OpenAI’s Dota 2 bots win over team with mostly professionals

Spread the love

A team made up of five bots trained for the game Dota 2 are pitted against a professional team of human opponents. The bots won the best of three by convincingly winning the first two rounds.

Christy Denninson of OpenAI announced that the team with bots, called OpenAI Five, won the first two matches and thus also the best of three. The human opponents team, made up of former pros and a Dota 2 commentator, did win the third and final round, which was actually played purely for fun. Prior to this round, the public chose which heroes the people could join in the battle.

The first match was a simple win for OpenAI Five, with the human opponents unable to destroy a single defensive tower. In the second round things went a little better and a tower was taken, but they didn’t get much further. The OpenAI Five matches took place as part of The International, the annual, prestigious Dota 2 tournament.

However, some restrictions were placed on the bots’ matches. For example, Dota 2 players can normally choose from more than 100 different heroes, but that number was reduced to 18 for OpenAI Five matches. fairly close to the reaction time of human players. However, according to the makers, this did not lead to changes in gameplay; they say the bots’ strength lies in teamwork and coordination, rather than quick reflexes.

OpenAI Five actually consists of five neural networks. The bots have been trained for several months by playing against themselves every day; in an accelerated manner, that was good for 180 years of playing time every day. With the help of 128,000 regular CPU cores and 256 Nvidia P100 GPUs, the bots were trained, learning through trial and error which actions yielded maximum virtual rewards.

It was already clear that OpenAI Five could score high. For example, the team has already won several times in matches against semi-professionals. Last year, OpenAI had already developed a bot that could play Dota 2 and even beat a world champion like Dendi, but this was still under quite restrictive circumstances and in a one-on-one situation.

You might also like