After defeating world Go champion Ke Jie in late May of this year, the Google DeepMind team hasn't rested on their laurels. They recently proposed allowing AlphaGo to continuously improve through self-reflection and learning. With the recently announced new version, AlphaGo Zero, the DeepMind team went back to basics, emphasizing that in just three days of self-play, the system nearly covered the entire millennia of Go history accumulated by humans, even exploring new strategies along the way. Compared to the AlphaGo version that defeated South Korean Go champion Lee Sedol last year, the system achieved a perfect 5-3 victory.
According to DeepMind CEO Demis Hassabis and AlphaGo project leader Dave SilverExplainThe new version of AlphaGo Zero makes its decisions entirely based on the positions of black and white pieces on the board. This is different from previous approaches that still refer to human playing habits. At the same time, it merges the previous strategy network and value network prediction methods, and judges all chess paths through a single type of neural network calculation method. It no longer only determines the position of the pieces in the fastest way. The new system may even have a longer "thinking" time for different chess paths.
In addition to the previous DeepMind teamSelf-examinationThis continuous self-training mode, combined with its ability to remain dormant, may have also enabled AlphaGo Zero to develop sharper judgments about its moves. Through three days of continuous training, it rapidly learned from scratch the thousands of years of human experience in Go, achieving a 3-100 victory over the previous AlphaGo version that had played against South Korean Go King Lee Sedol. Furthermore, over the course of 0 days, it surpassed the Master version of AlphaGo that had previously defeated World Go King Ke Jie.
During the game, AlphaGo Zero not only learned all the moves humans had made in the past, but also deduced new moves that broke the mold. This means that artificial intelligence will be able to help humans rediscover valuable data from existing data that had not been noticed before, thereby helping to promote new developments.
However, this may cause more people to worry again about whether computers will dominate everything in the future. However, many people believe that the development of artificial intelligence technology must inevitably require many ways to prevent system conflicts and allow humans to intervene and terminate calculations when necessary to prevent the development of applications that were originally expected to bring convenience from turning into a crisis.
