This is how AI defeats the best Dota 2 players

On August 6, 2018, Elon Musk's OpenAI artificial intelligence was officially defeated, five veteran Dota 2 players including Blitz, Capitalist, Fog, Merlini and MoonMeander. Earlier, 5 OpenAI bots (also known as OpenAI Five) also won against 5 players with a high rank of 6.500+ MMR (top 0.05% of the world).

At 'The International 2017' - Valve's tournament to select Dota 2 world championships, OpenAI was first introduced to the public. And in this tournament, OpenAI defeated all the best professional gamers of the time, including Canadian Dota 2 Artour "Arteezy" Babaev, in the 1 vs 1 Shadow Fiend format.

Dendi when fighting against AI in 2017.

Dota 2 is a complex game, requiring players to have a deep knowledge of all aspects of the game, to guess the opponent's decision-making abilities. To win, 5 members of a team must work together. When playing Dota 2, OpenAI must also act, support each other and judge themselves as 5 normal players. Perhaps this is exactly why OpenAI chose Dota 2 as a destination, not any other game.

In 2017, Bot only knew 1 vs 1 confrontation and ran on Azure platform with 60,000 employees, 5 'OpenAI Five' Bot of 2018 were developed on the 128,000 Google Cloud platform. Each day they confront themselves in a game volume of 180 years. In the process, the Bot will learn from its previous version. And after thousands of matches, an artificial intelligence with the ability to calculate every aspect of the game and find the factor that increased the winning rate was formed.

Currently, OpenAI can only play with 18 heroes (generals) out of 115 generals in Dota 2 but that is enough for it to win 5 veteran players representing 99.95% of players worldwide.

In a recent match, a BOT representative of the AI ​​team made the comment 'We estimate the winning rate to be over 80%' as soon as we enter the game. And just a few seconds later, predicting their winning rate has increased to 90% and 95%.

In fact, right from the selection of a prime minister - one of the most difficult things in Dota 2, OpenAI calculated his winning ability. OpenAI can make the decision to choose the best team possible because the BOT will learn up to 11 million abilities to combine heroes together.

As a result, OpenAI easily defeated people in the first two games in less than 40 minutes.

In the match, they constantly create pressure, focusing primarily on specific goals that make the human team fall into a passive position. Normally, in a Dota team the 2 most farmed positions are the 'carry' position (most likely to cause damage in the team) but OpenAI does something completely different. It gives 'support' (farm position) to farm first, even more gold and experience than carry.

AI gives a winning rate through each turn of ban / pick of the form team.

As soon as he saw the opponent, OpenAI took a continuous move to put pressure to draw the opponent's blood from a long distance. This makes it easier for AI to finish off the opponent and offers many advantages for his team when a 5-on-5 fight breaks out.

However, OpenAI still has the weakness that is the ability to last-hit (to throw a finishing blow on soldiers for a very low amount of gold). They often skip this stage or if there is a blow, it is not very accurate.

After the first two games, OpenAI was given 5 extremely disadvantageous heroes, without any slow, stunnering, time-consuming moves. And as soon as the game started, OpenAI predicted its winning rate of only 2.9%. However, it takes more than 30 minutes for the human team to defeat them.

In addition, at some point of the game, BOT Slark (a very weak hero in the early stages of the game) of Opne AI still shows up to be destroyed, but does not know to escape from the opponent and go to farm.

See more:

1. Microsoft artificial intelligence application can admire the poetic landscape, invite you to enjoy
2. Google artificial intelligence will redraw your image in a new, strange way, try it out
3. Train AI with Ninja game do you want to try?
4. AI can fake anyone's voice in just a few minutes