[Pachi] Better semeai handling ?
pasky at ucw.cz
Wed Apr 27 11:32:50 CEST 2016
Remi trains the patterns as teams of features. We just do frequentist
statistics on patterns as a whole.
There should be some pretty rusty code that can feed patterns to
Remi's MM tools and train the teams of features, which was part of the
effort for the probdist playout policy. I didn't bother trying this
approach for the pattern priors. Based on what you noticed (I never
measured prediction rate - if you have some scripts/code for that, it
could be nice to have merged!), it could be a pretty nice boost!
On Tue, Apr 26, 2016 at 10:14:28PM +0200, lemonsqueeze wrote:
> Wait, something's wrong.
> Shouldn't large patterns get around 34% prediction rate as in Remi's paper ?
> On 04/26/2016 15:13 PM, lemonsqueeze wrote:
> > Thanks, this makes more sense now.
> > I guess i should apologize for my previous comment on moggy:
> > I just ran some tests on kgs 6d games to get an idea of moggy's
> > prediction rate and it does very well actually: 24%
> > (i made the engine return the most frequent move from a number of
> > moggy runs)
> > It even beats the large pattern predictions which i thought would be
> > better (around 21%)
> > Does it sound possible that alphago's fast policy could have same
> > prediction power and yet be significantly stronger wrt semeais ?
> > Seems even more puzzling than before !
If you have good ideas, good data and fast computers,
you can do almost anything. -- Geoffrey Hinton
More information about the Pachi