AlphaZero Explained

If you follow the AI world, you've probably heard about AlphaGo. The ancient Chinese game of Go was once thought impossible for machines to play. It has more board positions () than there are atoms in the universe. The top grandmasters regularly trounced the best computer Go programs with absurd (10 or 15 stone!) handicaps, justifying their decisions in terms of abstract strategic concepts – joseki, fuseki, sente, tenuki, balance – that they believed computers would never be able to learn. And they spent three years painstaking years trying to prove this belief; collecting Go data from expert databases, tuning deep neural network architectures, and developing hybrid strategies honed against people as well as machines. Eventually, their efforts culminated in a dizzyingly complex, strategic program they called AlphaGo, trained using millions of hours of CPU and TPU time, able to compete with the best of the best Go players.