Starting from nothing, a computer taught itself to read the instruction manual for Civilization and saw its rate of victory jump from 46 percent to 79 percent.
That is according to researchers at University College London, who developed "meaning-inferring algorithms" that, when applied to the computer, took it from zero understanding of its task to winning strategic play.
The process of learning is extremely technical to describe but what happened, from the sound of it, is the machine was given a list of actions it could take, and could understand information displayed on the game screen, and then was told to carry on.
The computer began with completely random behavior. In the trial-and-error process, different words would appear on the screen as it took actions, and then the computer could search for instances of those words in the instruction set, and for associated words in the surrounding text, and form hypotheses based on that.
In one test, a software installation, the computer was able to reproduce 80 percent of the steps that a human would perform if they read the same instructions. In Civilization, it won 79 percent of its games, compared to a program that won only 46 percent without relying on instructions.
In the case of software installation, the system was able to reproduce 80 percent of the steps that a human reading the same instructions would execute. In the case of the computer game, it won 79 percent of the games it played, while a version that didn't rely on the written instructions won only 46 percent.
Commence Skynet jokes now.
Machine-Learning System Learns Language by Playing Games [Kurzweil AI, thanks Steve]