OpenAI Five beats Dota 2 world champions OG

Background

Dota 2 is a five-a-side team game in which two sides of heroes fight over a map, and each match runs for tens of minutes. At any moment a player sees only part of what is happening, the number of possible actions is enormous, and a decision made early can pay off much later. Those properties — long time horizons, imperfect information and a very large action space — are what made the game an interesting target for AI research.

In August 2018 the bots, called OpenAI Five, appeared at The International, the game's best-known tournament, and lost both of the games they played against professional teams there ^[1]. The system learned entirely through reinforcement learning, without imitating human play: it improved by playing against copies of itself, and the version at The International in August 2018 had played about 10,000 years of self-play over 1.5 realtime months ^[3]. In its 2018 version the system was playing roughly 180 years' worth of games against itself per day ^[1].

The match

On 13 April 2019, in San Francisco, OpenAI Five played a best-of-three series against OG, the reigning champions of The International 2018, and won it ^[1]. The opponent was therefore not a casual one: it was the team that had most recently won the biggest event in the game.

Key details

The training did not stop between the two appearances. By April 2019 the total amount of self-play had risen to about 45,000 years of game time accumulated over 10 months, an average of about 250 years per day ^[3]. That scale — many lifetimes of human play compressed into months of wall-clock time — is the practical reason a reinforcement-learning system could reach a professional level at a game this complex.

The public arena

After the match, OpenAI held a four-day public online event in April 2019 in which anyone could queue up against the bots. According to Wikipedia's account, the bots played 42,729 games during the event and won 99.4% of them ^[1]. The paper describing the system reports the same win rate for the arena held on 18–21 April 2019, where the bots met 3,193 teams across 7,257 games ^[2]. In other words, the win against the world champions was not a one-off: against a much wider field of ordinary players the system kept winning almost every game.

Why it matters

The paper describes OpenAI Five as the first AI system to defeat the world champions at an esports game, referring to the win over Team OG ^[2]. That claim is narrower than "beat humans at video games" — it is about a specific, well-defined contest against the best team in a popular competitive title — but it is exactly the kind of milestone that is easy to check and hard to argue with.

The technical account of the system was first submitted to arXiv on 13 December 2019 ^[2], so the methods behind the run were published in detail rather than only demonstrated in a exhibition. For readers following AI, the sequence matters: a loss to professionals in August 2018, a win over the world champions in April 2019, a public arena with a 99.4% win rate, and then a paper at the end of that year.