OpenAI's machine learning bots defeat Dota 2 champion team
A team of AI agents created and trained by the artificial intelligence startup OpenAI has taken down the reigning Dota 2 The International champions Team OG.
It’s the latest in an ongoing trend of pitting trained AI agents against professional esports teams to test their mettle, and one that showcases both the effectiveness of video games as AI training tools and how the machine learning tech itself is progressing.
ArsTechnica has a writeup and analysis on exactly how the weekend showdown transpired, but in short a team of five machine-learning-driven bots dubbed the OpenAI Five faced off against Team OG in a best-of-three match. Both of the matches—the OpenAI Five went 2-0—can be viewed in their entirety over on Twitch, complete with live commentary throughout each game.
Following the match, OpenAI has opened up sign-ups (for a limited period of time) to allow Dota 2 players to take the bot on for themselves, an opportunity that some AI-minded game devs might be interested in checking out.
The bots themselves have been gradually trained across years by playing millions of matches against themselves, and had a few limitations compared to the team of human players. While a five-person team playing Dota 2 would typically communicate with each other throughout the match, the OpenAI bots were treated as five independent players without the ability to communicate amongst each other. The bots also saw their reaction times restricted to realistic levels, and was only able to use a predetermined set of characters and items.
As Ars points out, there are other differences in technique between the human and AI players that become apparent as the matches go on, such a lack of understanding of standard MOBA techniques like pulling or last-hitting minions.