Dota2 OpenAI: captured musk end abuse the strongest human robot game player

Home > Sci-Tech

Dota2 OpenAI: captured musk end abuse the strongest human robot game player

2017-08-12 22:52:38 206 ℃

The new report Zhi Yuan

Editor: Bear laugh

The new guidance Zhi Yuan]OpenAI robot has won the top human occupation in Dota2 1v1 game Denti game player. In general artificial intelligence construction safety responsibility of the OpenAI, through the "Self-Play" mode, starting from zero training out of the robot.

The fall of Dota2

After sweeping the top human chess master and chess master computer, now in the worldwide electronic game Dota2 victory over the world class occupation game player. Had not disclosed this battle occurred in the Valve $24 million bonus Dota2 International Invitational "The International".

In a one-on-one battle, the OpenAI design of a robot defeated occupation game player Danylo "Dendi" Ishutin, Dendi has won a prize of $735449.4 accumulated in the previous occupation career. The OpenAI robot in the first battle in ten minutes after defeating Dendi, Dendi in the second inning of war and refused to fight the third exit.

"Please let me go," the match between Dendi of robot opponent said.

OpenAI for the first time in the field of E-sports human beating top game player. This is the international chess and go game is much more complicated.

OpenAI CTO Greg Brockman in the game called it a robot trained by thousands of times to guide the self confrontation, and that the robot has defeated a number of Dota2 occupation game player. Brockman said in a blog: "in the past few weeks, we have overcome the robot including SumaiL (top player in the world (Arteezy 1v1) and overall's top players in the world) and many other top game player".

"Self confrontation" concept is the key to the research and development of OpenAI. This is a AI system to study the effective method to solve the complicated tasks, and too weak or too strong opponent of war, it can learn from it, but it is their own valuable rivals. Brockman said: "you can see the AI from a random state of complete development step by step to the top level today".

AI used to use electronic games to test their technology, such as Google DeepMind to capture the "Star 2", the Microsoft AI team today is that they get high score in the pac man game.

OpenAI did not intend to stop in the footsteps. They hope that the next part of the formal five to five games. At the same time they also released this robot, so anyone interested can be with the battle. The first beat of this robot game player you can get Valve dedicated.

We are created to be able to beat the world top occupation players AI in Dota21v1 games. It is completely self-play the way of training, without the use of imitation learning (imitation learning) or tree search. The construction of AI system can complete well-defined goals in complex scenes of chaos, including human behavior, is an important step in the.

Dota1v1 is the game contains hidden information complex. The need of learning plan, attacking, tricks and misleading the opponent. Player skill and hand speed (action per minute) the association is not so strong, in fact, our AI "hand speed" is the general level of human players.

You want to win in the Dota, require players to cultivate the opponent's intuition, and take corresponding strategies. In the video, you can see that our robot has been completely through self-play to predict the other players in the action, and play improvisation in unfamiliar situations, and how to act on the player's units.

All the way through the Self-Play training, from zero to the top level in the world

The goal of OpenAI is the general artificial intelligence construction safety. For AI, Dota is an ideal test field, competition is intense, complex game. If you want to succeed in this game, we must advance the existing technology, technological breakthroughs. We developed a AI now, can beat the top occupation players in the Dota 1v1 game.

The rules of Dota are very complicated, if it is used to construct the rule method, then AI must be a bad player. So, we have full use of the self game (self-play) training method. At the beginning of the training, AI for the world knew nothing, but their copy game, which means that it and its rivals are always well matched. A little bit of lifting it with this method, until it reaches the world's best athlete occupation level. TheInternational (TI Dota) is the world championship, attracted 20000 viewers around the world to watch the athletes compete for 24 million bonus occupation. In TI, we sent AI and the number of master of the game, in the game, AI show the robustness and strong skills, many players want to play for occupation and our AI, and consider it as a means of training.

The next step for the research will be the 5v5 game, we hope to make 5 cooperation AI, beat the human team. In addition, we also look forward to try to get AI and human in the same team, cooperation win.

Collection reportPublish

Binding mobile phone number can comment! To bind

The latest comments

    See more comments failed to load, please click Retry loaded all commentsDotA

    58 thousand subscribers

    Related news

    Beijing Public Security No. 11010502033561 Internet drug information service qualification certificate Beijing Beijing ICP license No. [2015]0377-157 Article No. 140123

    Report Mailbox Company Name: Beijing a fabric technology Co. Ltd.