At OpenAI, we’ve used the multiplayer video game Dota 2 as a research platform for general-purpose AI systems. Our Dota 2 AI, called OpenAI Five, learned by playing over 10,000 years of games against itself. It demonstrated the ability to achieve expert-level performance, learn human–AI cooperation, and operate at internet scale.
November 9, 2016
First commit of OpenAI's Dota 2 project.
March 9, 2017
August 11, 2017
1v1 Shadow Fiend bot beats top professional Dota 2 players at The International 7.
September 7, 2017
February 28, 2018
First 5v5 results: RL agent beats OpenAI scripted bot at tower minigame.
April 3, 2018
RL agent beats in-house OpenAI team at net worth minigame.
June 6, 2018
RL agent defeats in-house OpenAI team at fairly restricted 5v5.
June 30, 2018
OpenAI Five’s parameters initialized.
August 5, 2018
OpenAI Five defeats popular casters at the Benchmark in front of a live audience and 100k livestream viewers, with somewhat restricted 5v5.
August 9, 2018
August 17, 2018
August 22–24, 2018
August 26, 2018
More model capacity: OpenAI Five’s long short-term memory (LSTM) size is doubled and number of parameters quadrupled.
October 5, 2018
December 10, 2018
January 16, 2019
February 1, 2019
April 5, 2019
April 13, 2019
OpenAI Five wins back-to-back games versus Dota 2 world champions OG at Finals, becoming the first AI to beat the world champions in an esports game.
You play against [OpenAI Five] and you realize it has a playstyle that is different. It’s doing things that you’ve never done and you’ve never seen. Sometimes it looks extremely silly. But then again, are you going to be human and be like "Hey, this looks very stupid, this is bad" or [do] you try to take it to next steps, like "Why is it doing this?"
One key learning that we took is how it was allocating resources. It’s just allocating resources as efficiently as possible. And then you realize that we’re guilty of being stuck in a team dynamic, whereas sometimes we have to be way more flexible. […] If OpenAI does that dynamic switch at 100%, we maybe went from 5% to 10%? But that is already a difference—we’ve noticed it.
April 18–21, 2019
July 12, 2019
Finished training of a new agent, Rerun, which reached a 98+% win rate against the agent that played at Finals. This was completed in 2 months, without surgery, and while utilizing only 20% of the resources.
August 25, 2019
I don't believe in comparing OpenAI Five to human performance, since it's like comparing the strength we have to hydraulics. Instead of looking at how inhuman and absurd its reaction time is, or how it will never get tired or make the mistakes you'll make as a human, we looked at the patterns it showed moving around the map and allocating resources.
In terms of what OpenAI has done for us and how it influenced our run at TI9, one of the many curious patterns was the buyback and pressure play that happened in most of the games. We had a lot of talks about fighting and pressuring and how it used a different approach from any human in the past. As people, it's about being realistic and learning from the brain of the AI and not the hydraulic strength that machines have.