Looks like the Maia Chess neural net, released last year, is exactly what I'm looking for. It matches human play at lower skill levels better than other engines do.
"Maia is a human-like neural network chess engine trained on millions of human games."
"A collection of chess engines that play like humans, from ELO 1100 to 1900."
Main website:
Code repo:
https://github.com/CSSLab/maia-chess
Included in the latest release of Lucas Chess:
https://lucaschess.pythonanywhere.com/downloads
The Bad Gyal neural networks are also trained on human play, but they're not targeted at matching human play at specific ratings like Maia is. Rather, it looks like various sizes of neural net are all trained on the same human play:
https://github.com/dkappe/leela-chess-weights/wiki/Bad-Gyal
In theory it should be possible to train a neural net on the millions of games in chess.com's database to make it realistically play like a human with a given rating.
I've been looking for something like this but so far haven't found anything. Anybody know if it has been done?