Danil 'Dendi' Ishutin attacks the positions of the OpenAI bot at the tournament The International, August 11, 2017. Screenshot from the live broadcast of the tournament
Yesterday in the championship of Dota 2 International in Seattle, the bot created by the organization OpenAI defeated one of the best players in the world Dota 2 Danila Ishutina, a well-known professional circles under the name Dendi ($ 735,449 Prize in the career). The 27-year-old Ukrainian surrendered in the middle of the second game.
According to the rules of a one-on-one match, a player who committed two murders or destroyed the enemy's tower was considered the winner in each game. In the first game, OpenAI dominated and won in less than 10 minutes, and Ishutin seemed to marvel at the bot's capabilities. At the beginning of the second game, the bot made a killing, and soon Dendi stopped the game, admitting defeat. "This thing scares," Dendi said to a huge crowd of spectators. Ilon Mask rejoiced.
Thus, the OpenAI bot was unbeaten in the confrontation with the best players in the world in Dota 2. Earlier he celebrated the victory over Arthur 'Arteezy' Babaev (No. 1 in the overall rating) and Sayed 'Suma1L' Hasan (No. 1 in the rating of 1v1).
Dendi against the bot OpenAI
Dota 2 is a complex game with hidden information, where players have to plan actions, attack, cheat and deceive the enemy. There is no explicit correlation between the player's abilities and the number of actions per minute, although the bot has the same number of actions per minute as people's. Nevertheless, the players note that the bot gained an advantage due to a faster ration and exceptionally accurate movements, compared to a live person who clicked the mouse.
A member of the maintenance staff of the championship The International raised his hand with a USB flash drive on which the bot was recorded. Professional Dendi stands in the background, waiting for the start of the bout. Photo: OpenAI via YouTube
"What we've shown here is called a common learning system," explains Greg Brockman, co-founder and technical director of OpenAI. – It still has a number of limitations, but it is already capable of defeating the best professionals in Dota. This is a step towards building more general systems that can be learned in more complex, confusing and important real world problems, such as the surgeon's profession. "
The OpenAI bot learned to play Dota 2 by conducting a large number of gaming sessions against itself. The training took two weeks. During this time, the path from random random actions in the game to skills sufficient to beat the best professionals was passed. The developers did not put into the program any strategies, did not use the help of experts. The bot just started from scratch and played with itself, step by step making small improvements in the game, until it reached the professional level.
However, in the current state, the bot is unlikely to compete in a big game, where teams of five players usually play. Still, one-on-one matches are a simplified version of Dota, but in team games there are many more different strategies and specific techniques. In an official blog, the OpenAI organization stated that creating a group of bots to play against a team of people is the next goal.
For OpenAI, this is a definite achievement. This non-profit organization was founded in December 2015 by well-known entrepreneurs Ilon Mask and Sam Altman, executive director of Y Combinator startup incubators. Among the sponsors – a number of influential figures Silicon Valley, including businessmen Peter Til and Jessica Livingston. The organization aims to create a safe (that is, public and open) Artificial Intelligence.
In December 2016, OpenAI introduced the Universe middleware for training and training strong AI. Theoretically, training can take place on all the information of humanity accessible via the Internet. These are games, websites and other applications.
OpenAI believes that reinforced learning is an important method of machine learning that will greatly improve AI. In the process of learning this method, the test system (agent) learns by interacting with some environment. Unlike traditional training with a teacher, the response to AI decisions is reinforcement signals, while some reinforcement rules are formed dynamically and are difficult to understand, that is, they are based on the simultaneous activity of formal neurons.
"Our ultimate goal is to The development of a single intelligent agent that is able to flexibly apply the experience accumulated in the Universe to meet new challenges and quickly gain new experience, which will be an important step towards a strong AI, "say axis while the statement OpenAI.
AI developments are now engaged in commercial corporations such as Google, Facebook and Microsoft. Of course, they put their financial benefits above the interests of mankind. The AIs created by them will act accordingly. Non-profit organization OpenAI with an open-source alternative to AI tries to resist corporations. All research within the OpenAI Institute is published in the public domain. In the official announcement about the foundation of the organization it is said: "In connection with the unpredictable history of AI, it is difficult to foresee when the AI of the human level can appear. When this happens, it will be important to have at the disposal of mankind a leading research institute that is able to prioritize the gain for all over its own interests. "