Now, he's once again dealing with Starcraft. But this time he's not a player, he's trying to teach a robot how to play the game, which means that after the AI becomes the best Go player in the world, Interstellar becomes the next target for it to take on.

DeepMind has partnered with Blizzard, the developer of StarCraft, to allow AI researchers to learn from millions of previous matchups. One of their goals is to develop an AI system that can beat humans, like AlphaGo, which has defeated Go world champions Lee Se-dol and Ko Jie. And the ultimate goal is to apply this technology to the real world, not just at the gaming level.

We're trying to understand how the human brain works," said Jacob Repp, Blizzard's lead software engineer. If we could get this kind of high-quality data stream - the raw input from humans playing the game and its results - it would be very useful data for people in behavioral research. "

StarCraft 2 presents an interesting challenge for AI researchers. Unlike chess or Go, the Star player does not have perfect information. This "fog of war" means that players (real or virtual) must make plans, make decisions, or respond to actions that will only have consequences in a few minutes. As DeepMind's researchers put it, the results "will face substantial challenges in time credit allocation and exploration."

Blizzard already uses neural networks to assess player skill based on information about their keyboard and mouse inputs, their approach to lining up, and how efficiently they play the game, signals that can make the game more interesting or make the matchups more balanced.

However, for an AI to play StarCraft 2, they must be able to "see" the game's 3D maps and interpret them quickly and accurately.

DeepMind's first test involved the training of neural networks and artificial intelligence before putting their application into the game. Even without further instructions, the AI can walk around the map, move the camera and even line up troops at will.

Before joining DeepMind, Vinyals worked on image search and Gmail's "Smart Reply" feature. The team also worked on voice recognition, allowing the AI to remember how different people speak, so that recognition can be accomplished when that voice is encountered again.

"In Starcraft 2, this is something that needs to be addressed as well." Vinyals said. For artificial intelligence, remembering what they encounter while understanding the behavioral meaning requires the use of LSTM neural networks. "Computers can retain the memory of a particular piece of data for decades, but now that memory needs to be not only preserved, but also completed to retrieve information when needed in the future."

Recommended>>
1、IOTA Forum for Automotive Supplier Cooperation
2、Big upgrade for Jovi Vivos smart voice assistant all in one AI feature
3、Python Data Visualization Magic Matplotlib Plotting Explained I
4、MachineLearning
5、Ma Huateng starts bragging invest in 2 big business opportunities in 2018 and earn 1 million a month in a year

已发送

发送中