Get r2 back up to pre-auto-skew levels

jkomoros / sudoku

A sudoku puzzle solver, generator, and difficulty-rater built in Go

Apache License 2.0

5 stars 1 forks source link

Get r2 back up to pre-auto-skew levels #264

Open jkomoros opened 8 years ago

jkomoros commented 8 years ago

We went from an r2 of roughly .78 to .64 with the switch to the new auto-skew relative difficulty calculation (the resolution of #253). Although the resulting model is "better" in that it outputs numbers more smoothly across the range (which was a serious problem before), it's significantly less accurate.

This bug will track efforts to get r2 back up to as least as good.

jkomoros commented 8 years ago

Early tensorflow experimentation is apparently going somewhat well (not yet in comparable terms).

Another thing to look into is a higher-level TF library that helps do cross-validation etc

jkomoros commented 8 years ago

Playing around with SGD in Orange implies that it might just not work in this context, which would explain why my numbers I'm getting in TensorFlow are so crazy bad...