[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Bug-gnubg] Pruning nets and SSE

From: Øystein Schønning-Johansen
Subject: Re: [Bug-gnubg] Pruning nets and SSE
Date: Thu, 17 Feb 2011 11:18:51 +0100

On Wed, Feb 16, 2011 at 10:50 PM, Philippe Michel <address@hidden> wrote:
I have tried to use the patch to SSE code that Oystein posted here about one year ago to compute the pruning nets' outputs (using the current weights file padded with zeroes at the right places).

It works and is a little faster than the current way, so I suppose it could be used as is. The cost for non-SSE processors looks very small and they could always use the current nets.

But then the obvious question is : how about using properly trained nets with 8 and 12 (or 16) hidden nodes ? Could it make a difference in overall playing strength ? Could it allow faster evaluations thanks to more aggressive pruning ?

Do you (especially Joseph) have insights about this ? Were other parameters tried than 5 and 10 hidden nodes / 10 moves selected ? How did they do ?

I guess I should be able to retrain a pruning net to say 16 hidden nodes (I suggest 16 hidden as new AVX register systems are on the way, and having 512bit registers may not be far into the future!). I'm not sure how much a increase from 5 to 16 hidden inputs mean, but it could be interesting to at least run some experiments. I can at least do some checking to see how much work this will be. I guess a training session of such small nets will be just some days, and I have the tools to do it.


reply via email to

[Prev in Thread] Current Thread [Next in Thread]