem to lose more with him, better pick something else). 2) The AI learns to focus more on general strategies that work well in general and are more resistant to patch changes. | ||
|
YBBUB
70 Posts
On April 16 2019 03:30 polgas wrote: AI doing their utmost in microing illusions & summons would be just as unrealistic/unfun as WC3 DotA AI map where computer always know which illusion is real, and computer Puck always phase-shift dodges my Luna Lucent Beam (as computer knows my keyboard/mouse input before it actions).Would humans be better than the AI in microing illusions and summons? OG was also informed in advance about the rules, not like they sprang a surprise during the event that oh sorry you can only pick between these 17 heroes. It may be hard to program AI to use illusions & summons like a human without seeming too "superhuman"? I'd like them to put more heroes in, but at same time still simulate what human players do as it's super interesting to think AI can come up with some interesting combos/strats. OG plays a lot of those micro kinda heroes compared to other teams & it's not about whether they know the hero pool or not - there's probably at least 8 other teams that would be more suited to the AI hero pool 'meta'. If Paris Disney EU quals were restricted to those 17 heroes, I'm not sure if OG would have qualified. | ||
|
lolfail9001
Russian Federation40190 Posts
On April 16 2019 21:57 YBBUB wrote: AI doing their utmost in microing illusions & summons would be just as unrealistic/unfun as WC3 DotA AI map where computer always know which illusion is real, and computer Puck always phase-shift dodges my Luna Lucent Beam (as computer knows my keyboard/mouse input before it actions). That stuff won't happen simply because OpenAI is not actually a WC3 Dota AI. If anything that illusion/summons is just another restriction that makes sure that what OpenAI won at was a custom map OG never played on, that's about it. | ||
|
YBBUB
70 Posts
That stuff won't happen simply because OpenAI is not actually a WC3 Dota AI. The emphasis from my post should really be on "as unrealistic/unfun as WC3 DotA AI map" - meaning it's not fun when computer cheats by doing stuff humans can't - like if they program a computer player that can almost immediately tell without error which one of you is real vs illusion , or if they can give multiple simultaneous commands to main plus multiple illusions at the same time. Pros are really good at detecting which is the real one & I wouldn't want a computer that is much better at that. I was responding to polgas's statement about "Would humans be better than the AI in microing illusions and summons?" with "I don't want AI to be much better in microing illusions and summons to the extreme of above which feels like cheating" - AI map is better stated as "map where computer player had built-in cheats", not "DotA AI" Experimental Custom map with 17 heroes & no illus/summons - yes | ||
|
Murlox
France1699 Posts
Am I the only one noticing many more buybacks in midgame? I've had it happen quite a few games, and just seen it on some competitive match as well. I think this is what it boils down to, for me. What will our AI overlord teach us. Because I'm pretty sure it will outperform us (to infinity), so even if it is interesting to see that happening, I think the real benefit is to see what we can learn from it. Just like the first iteration changed the way the mid matchup is played (much more regen ferrying), I'm curious to see if this iteration will impact the pro teams, and therefore the plebs after that. For instance more buybacks, or even... a more horizontal farm spread (ie AI farms all their heroes, no pos6 shinanigans), into, why not, death-ball push. | ||
|
polgas
Canada1770 Posts
| ||
|
lolfail9001
Russian Federation40190 Posts
On April 17 2019 23:12 YBBUB wrote: The emphasis from my post should really be on "as unrealistic/unfun as WC3 DotA AI map" - meaning it's not fun when computer cheats by doing stuff humans can't - like if they program a computer player that can almost immediately tell without error which one of you is real vs illusion , or if they can give multiple simultaneous commands to main plus multiple illusions at the same time. Pros are really good at detecting which is the real one & I wouldn't want a computer that is much better at that. I was responding to polgas's statement about "Would humans be better than the AI in microing illusions and summons?" with "I don't want AI to be much better in microing illusions and summons to the extreme of above which feels like cheating" - AI map is better stated as "map where computer player had built-in cheats", not "DotA AI" Experimental Custom map with 17 heroes & no illus/summons - yes But that's the point: OpenAI is not going to actually do stuff like that (instant identifying of illusions unless they attack something or get attacked by something is impossible, period). At best they can probably learn to use summons to tank aggro perfectly while tower diving or box-in heroes, but is it something unrealistic for humans to do? | ||
|
polgas
Canada1770 Posts
https://www.twitch.tv/openai | ||
|
spudde123
4814 Posts
Didn't see either game from the beginning so not sure exactly what humans did, but at least the end result was that the humans pretty heavily outfarmed the bots despite not necessarily being up in kills | ||
|
polgas
Canada1770 Posts
| ||
|
spudde123
4814 Posts
| ||
|
polgas
Canada1770 Posts
| ||
| ||
