Re: [gnugo-devel] hash collisions

[Arend Bayer]

> > In a web search I found that some chess programmers use fairly short
> > hashes, and accept a certain level of collisions. Even though this does
> > cause misevaluations due to false cache hits, it's a fairly small fraction
> > of nodes which will effect the evaluation of the root node, and can be a
> > win overall.
> 
> Sounds reasonable, at least if there are effective methods to
> determine whether a bad move was caused by hash collisions. (It would
> be a nightmare to debug otherwise.)

I am not against it, but with the current code  this is out of the
question, since a hash collision is too likely to trigger an assertion
failuer somewhere.

> I suppose the main reason for this
> is to allow bigger hashtables in the same amount of memory, which
> maybe is more useful in chess than in go. At least GNU Go doesn't seem
> to benefit much from a huge cache.

Another reason is probably that smaller hashes are faster to compute and
to compare.

Arend