In recent years, OpenAI launched the mission "artificial general intelligence benefits all humanity", has developed a machine-learning bot to play Dota 2 the largest game in the universe. Starting from a very slimmed down version of the entire game, the bot has been developed over the years by playing millions of millions of games against itself. It was not only learned how to play the five-on-five team, but how to win. resistant.
We were able to follow the development of the bot in a series of show matches, each using a more complete version of a game and more skilled human opponents. This culminated in the presumably final game of the show over the weekend, when OpenAI Five was played in a best-of-three match against OG, the team that won the biggest competition in all esports games last year with The International. [1
The games can be viewed here. The first game even looked like about 19 minutes. The humans had a small gold advantage, but the bots had better territorial control. The bots advanced in a team fight, killing three human players while losing only one player. The game still looked like a knife, but the bots disagreed: they said they had a 95 percent chance of winning and when they made that statement, they immediately took advantage of it to inflict serious damage to the human base. This further improved their territorial control and also gave them a significant lead in gold.
This turned people on their backs, and although they managed to pull the game out for another 20 minutes, they did not manage to break the bots' leadership, giving OpenAI a 1-0 advantage.
Things were not even close in the second game; The bots took the lead early and broke through the human base within 15 minutes. Seven minutes later, they took the victory.
Overall, it was a dominant achievement from OpenAI: a 2-0 victory over an established human team used to playing together at the highest level the game has to offer. This performance was by far the most powerful feature of OpenAI.
Coordinating the bots is scary: although they can not communicate, all five computer-controlled players think in the same way. If you think that this is a good opportunity to attack a human player, the other four will think the same and join the attack. This gives the appearance great coordination in team fights – coordination with a precision and accuracy that human teams can not match.
OpenAI, however, looks beatable. There are clear, albeit surprising, weaknesses – it's not very good to score the last few hits, the fatal blows at computer-controlled units that accumulate gold in the game. This gives people the opportunity to gain an early gold advantage. The bots also had difficulty counteracting invisibility on the human side. They also seemed to be ill-adapted to certain heroes' spells, especially Earthshaker's Fissure, a spell that temporarily creates an insurmountable barrier on the map. People were effective at catching bot players and limiting their movements, and this seemed to confuse OpenAI.
The behavior of the bots is also a lesson in the big gap between this machine learning system and a complete system of general artificial intelligence. While AI Five is apparently successful in winning games, it obviously does not know how to play Dota 2 . Game human players use a technique called "Drag" to divert the flow of their side's computer-controlled minions (known as Creeps in Dota 2 ) to deny the opposing team both gold and experience , Human players can recognize that this has happened because crawling does not occur at the right time. Human players have a mental model of the entire game, an understanding of their rules, and therefore can see that something is wrong. They can think about where goose bumps must have gone and hinder the train. In contrast, in this scenario, the computer wanders around aimlessly.
In his millions of games played against himself, OpenAI seems to have never picked up on the traction technology he learned to play. So when a human team begins to pull, the bot does not recognize the situation and does not really know what to do. It can not think about what the game should look like and it can not be speculated about why the game behaves unexpectedly. The bot can only search for patterns it recognizes and select the action that most likely leads to the best result. Give it a pattern that it can not recognize and its performance worsens.
So far, the OpenAI bot has been restricted; some pros and streamers were allowed access to the game, and some live events also played against them. But for a few days that will change: Dota 2 Players can log in here to play against the bot or with him for three days. Unfortunately, this public period does not look like it would lead to a new and improved bot: beating a top-notch human team was the goal OpenAI set for its bot, and that's the end of the experiment.