A research lab named OpenAI, co-founded by Elon Musk, has announced its latest achievement: a team of AI bots that are gunning down human Dota 2 players and winning.
OpenAI has been pitching its AI bots against professionals players in 1v1 matches since last year, but such games greatly reduced the actual difficulties of Dota 2.
So for the first time, a team of upgraded AI players, ‘OpenAI Five,’ entered the battleground in 5v5 matchups against amateur Dota 2 human players which requires more coordination and long-term planning.
For now, they managed to win the game. However, they are yet to challenge Dota 2’s very best players. But the AI agents will be doing so later this year at the Dota 2 world championship, The International 2018.
The motivation behind this research is straightforward: a complex video game like Dota exhibits the messiness and continuous nature of the real world. If AI systems manage to learn the skills required to play such a game, it can be used to solve real-world challenges too.
To achieve this, OpenAI took the Reinforced Learning approach where the AI bots are thrown into a virtual environment. Here, they teach themselves how to complete their objective through trial and error instead of feeding them data accumulated by humans.
The developers set what they call “reward functions” where bots are given points for killing an enemy and leave the AI bots to play with themselves over and over again.
OpenAI Five runs on more than 100,000 CPUs, and its neural networks perform around 150-170 actions per minute. Its average reaction time is 80ms, which is faster than humans.
In a matter of few hours, the bots end up playing more games than a human in its lifetime. The AI plays up to 180 years’ worth of games against itself every day, gaining ample opportunity to learn.
But winning Dota isn’t the ultimate goal. “Our underlying motivation reaches beyond Dota,” writes OpenAI. For now, it will be worth watching whether these AI bots can beat the top human players at the game.