So I’ve been reading about how in chess there are an extremely high combination of moves that can be made and trying to write out every combination is impossible. In an eli5 style, how are algos like alphago able to overcome this fundamental issue? How do they go about solving this problem?
In: 0
By exploring much less of the tree, because they focus on the better ideas with beefier evaluation. Neural Nets such as AlphaZero or Leela Chess Zero explore far fewer nodes/second, a few thousand nodes/sec on my *machine*(full CPU and GPU) versus about a million nodes/core/sec for SF. Basically, they spend more CPU cycles determining how good positions are, and explore far fewer positions. This is also, IMO, the correct human approach to take. We can’t explore the breadth of chess, but we don’t need to, we can just focus on the correct ideas.
Latest Answers