>> where indeed they use machine learning to induce the rules of chess from a large number of played games (and then learn to play better than any human).
They didn't induce the rules of chess! What they did was learn an evaluation function for chess moves, from scratch, i.e. without giving it the rules or any hand-crafted features.
So they learned a classifier for good/bad board states, which was later used in an alpha-beta minimax search. However, the classifier didn't learn the rules of the game, only a mapping from board positions to labels, good/bad. This is not made explicit in the text, but minimax typically has a rule-base to generate moves as successor states for the search; so the rules of the game were hand-crafted, but did not contribute to the learning of the evaluation function.
It's extremely hard to induce a set of rules with neural nets, or indeed any statistical machine learning algorithm. I mean, imagine representing the rules of chess as a function; let alone trying to learn that function from data.
(Full disclosure: I study algorithms that learn rules from data; they're not neural nets).
They didn't induce the rules of chess! What they did was learn an evaluation function for chess moves, from scratch, i.e. without giving it the rules or any hand-crafted features.
So they learned a classifier for good/bad board states, which was later used in an alpha-beta minimax search. However, the classifier didn't learn the rules of the game, only a mapping from board positions to labels, good/bad. This is not made explicit in the text, but minimax typically has a rule-base to generate moves as successor states for the search; so the rules of the game were hand-crafted, but did not contribute to the learning of the evaluation function.
It's extremely hard to induce a set of rules with neural nets, or indeed any statistical machine learning algorithm. I mean, imagine representing the rules of chess as a function; let alone trying to learn that function from data.
(Full disclosure: I study algorithms that learn rules from data; they're not neural nets).