When the AlphaGo program became the world champion at the complex game of Go, it did so by evaluating positions and selecting moves using neural networks that been trained by human experts and reinforced with self play. It wasn’t much of a step from there to move onto a new version of AlphaGo becoming its own teacher (Nature, 19th October 2017). This new program, AlphaGoZero, obtained superhuman performance by teaching itself with no human guidance and went on to trounce AlphaGo 100 games to zero.
Comments
Post a Comment