A bot for DOTA 2 can now defeat entire teams. To do this, it learns a lot… namely a proud 180 years per day.
The OpenAI project is making waves again. After the self-learning artificial intelligence managed to defeat Dota professionals in “1 against 1” almost a year ago, the system is becoming increasingly sophisticated and smarter. Currently, the “OpenAI Five” already competes against 5-player teams, although so far only against amateurs.
Millions of matches per day: To improve, the self-learning AI must collect data. This is done by competing against itself repeatedly in simulations. Per day, it plays as many games as fit into 180 years – an almost unimaginable amount. 80% of the games are played against itself, and 20% against older versions of itself.
They’re not even playing DOTA, critics say: Although most players show impressiveness at how the bots are developing, there are also objections. Many players criticize that the bots do not actually play real DOTA 2 because they operate under too many constraints. The bots only play in “Mirror Matches”, meaning they always face identical heroes, wards cannot be used, summons and illusions are also not allowed. Some items are also banned, as is invisibility. These are conditions that would never occur in a real match, hence the results should be taken with caution.
From the independent development of OpenAI Five, they hope to gain important insights for the general creation of AIs. Because the researchers believe that “AI has the potential to be the most positive thing ever created by humanity.”
On July 28, a live match of OpenAI Five against professionals will be held at the big “The International” tournament of DOTA 2. How the AI performs against the professionals will be revealed then.
What do you think of OpenAI and the successes of these bots? Are they still just at the beginning? Or is it already impressive what these programs learn on their own?
