I guess you're saying that it's still a legitimate subject, and that progress is made by building on previous results. Nobody is disputing that.
But the TFA doesn't mention previous results or even the existence of the field, I have the impression that Norvig is not aware of it. So this is not contributing to the body of knowledge re automata induction, this is recreative CS. Nothing wrong with that per se, but there's also nothing wrong with providing some scientific background.
Btw in the blog post, the approximative approach is motivated by NP-hardness of the exact problem. Given the link (parameter-preserving reduction) with graph coloring already mentioned, the problem can't be approximated in polynomial time with arbitrary error, unless P=NP. This theoretical result is backed up by practical experiments with approximating coloring algorithms, which often find solutions 100% off from the optimal nr of colors. So good luck with that.
Also, informed regular inference is already NP-hard when restricted to DFAs with just 2 states:
Perhaps surprisingly, the problem becomes tractable when the data is `complete', in the sense that it is consistent with just one single automaton (google RPNI).