OpenAI's robots have just defeated human top professional Dendi in the Dota2 1v1 competition. OpenAI, which built the safe universal artificial intelligence, trained the robot from scratch by the "Self-Play" method.
Dota2 fall
After sweeping the top human chess masters and masters of Go, the computer now beats world-class pros in the world-famous video game Dota2. The previously undisclosed matchup took place at Valve's $24 million Dota2 International Invitational "The InternaTIonal."
Dendi (pictured left) communicated with the OpenAI team before the game.
In this one-on-one match, a robot designed by OpenAI defeated professional player Danylo “Dendi†IshuTIn, who has won a total of $73,544.94 in his previous career. OpenAI's robot defeated Dendi in the first game in ten minutes, after which Dendi quit in the second game and refused to play the third game.
Dendi
"Please let me go," Dendi said to the robot opponent during the game.
Elon Musk published a tweet to praise OpenAI's record and called it "the first robot to defeat the top players in eSports."
OpenAI beats top human players for the first time in eSports. This is much more complicated than games like chess and go.
OpenAI's CTO Greg Brockman said before the game that the robot was trained through thousands of guided self-defense, and that the robot had beaten several Dota 2 pros. Brockman said in the blog: "In the past few weeks, our robots have already defeated many top players including SumaiL (the world's top 1v1 player) and Arteezy (the world's top overall player).
The concept of “self-defeating†is the key to OpenAI development. This is an effective way for AI systems to learn to solve extremely complex tasks: against a weak or too strong opponent, it can't learn from it, but it is a valuable opponent. Brockman said: "You can see that the AI ​​has evolved from a completely random state to a top level today."
AI companies used to use e-sports games to test their technology, such as Google's DeepMind to capture "StarCraft 2", Microsoft's AI team today announced that they scored high points in the Pac-Man game.
OpenAI is not planning to stop the battle. They hope to participate in the official five-on-five competition next year. At the same time, they also released this robot, so anyone interested can play against it. The first player to beat this robot will receive a special reward from Valve.
We created the AI ​​that beat the world's top pros in the Dota2 1v1 competition. It is completely self-played and does not use imitaTIon learning or tree search. This is an important step in building an AI system that can accomplish well-defined goals in complex scenarios involving chaos and human behavior.
Dota 1v1 is a complex game with hidden information. Agents need to learn plans, offenses, tricks, and misleading opponents. The association between player skill and hand speed (per minute action) is not so strong. In fact, our AI "hand speed" is only the level of the average human player.
To win in Dota, players are required to develop their instincts and adopt appropriate strategies. In the above video, you can see that our robots have learned to predict the movements of other players through complete self-play, improvise in unfamiliar situations, and how to act on the player's unit.
Trained entirely through Self-Play, from zero to the world's top level
The goal of OpenAI is to build secure, universal artificial intelligence. For AI, Dota is an ideal test site with intense competition and complex games. If you want to achieve success in this game, you must advance the existing technology and achieve technological breakthroughs. We have now developed an AI that can beat top pros in Dota's 1v1 game.
Dota's rules are very complicated. If you use the rule method, then the built AI must be a crappy player. Therefore, we completely adopt the self-play training method. At the beginning of the training, AI had no knowledge of the world in which he lived, but only played with his own copy, which means that it and its opponents will always be equal. It uses this method to improve a little bit until it reaches the level of the world's best professional athletes. The InternaTIonal (TI) is Dota's World Championship, attracting 20,000 spectators from all over the world to watch professional athletes compete for 24 million prizes. At TI, we sent AI and many of them to compete. In the game, AI showed strong robustness and skill. Many professional players hope to continue playing with our AI and consider It is regarded as a means of training.
This research team is now.
The next step will be the 5v5 competition, and we hope to make five AIs that work together to defeat the human team. In addition, we also look forward to trying to put AI and humanity in the same team and achieve victory in collaboration.
Special equipment for university physics laboratory
Teaching Equipment,Optical Bench Kit,Optical Instruments In Physics,Electrooptical Modulator Experimental Instrument
Yuheng Optics Co., Ltd.(Changchun) , https://www.yuhengcoder.com