This AI trains on 180 years of gaming every day, and is now beating humans
Artificial intelligence is getting better and better at complex gaming. The OpenAI initiative backed by Elon Musk is now regularly beating teams of five human gamers at Dota 2, its developers have announced, and will be taking on some of the best esports players in the world in a tournament in August.
OpenAI beat human opponents at last year's The International event, but in that case it was only taking on single opponents under some tight restrictions. This time around, it's battling as a team of five in the full version of the game.
OpenAI Five, as the new AI system is called, has already claimed victories against amateur and professional teams, thanks to a daily training regime in which it crunches through around 180 years of gameplay. Playing against itself, the bot learns the best strategies through trial and error, a process known as reinforcement learning.
"Relative to previous AI milestones like chess or Go, complex video games start to capture the messiness and continuous nature of the real world," says the OpenAI team in a blog post. "The hope is that systems which solve complex video games will be highly general, with applications outside of games."
Dota 2 is a real-time strategy game with a sophisticated set of rules, pitting two teams of five players (each with different skills and strengths) against each other at opposite ends of a map. The aim is to capture the opposition team's territory, and to destroy a structure called the Ancient.
Those elements make the game much harder to grasp for a computer than something like Go, which AI is already unbeatable at. With the help of 256 GPUs and 128,000 CPUs to process the learning process and repeated gameplay, the OpenAI Five system is learning fast.
The underlying AI deploys a finely weighted set of balances and measures to decide how to prioritize individual bot player progress against the good of the team. It also has to calculate when to experiment with its strategy as the game evolves, and when to stick with what has worked in the past.
Another balance needs to be struck between short-term and long-term gain, and the AI is improving at getting the combination right. At any given moment, the researchers say, OpenAI Five's players have to choose between around 1,000 possible actions.
With the system now regularly (but not always) beating amateur and semi-pro teams in testing, the challenge is to take on the best of the best at The International, and perhaps even walk away with some of the US$15 million prize money.
Certain elements of Dota 2 have been turned off during testing, and the OpenAI Five system is limited to using five specific characters out of 113. What's more, its bots can see the whole map at once, something which human players aren't able to do. Even given those caveats though, OpenAI Five is making impressive progress – and we fancy its chances at the Dota 2 tournament in August.
You can check out a video of OpenAI Five in action below.