chess.com personality bots research

[dkappe]

Check out my old distilled nets (as small as 16x2) and small nets trained on bad gyal data. You can really turn the knob for nps or total nodes much more than on a big net.

https://github.com/dkappe/leela-chess-w ... d-Networks
https://github.com/dkappe/leela-chess-w ... -style-net

Thanks. They still seem to strong, though? At 2408 rating it's still going to beat most people.

The idea here is that to get this strong, in the process of the training, the engine was at various strengths (so at some point it was 1000 rated, 1200 rated, 1400 rated...) we'd take a "snapshot" of them (stop their training) and then we'd have many bots at different strengths that people interested could play with (for free, because chess.com's bots are commercial.) I don't know if decreasing to, say, 1 node would be enough to play at some 1000 rated human player level.

You can limit the nps or just the total number of nodes per move to get down to 1000. Like I said, much more of a knob to turn than for the bigger nets.