An artificial intelligence has beaten one of the world’s top Dota 2 players in single combat today. Danil Ishutin, better known by his gaming handle “Dendi,” threw in the towel in the middle of a second game against a bot that OpenAI created, one that had been beating him handily.
The two squared off in a pair of 1-on-1 matches in this multiplayer online battle arena (MOBA) game at The International, one of the biggest esports events in the world. The first character who scored two kills or destroyed an in-game tower would be crowned the winner. In the first game, OpenAI’s bot appeared dominant, and Ishutin appeared flustered by its capabilities. It scored first blood against Ishutin fairly early on, and the two subsequently traded kills, crowning the bot the winner.
After the bot scored a kill on him early in the second game, Ishutin threw in the towel, meaning that OpenAI’s system remained undefeated against a group of top players, including Artour “Arteezy” Babaev and Syed “Suma1L” Hassan.
It’s a milestone for OpenAI, an organization that’s trying to make sure future artificial general intelligences are positive additions to humankind. Greg Brockman, the organization’s cofounder and CTO, said in a streamed interview that this validates a step toward more impactful systems.
“What we’ve built here is a general learning system, which is still limited in a number of ways, but it’s still capable enough to beat the best human pros at Dota,” Brockman said. “This is a step toward building more general systems which can learn more complicated, messy, and important real-world tasks like being a surgeon.”
The OpenAI bot learned how to play Dota 2 from running game after game against itself. The bot learned its behavior over two weeks of training, starting from random behavior and reaching the point that it could conquer pros.
“We’ve coached it to learn just from playing against itself,” said Jakub Pachocki, a researcher at OpenAI. “So we didn’t hard-code in any strategy, we didn’t have it learn from human experts, just from the very beginning, it just keeps playing against a copy of itself. It starts from complete randomness and then it makes very small improvements, and eventually it’s just pro level.”
This doesn’t mean that the bot could go out and conquer the larger game of Dota 2 just yet, though. The contained 1-vs.-1 exhibition helped reduce the usual complexity of a match, which typically involves two teams of five players each working together to complete a variety of objectives that lead toward victory.
OpenAI said in a blog post that its next step is to build a group of bots that can compete against and with a team of top human players.
This is the latest in a string of gaming victories by machine learning systems. Earlier this year, DeepMind’s AlphaGo bested the world’s best Go Player in a best of three series.
There’s more gaming to come: Blizzard and DeepMind (which Google owns) just released a set of tools designed to let machines compete at the real-time strategy game StarCraft 2.