- An AI opponent beat a champion Dota 2 player at last year’s championships
- Has been training for a five-on-five bout at this year’s main event
- Plays the equivalent of 180 years’ worth of Dota 2 games every day
Elon Musk is not on the board of his non-profit AI watchdog OpenAI any more, but he still loves watching its bots outlearn humans.
At last year’s Dota 2 tournament, The International, a bot from OpenAI made a surprise appearance for three one-on-one exhibition matches against pro gamer Danylo “Dendi” Ishutin. It beat him in the first, Dendi resigned from the second match, and declined to play a third.
Musk hailed the victory on Twitter:
OpenAI first ever to defeat world's best players in competitive eSports. Vastly more complex than traditional board games like chess & Go.
— Elon Musk (@elonmusk) August 12, 2017
His claim that Dota was “vastly more complex” than chess and Go, where AI had in the past made headlines by beating the world’s best human competitors, was based on the fact that to win Dota, a bot can’t simply sit and ponder the next best move.
The games Open AI Five – the current iteration – train for produce 80,000 frames of action in 45 minutes. Stop moving, and you’re dead.
But one-on-one victories are obviously less complex for an AI bot to figure out than the teamplay games that millions of fans tune in for at the Dota 2 world championships.
And after last year’s ambush, OpenAI CTO Greg Brockman promised the company would be ready to take on teams in proper five-on-five matches this year.
It’s just published a post hosted by machine learning engineer Christy Dennison which outlines seven behaviours OpenAI Five has learned in preparation for Dota 2 teamplay.
To do that, Open AI Five has been working as a team of five distinct neural nets, and to train, has been playing against itself.
All up, it has completes about 180 years’ worth of Dota 2 games every day.
So, is OpenAI Five ready? You decide.
See the graph at the bottom left? That’s OpenAI Five checking its chances of victory in real-time as it shifts tactics during a fight:
Here the human Crystal Maiden flees into the trees and hopes to hide:
But OpenAI Five knows it will find her eventually by tracking the most likely escape route. Death follows swiftly:
In the top right, watch OpenAI Five deliberately use a character to lure a sniper and his mates back toward his tower:
Then watch OpenAI Five move in behind to finish the team off:
OpenAI Five is struggling to take a human base, so when one bot character gets low on health, it makes a break for it. The human team abandons its base to follow and finish the bot character off:
And OpenAI Five moves in to take the barracks:
There are more examples of OpenAI Five learning to focus all its energy on one outcome, hunting down important opponents and fighting as a team in this video.
While these are tactics most reasonably competent humans employ, remember, OpenAI Five learnt these just by playing itself. There was no human interference in the learning process at all.
It will undoubtably be the showcase event at this year’s The International.
The International main event will be held on August 20 at the Rogers Arena in Vancouver, Canada. The prize pool is currently at the $US15.5 million mark, tracking just behind last year’s total at the same time, which went on to grow to a record $US24 million pot.