On 24 June 2012 13:38, Joseph Heled <address@hidden>
It is heart warming to see that the race net can be improved still :)
The benchmark rates the new crashed net as stronger, which is great too, but the new contact net is rated weaker in moves (0.0104453134853 vs 0.0104996763558), stronger in cube actions.
I take this back. After updating the contact benchmark to include all missing positions, the new net comes out ahead. The crashed benchmark need to be updated as well.
I am very interested to know how those nets were generated?
The attached files are drop-in alternatives to gnubg and gnubg-nn weights files.
The crashed net is substantially stronger, the racing net should be slightly better and the contact network should benefit from that for positions soon to be crashed or non-contact but remain relatively unchanged elsewhere. There could be gross regressions in certain cases of course.
I'd be interested if some of you have private benchmarks or some process that could quantify the difference in strength (in ppg or Elo) between these and the 0.90.0 nets for "normal" usages like 2ply cubeful play.
In the gnubg-nn benchmarks, the improvement for 1ply checker play is important compared to that of 0- and 2ply. It would be interesting to check if this translates to a clearer strength hierarchy between 2ply and 3ply than with the current nets.