OpenAI, the non-profit research organization founded by Elon Musk, can boast a world first: its artificial intelligence system trained to play the complex strategy game Dota 2 beat a world champion e-sports today. The competition was held in San Francisco and was named "OpenAI Five Finals", putting an end to the public demonstrations of his organization. dotagame technology on a high note.
The competition on the human side included five top Dota 2 of the OG team, which won the world's most coveted eSports award last year, winning first place at The International, the first Dota 2 tournament with prizes now totaling $ 25 million. OG competed against OpenAI Five in a competition of the best of three, all trained using the same deep reinforcement learning techniques independently controlled by different layers of the same system. Reinforcement learning is actually a trial-and-error approach to self-improvement, in which artificial intelligence is dropped into the gaming environment without any understanding of how the game works.
Today's performance is by far the demonstration of the highest quality OpenAI Five capabilities to date, with the system losing two games against two less-performing e-sports teams last August. According to Greg Brockman, chief technology officer at OpenAI, OpenAI Five is improving by playing itself in an accelerating virtual environment. "OpenAI Five is optimized for deep learning, which means we have not coded it to play. We coded how to learn, "Brockman told the crowd before the competition. "In 10 months of existence, he has already played 45,000 years of Dota 2 gameplay. It's a lot, it's not bored yet. "
Dota 2 is an extremely complex strategy game, involving over 100 unique characters, skill trees and extensive item lists, as well as an impressive number of variables unfolding on screen at any time Match. As such, OpenAI imposes certain limits when its artificial intelligence system plays against professional players, mainly by limiting the number of heroes used by the two teams of five players.
In this case, each team had 17 heroes to choose from. OpenAI has also chosen the so-called "Captain" game mode, which allows each team to strategically ban heroes to prevent the other team from selecting those characters before using a different selection order. This allows the captain to build strengths between hero combinations and take advantage of the enemy hero's weaknesses through powerful scorers once the teams begin to complete the one-to-one alignment. Like previous games, OpenAI also disabled the Summoning and Illusion features, both of which involved additional variables in the form of copies of heroes and unique creatures for which OpenAI had not trained. his system.
Beyond, the game is played like a normal Dota 2 match, with the ultimate goal of destroying the "old" group of the enemy team or a great turn at the end of the territory of each team that becomes vulnerable only when the enemy team succeeds in destroying small towers throughout the match, between heroes – hero team is fighting.
In the first match of the day, OpenAI Five surprised OG and won with a number of aggressive tactics, including the special decision to spend money on the game to instantly revive the heroes. dead even at the beginning of the game. As OpenAI's Chief Technology Officer Greg Brockman notes, OpenAI loves strategies that drive short-term gains. It reveals its weaknesses in mastering the type of long-term planning on which humans are excellent and on which they usually rely to win. However, in this match, the first redemptions paid off and OpenAI Five gained an advantage that OG simply could not beat, the match being extended within 30 minutes.
This happens all the time in the test games: robots redeem, humans laugh, then humans lose. Hard to know if it will happen here too …
– Greg Brockman (@gdb) April 13, 2019
In the second match, OpenAI achieved an even better performance, gaining a quick advantage over OG in the first few minutes, and ruthlessly outsmarting the human players to win in just over half the time needed win the first match. Mike Cook, a passionate Dota 2 A player and spectator specializing in the fusion of artificial intelligence and game design, he pointed out that the unusually aggressive OpenAI Five had started playing in the second game, and that few OG players were succeeding. to fight against his advances on the map. In particular, Cook noted how OpenAI Five had benefited from his choice of specific heroes.
It's probably already finished, unfortunately. OpenAI has four of the top five ranked heroes by their net worth. At ten minutes, against the bots when running OpenAI, it's really a shame. #openaifive
– Mike Cook (@mtrc) April 13, 2019