快猫短视频

An artificial intelligence has learned to play the Atari computer game Skiing 6000 times more quickly just by reading the instruction manual first. The same approach could help teach AIs to drive cars or robots to use household appliances.

Atari computer games from the 1980s have become a benchmark problem for AI research because they have simple controls, but greater complexity when it comes to completing levels.

The core approach is to give an AI a target, such as maximising a score or completing a level. To train the AI, you give it individual, static snapshots 鈥� or frames 鈥� of the screen from games and it has to push a virtual button to make a move in response. Gradually, after trying millions or billions of approaches, it builds up a statistical model of what order moves should be made in.

顿别别辫惭颈苍诲鈥檚 took about 80 billion frames to learn Skiing, in which the aim is to steer left or right to avoid trees and go between pairs of poles as you descend.

Now, at Carnegie Mellon University in Pennsylvania and his colleagues have given an AI a head start by getting it to read the instruction manual and extract lessons that would have taken billions of attempts at playing the game to learn. They enabled their game-playing AI to read the instructions by using another AI 鈥� a 鈥� to help it recognise what objects in the game the words 鈥渢ree鈥� and 鈥減ole鈥� referred to.

This resulted in it requiring about 6000 times fewer training iterations to master Skiing than Agent 57, needing some 13 million rather than 80 billion frames. So, because the game runs at 30 frames per second, this is equivalent to the new AI playing for five days, when Agent 57 needed the equivalent to 85 years of game playing, says Wu.

Skiing is a simple game that people can get the hang of in about 5 minutes, so it is absurd that an AI would need many years to master it, says Wu. He says his approach leverages the growing ability of AI to extract information from text, without needing the vast computing resources of a cutting-edge AI company.

While DeepMind can afford to throw 1000 high-end computers at training the AI, this is beyond the reach of academics. 鈥淭he things that DeepMind is doing are also very clever, but we have to attempt more efficient ways,鈥� says Wu.

Reference: