This article is originally published by AI Frontier. The original link is t.cn/RTnmxxF
The author | EmTech CHINA
Edit | Emily
StarCraft, top esports players, artificial intelligence and Google are the key words of Oriol Vinyals’ life.
Who is Oriol Vinyals? Maybe it felt strange to some people, but a group of people have long heard his name, and when she was younger, he was a Spanish starcraft games e-sports, ranked first, you use Google Gmail, image recognition services also have from his written procedures, he also got a bot, you can chat with you what is the meaning of life, Last year, he won the MIT TR35 Award and was named one of the 35 most innovative young people in the world.
Now, as a research scientist at Google’s ARTIFICIAL intelligence company DeepMind, he’s turning game worlds into a monastery for AI robots, ready to take on human masters once again. This is one of the key moments in the development of artificial intelligence after AlphaGo.
Figure | Oriol Vinyals
After AlphaGo’s defeat of the human champion had built an impenetrable wall in the go world, DeepMind had moved into another arena: gaming. Starcraft, which is popular and respected in the video game industry, was chosen as the AI training environment.
Starcraft is a game launched by Blizzard Entertainment in 1998. It sold 1.5 million copies in 1998, and became the pioneer of esports. In 10 years, it sold more than 9.5 million copies in total. Broke sales records for strategy games. Set in the 26th century, three races, including Terran, Zerg and Protoss, vie for supremacy at the center of the Galaxy to open the game.
BlizzCon 2016 at the end of last year, Blizzard Entertainment announced a partnership with DeepMind to train AI in the StarCraft II environment to aid AI research and development. The news has thrilled gamers and AI researchers around the world.
Immediately after the news was announced, Oriol Vinyals wrote on Twitter: “StarCraft is back in my life :)”.
Into artificial intelligence experts from the e-sports person that’s right, like most boys in childhood are deeply attracted by computer games time, hoping that the birth Oriol Vinyals, too, started to play starcraft he was 15, also with super performance, as was once ranked first in Spain, the famous e-sports 16 players across Europe, But instead of becoming a professional esports player like most of the best, he chose to devote himself to computer research.
“I’m fascinated by the ai issues in games.” He said. From then on, he started the process of becoming an AI expert from an e-sports master.
After completing a double major in telecommunications engineering and mathematics at the University of Catalonia in Spain, he went to the US to complete his undergrad thesis on machine learning and computer vision at Carnegie Mellon’s Robotics School. Then I went to the University of California, San Diego, where I earned a master’s degree in Computer Science and Engineering. In 2009, he entered UC Berkeley for a PhD in Electrical Engineering & Computer Science, where he also participated in Berkeley’s Overmind program. Photo: Oriol Vinyals tweeted, “StarCraft is back in my life.”
Berkeley Overmind is well known in the field of game AI. Overmind is actually named after the Overmind character of the Zerg in the StarCraft game. The project delves into the use of various AI computing techniques to figure out what steps and strategies to take to win matches — in short, to develop an AI robot that knows how to play real time strategy games. Overmind made a name for itself by winning the Starcraft AI Competition in the full game mode category at the 2010 AI and Interactive Digital Entertainment (AIIDE) conference for the first time.
After he asked Google’s Album app to “read” photos, Oriol Vinyals joined the Google AI team to build new technology for the translation system. “I just wanted to see if a computer could accurately describe an image, from pixels to interpretation, which is also a form of translation.” Oriol Vinyals said. With this in mind, he changed a line of code, replacing the French translation program with his own image program.
The next day, he showed the computer an image of a busy market with lots of bananas on the ground next to the stall. The surprising result was: “A group of people buying fruit at the market.” “Success! It doesn’t just say people are standing on the street.” Oriol Vinyals vividly remembers the excitement of the moment. This means that machines can read complex images, a technology now used in Google’s Image search service. When a user enters a keyword, the system begins to “read” the image and present the results.
In addition, Gmail, which almost everyone uses, also has his work in mind. For one thing, the “Smart Reply” feature, which automatically responds to an email with three brief suggestions, will be answered by clicking on it when you are busy.
In 2015, a team of Google researchers published a paper showing that chatbots they had trained through movie dialogue were able to talk abstract life issues to humans. The researchers began by asking simple questions, such as my VPN connection was not good ENOUGH for IT, and the machine did its job as if IT were an IT professional. But then they asked IT metaphysical questions like “What is the meaning of life?” and “What is the purpose of being alive?” and the machine said: “In pursuit of the Greatest good”, “For Eternal Life”…… And so on. Conversations go from feeling funny to a little scary. The chatbot was developed by Oriol Vinyals and his colleague Quoc V. Le. Photo: A chatbot talks with a human about the meaning of life. Image credit: Wired
Oriol Vinyals joined DeepMind in January last year, after Google Brain developed a number of services with real-world applications, to challenge artificial intelligence — which is harder than Go — to teach computers to play video games without typing in rigid coding rules. Instead, the computer learns by experience, and the environment used to train the machine is what Oriol Vinyals does best in Starcraft.
DeepMind and Blizzard entertainment released starCraft II’s training AI tools in August, including a machine learning API, footage of up to 65,000 StarCraft II esports matches, and open source tools. Let the AI learn the tactics of the human player, and they also break down certain game elements into mini-games, such as map exploration, resource gathering, etc., to help the AI train specific skills.
Apart from the popularity of the game itself, the most important reason for choosing StarCraft II is the complexity of the content, which has three characteristics: incomplete information, resource management, and immediacy.
In go, players can only look at one piece of the board, so the AI robot can think within a few hundred moves. However, too much higher than that of difficulty of the game, especially real-time strategy game, players must start long time scale of a comprehensive strategy, and the players can see map within your own unit, unable to see the whole game scene at the same time, also is the incomplete information environment, must rely on memory to remember the opponent’s message, but the game at the same time, competitors also in play, Messages are fluid rather than permanent, and test the player’s ability to adjust plans based on new information.
In addition, the game involves economic planning such as resource collection and troop production, testing players’ long-term strategies and skills, so it is regarded as the best field for AI training. Once a breakthrough is made, AI development will reach a new milestone.
DeepMind points out that the dimensional action space for StarCraft is very different from previous studies on reinforcement learning. For example, when players do simple tasks, Expanding a base to a location, for example, requires coordination of mouse clicks, cameras, and available resources — a hierarchical process of action and planning that can be challenging for reinforcement learning.
However, just as before AlphaGo beat the human king, many people did not think that the machine would win, while others thought that the machine had a good chance, there is now a divided opinion in the gaming world: “I don’t think AI can beat professional players, at least not in my lifetime.” Byun Hyun Woo, a Korean esports expert who won the 2016 StarCraft ii WCS World Championship, previously told MIT Technology Review in an interview. He further explained that starCraft is a real-time strategy game that has to react quickly to a lot of uncertainty and variability, but AI still seems to be bad at dealing with the unexpected. Another reason is long-term strategy. “It’s so hard to build an AI that’s good at making both short-term and long-term decisions.”
On the other side, machines might have a chance. LEE Jae-dong, once the world’s top Zerg player and a subject of the documentary State of Play, thinks AI might have a chance against humans because computers don’t get tired and can be much faster than humans. In StarCraft, the speed of Actions Per Minute (APM) is key. The advantage of a machine is that a top player can do hundreds of commands a minute, whereas an AI can do tens of thousands.
But putting aside the question of whether machines or humans will win and getting back to AI research, using games to train AI “requires innovative strategies, memory and the ability to deal with uncertainty, which is definitely a big leap forward”, says Oriol Vinyals.
Memorization, for example, can be applied to a variety of environments, such as a computer in a data center. When the machine sees that Sunday consumption is always high, it remembers that it will be the same the following Sunday, helping to implement power management or power-saving designs in a data center, he explained.
Oriol Vinyals once said, “There is still a long way to go to beat human professionals.” That may not apply, though.
In August, OpenAI, the artificial intelligence nonprofit founded by Tesla CEO Elon Musk, shocked the esports world by beating Dota 2 pro Dendi one-on-one in an esports performance match.
In the past, Berkeley Overmind researchers trained AI robots to play against Oriol Vinyals. Even though he was already a world-class player, he was beaten by Overmind AI. “Overmind Triumphs over Professional Players” was the headline. Now Oriol Vinyals is the lead scientist in training DeepMind’s GAME AI, and it looks like starcraft’s esports masters will soon be given a shock education.
Chinese tech and gaming enthusiasts will soon have a chance to meet Oriol Vinyals, an artificial intelligence expert who is trying to overturn the confidence of gamers.
He will be a guest speaker at EmTech China, which will be held from January 28 to 30, 2018. This will be the first time that EmTech China, which has been hosted by MIT Technology Review for over 18 years, will be held in China.
Coming to China, EmTech Will not only continue the style and tradition of MIT Technology Review, but also combine it with the development of Chinese technology and market, and become a bridge for exchanges between China and the United States in the field of emerging technology.
AI Frontier, as the strategic partner of EmTech Global Emerging technology Summit, will jointly cover the conference and provide live broadcast service on the official website for AI related topics.
Follow our wechat account “AI Front “and reply to “AI” in the background to obtain the SERIES of “AI Front “PDF e-books