EPTOpenAI's Dota 2 bots vs. 5 top professionals in TI - Page 9
| Forum Index > Dota 2 General |
|
spudde123
4814 Posts
| ||
|
WolfintheSheep
Canada14127 Posts
| ||
|
FreakyDroid
Macedonia2616 Posts
| ||
|
spudde123
4814 Posts
Sometimes the bots seem a bit confused though | ||
|
WolfintheSheep
Canada14127 Posts
And there's that typical "losing base but probability says we can't fight so just ignore base" behaviour. | ||
|
polgas
Canada1770 Posts
![[image loading]](https://i.imgur.com/OcYthOG.jpg) | ||
|
FreakyDroid
Macedonia2616 Posts
| ||
|
Murlox
France1699 Posts
On August 06 2018 08:38 FreakyDroid wrote: They seem to have a rather simplistic concept of the game at this stage, its mostly pick heroes with long range nukes that enables them to make kills and/or zone out opponent heroes and then push as 5. Not sure how far they can get just by playing bot vs bot matches, perhaps allowing them to analyze top level replays from pro teams might give them more ideas/concepts that they can utilize. True, but I believe the goal of the team is to make a self learning machine. Being extremely good at dota can come second. So maybe there are willing to accept that the machine is not optimal in some situations because it didn't have enough time to encounter those situations yet, rather than "help it get good faster" by feeding it data. But... maybe they'll talk about it on the panel. Team members have not been very technical yet, though, I found. The blog was much more useful. | ||
|
Kometijanac
Serbia98 Posts
| ||
|
polgas
Canada1770 Posts
| ||
|
ASoo
2865 Posts
Of course, the bot will probably also be better by the time it plays a real pro team, so who knows? | ||
|
WolfintheSheep
Canada14127 Posts
On August 06 2018 09:32 polgas wrote: It was a treat to see how machine learning playing by itself can be this good. It felt like the AI always has the initiative and coupled with its efficiency it just snowballs to a win. Also quite scary to actually see AI's aggression. I think this was the most boring part of the AI strat, IMO. Coordinating spells and calculating damage is fairly basic behaviour, and should be a fairly direct line of learning once kill rewards are input as variables. Ditto to charging down towers, it should just be reward based decision making, since those are the most valuable targets at that point in the game (and that's not even considering abstract value like map control). How that factors into draft was interesting, but ultimately limited by the hero pool. We saw from the draft probabilities that almost everything for the bots came down to a single gameplan. The 3rd game split-pushing and tactical feeding was much more interesting overall, as that goes well beyond X > Y decision making. Obviously they need some further development to learn how to not feed, and how to prioritize saving major objectives over staying alive, but it showed some actual reactionary strategy and game-state evaluation. | ||
|
spudde123
4814 Posts
For pros it's also very useful that they've now seen the bots play so they have some idea of what to expect even if it's not the same version. Pretty clearly today the human team was taken by surprise by a lot of things and they played various situations as they would against humans and just got owned by the spell casting from the bots. | ||
|
SimplyPanda
United States15 Posts
| ||
|
polgas
Canada1770 Posts
On August 06 2018 09:47 WolfintheSheep wrote: I think this was the most boring part of the AI strat, IMO. Coordinating spells and calculating damage is fairly basic behaviour, and should be a fairly direct line of learning once kill rewards are input as variables. Ditto to charging down towers, it should just be reward based decision making, since those are the most valuable targets at that point in the game (and that's not even considering abstract value like map control). How that factors into draft was interesting, but ultimately limited by the hero pool. We saw from the draft probabilities that almost everything for the bots came down to a single gameplan. The 3rd game split-pushing and tactical feeding was much more interesting overall, as that goes well beyond X > Y decision making. Obviously they need some further development to learn how to not feed, and how to prioritize saving major objectives over staying alive, but it showed some actual reactionary strategy and game-state evaluation. Agree 3rd game was interesting but in the end it felt like the AI was playing to delay the inevitable rather than to win. I believe if the bots were not scripted to buy items based on a guide it might have been more challenging. | ||
|
acie
United States247 Posts
| ||
|
WolfintheSheep
Canada14127 Posts
On August 06 2018 11:57 polgas wrote: Agree 3rd game was interesting but in the end it felt like the AI was playing to delay the inevitable rather than to win. I believe if the bots were not scripted to buy items based on a guide it might have been more challenging. Probably a lot less challenging for human opponents. If the bots were scripted to buy recommended items, then they probably have no idea how to evaluate item buys, or the recommended guide is actually superior to their current AI. Which makes sense. Future predictions are not a simple task, and getting an AI to evaluate the risk/reward of buying immediate item benefits vs saving for strong future items suited to the enemies is going to be a giant hurdle. | ||
|
polgas
Canada1770 Posts
| ||
|
Nymzee
3929 Posts
On August 06 2018 12:21 acie wrote: all the games were draft wins, the computer came up with the best strategy/draft for that tiny 18 hero draft pool. It's impossible to counterpick from such a small hero pool that people don't have experience choosing from. Maybe every win could be attributed to draft? Maybe that's actually a thing, where 1 team could theoretically have 80%+ chance to win based off draft but because they're human (and make mistakes) they don't actually have that 90%+ chance to win that the bots were often able to achieve. | ||
|
FreakyDroid
Macedonia2616 Posts
On August 06 2018 08:42 Murlox wrote: True, but I believe the goal of the team is to make a self learning machine. Being extremely good at dota can come second. So maybe there are willing to accept that the machine is not optimal in some situations because it didn't have enough time to encounter those situations yet, rather than "help it get good faster" by feeding it data. But... maybe they'll talk about it on the panel. Team members have not been very technical yet, though, I found. The blog was much more useful. Well yeah, but there's clearly a pretty big gap in their learning process, it seemed as if they do stuff that directly rewards them, ie take tower, last hit creep, make a kill, prevent them from hitting my tower etc. Dota is way more nuanced than that, the reward doesn't always have to follow a few simple linear steps, it involves a lot of prediction/foresight something which these bots didnt have. In the first two games I think our casters were intimidated by their aggressiveness and spell usage and didnt know how to respond to that, but I'll bet my life on it that if they play another bo3 they'll crush the bots. They kept saying the bots learn 180 years per day, or rather the equivalent of it, but I think that's a very misleading number of years, its simply hours of game play of bots doing stupid things, but the actual learning they get from that is very little compared to what a human can learn in the same time. We learned in 3 games or 90 minutes how to deal with them. What Im saying is that is a bit dishonest from them to equate machine learning time with human learning time and that 180 years per day is not a good metric. | ||
| ||
