OpenAI, the independent research institute that was co-founded by Elon Musk in 2015, will send its Dota 2 bots to The International 2018. There, the AI team will take on a professional team in a 5v5 match. And it plans to win.
Known as OpenAI Five, the bot team has taught itself the nuts and bolts of Valve's free-to-play MOBA by playing 180 years' worth of games against itself every day. OpenAI says this process requires 256 GPUs and 128,000 CPU cores—and is a scaled up version of the "much-simpler" variant that toppled pro player Danil "Dendi" Ishutin in 1v1 at last year's TI.
"Our team of five neural networks, OpenAI Five, has started to defeat amateur human teams at Dota 2," says OpenAI. "While today we play with restrictions, we aim to beat a team of top professionals at The International in August subject only to a limited set of heroes. 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 $40M prize pool."
This blog post explores the myriad challenges and obstacles the OpenAI team faces while brining OpenAI Five up to speed—not least Dota 2's complex rules. Read the post in full via the link above, and watch the following short which examines some of the learned behaviours the bots have picked up along the way below.
These include teamfighting, value prediction and, rather amusingly, ganking. Here's that:
And here's Dendi's 1v1 defeat to an AI opponent at The International 2017, which kicks off around the 7.30 mark. The pre-match testing segment is worth watching too if you've time—I was particularly tickled by the chap who says "the not being able to kill it part is so annoying" in reference to the AI's skills.
