Re: [Bug-gnubg] Re: Bug? Pruning On/Pruning Off gives significantly different results at 4ply

[Michael Petch]

The sigmoid calculation being more accurate (may the word I should use is “more preciise”) was a side effect of the new SSE2 performance enhancement. It has been discussed previously on the mailing list. The variations I have seen have been very small (maybe at most 0.1-0.2% MWC change when observing TOTAL errors of a match as expressed in MWC).

Oystein and Christian, any feelings on this? Is it significant enough that the net would have to be retrained?

Michael

On 31/08/09 3:07 PM, "Joseph Heled" <address@hidden> wrote:

> I am not sure if you are missing the point or me. Being "more accurate" may be nice but unless the training is repeated it is probably harmfull.