He claims it's "vastly more difficult" than chess or go.
Of course, humans say that they will still dominate 5-vs-5. Yeah, right.
https://www.businessinsider.com/the-international-dota-2-openai-bot-beats-dendi-2017-8He claims it's "vastly more difficult" than chess or go.
Of course, humans say that they will still dominate 5-vs-5. Yeah, right.
https://www.businessinsider.com/the-international-dota-2-openai-bot-beats-dendi-2017-8Of course bots will beat humans, when they don't even need reaction time for their calculations. Having played Dota 2, this is like having 0 ms start time in the 100m for every action. Add up the amount of reactions needed, in a game, and suddenly the bot is way ahead and more efficient.
Wonder how far aibots will go-perhaps they could be used in battle planning outside of video games. Interestingly this bot "learned" all its judgments, probably through multiple simulations, like the Google AI that runs. 5v5 would be more unpredictable since there is that much more strategy involved and possiblities. I would think humans have a better chance because of that, but realistically an AI that can learn different strategies in a short amount of time will probably beat the human brain. That doesn't mean that they can't mess up--even the top Kenyan runner messed up the first water jump today, with so much "learned" steeple experience.
The hard/insane bots have very quick reactions and chain stuns etc but they are easy to beat because they don't have tactics.
I'd still fancy a pro team.
The game's very hard bots (5v5) uses cheats to regain health or make skills become available faster. Unfortunately they are hard-coded bots, so performing an event that confuses the bots will lead to victory. I'm not sure if a learning AI will be able circumvent this problem--we will assume it does, since the 1v1 bot was able to grow out of dead-end thinking.
The 1v1 learning takes about 10 min per match. A team of 10 people programmed this bot, and who knows how many matches it needed to play to achieve such a master level. They could have been working on this bot for a year. 5v5 ranges from 20 to 70 min per match. The time it would take for a team of bots to be proficient would take 2 to 7 times longer than a single bot, based on time of match and assuming the same number of trials will achieve such a high level. I doubt that 10 people is enough to create a beast; they would probably need a team of 50-100 behind the scenes to create super bots. We have not seen 5 self-aware bots coordinate something. I bet they will run into some messed up confusion and be looped not being able to determine a decision. Look at the Youtube video of 3 AI talking to each other; someone managed to get Siri, Alexa, and the Google Talk bot to be in an infinite loop. Now, if it was 1 bot controlling all 5 characters rather than 5 individual bots, I would bet on the 1 bot.
Even so, I'll still bet on humans. Humans have outsmarted robots in the past with creativity. These bots have been learning by throwing eggs at walls and seeing which don't crack, rather than tholinking critically. Of course bots will always be better in terms of raw numerical calculations, but we have not seen a bot with the ability to derive Einstein equations yet.
You say "time" to proficiency, but really it is energy and/or cost. Just use 10x or 30x the computing power. Hire a wider team, etc. AlphaGo could have been done some years ago, if academia had the problem-solving model that DeepMind brought (and academia has plenty of resources, essentially gov-funded to near infinity, but just distributes them poorly, more involved in keeping people employed than making progress). But all the advances were coming from Facebook/Google, etc., due to wacky incentives.
An actual infinite loop is trivial to avoid, but making a random decision might indeed be a tricky consequence to deter.
That's largely because no one has tried to program a bot to propose physical laws, based upon observational data. One could start with Kepler's, I suppose, as ancient Greek geometry (conics) is pretty well-understood at the automated proof level.
When can we see this in League of Legends? Because nobody cares about DOTA 2.
I just check that match there https://hawkbets.com/ and unfortunately, pro teams don`t win. But i know that some guys (non-pro) get lucky to win it, not let the OpenAI to take pick that the almost every time build. You can also find this match on hawkbets.
This is the difference between Dota and other games. I really like how this game is made, namely its ideas. So you can't play with a bigger team than yours, or you can't play solo. This is a very important fact for any game worth playing. I like Dota what is possible to update the account. I know a good company that deals with this (https://dota2-boost.com). Has anyone tried this? How are your accounts improving?
I don`t like boosters. This AI works with the database of pro gamers but it still buggy. Just see some official pro games https://hawk.live/ and see that AI doesn`t learn. It just repeats the moves and doesn`t try something new. I think it`s fake info!