OpenAI Five beats the Dota 2 champions
On 13 April 2019 OpenAI Five, an OpenAI program that controls five heroes and was trained by reinforcement learning, won two games in a row against OG, the reigning Dota 2 world champions. A preprint of 13 December 2019 summed up the training: one continuous run from 30 June 2018 to 22 April 2019.
Why it matters
By OpenAI's account this was the first win of a program over world champions at an esports game, and the first over professionals on a livestream. OpenAI put the result down not to a new algorithm but to scale: an existing reinforcement learning method stretched over ten months and hundreds of petaflop/s-days of compute.
The authors named two limits on the game: 17 of 117 heroes were available, and items that let a player control several units at once were removed. On 18-21 April 2019 the agent was opened to all: 7,257 games against 3,193 teams, 99.4% won, but 3,140 of the 7,215 wins were games people abandoned, counted as wins. The preprint gives 770 +/- 50 petaflop/s-days and over 150 million parameters; the same maker's April release gives 800 petaflop/s-days and about 45,000 years of self-play. The same release recalls that OpenAI Five lost to professionals at The International 2018.