Man vs machine in Dota 2

Machine learning and deep learning are two expressions of artificial intelligence (AI) that have been widely written about. But, another approach to AI called reinforcement learning is coming to the fore. One of the first applications of reinforcement learning is in training AI bots to beat humans in the highly popular game Defense of the Ancients 2 (Dota 2).

OpenAI Five playing a top employee team.
OpenAI Five playing a top employee team.

Researchers at OpenAI have trained AI bots to beat the top one percent of amateurs at Dota 2 — that’s no mean feat considering that it’s not a mere one-on-one but a complex five-on-five game requiring teamwork and game strategies.

Founded by Elon Musk and Sam Altman, the research lab hopes to pit its team against top players at The International, the biggest Dota 2 event on the e-sports calendar in August.

“We may not succeed: Dota 2 is one of the most popular and complex esports games in the world, with creative and motivated professionals who train year-round to earn part of Dota’s annual US$40m prize pool (the largest of any esports game),” noted the OpenAI blogpost.

OpenAI Five plays 180 years worth of games against itself every day, learning via self-play. It trains using a scaled-up version of Proximal Policy Optimisation running on 256 GPUs and 128,000 CPU cores — a larger-scale version of the system built to play the much-simpler solo variant of the game last year.

Whatever the outcome in August, AI has certainly come a long way from beating humans in static games such as chess to competing in the highly-complex Dota 2.