After Chess and Go, AI is now confronting the best videogame players in the world.

OpenAI Five training on Dota 2:

Dota 2 is a real-time strategy game played between two teams of five players. Each player controls a ‘hero’ and must defend its base by collecting experience, items and by winning fights against the opponents.
Dota 2 is known to be a very complex game, hundreds of hours are needed to understand the multiple factors.
OpenAI Five (Five) is a team of five artificial intelligence players, created by a research company presided by Elon Musk and Sam Altman. As the team is really important in Dota 2, each bot must not be selfish but have to “care about their teammates so they can play together as one unit” says Christie Dennison, a machine learning engineer at Open AI.
Bots originally integrated in Dota 2 are programs that are told what to do in a specific situation. But the AI team learns from scratch by itself like a real human being. The technology is based on self-play learning, Five plays nothing less than 180 years of games against itself every day!
The interest is that, unlike Chess or Go, Dota 2 is not a full-information game. The map is not revealed to the players, except the area around them and their allies. The strategy of the opponent remains secret, so the team must play with incomplete data.

The results: will AI surpass a human being?

In 2017, Five had already beaten a bot, but this year they wanted to beat a real human. The first challenge was to defeat a team of the best non-professional players. In August 2018, 3 games occurred: the first two were won by Five and the last one by team Human. AI: 2 / Human: 1. Humans are defeated.
The second challenge led Five against two professional teams, which means ten of the best Dota 2 players in the world. If they cannot defeat AI, who will? Indeed, we won’t be surpassed by robots this time. In game 1, Five lost against paiN Gaming after 51 minutes (games usually last 45 minutes). Then, Five was also defeated by a Chinese team after 45 minutes.
But the gap is huge regarding the capacity of learning of an AI and a human. Blitz, an ex-professional Dota 2 player and coach, confesses “It took me 8 years to learn the strategy that the bot was intuitively doing”.


AI is developing really fast, it can beat humans in complex games but can always be improved. The researchers learnt from those matches. They know what can be improved and what is working well. Even if it is challenging to confront humans with the machine, the goal is not really to see who will win but how far we can push the ‘intelligence’ of the machine and make it capture the “nature of the real world”. According to the researchers, the purpose is that these systems “which solve complex video games will be highly general, with applications outside of games”. AI will open new opportunities for us. For example, Five developed new strategies in the game that were unknown for the players.
To go further:

A propos de Léa VERMERSCH