[Top][All Lists]

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

Re: [Bug-gnubg] The race training and benchmark datasets

From: Philippe Michel
Subject: Re: [Bug-gnubg] The race training and benchmark datasets
Date: Sun, 16 Jun 2019 22:33:05 +0200
User-agent: Mutt/1.12.0 (2019-05-25)

On Mon, Jun 10, 2019 at 02:49:26PM +0200, Øystein Schønning-Johansen wrote:

> I will try re-rolling out these positions. Do you have any experience of
> how to do good rollouts of race positions? Good rollout settings for race
> positions?

When I re-rolled out the benchmarks I mostly used the settings that had 
been used previously. I think I changed the rolloutLimit parameter 
somehow (the number of alternatives included for checker plays). It is 
currently set to 10 for doublets and 5 for other rolls. I don't remember 
exactly what I did ; maybe it used to be 5 in all cases.

I previously wrote it would be useful to have variance reduction in 
sagnubg, but this is not very important since it does 0-plys rollouts 
(VR works for them, but it is slow and simply doing more trials is about 
as good in terms of SD vs. time used). Doing 7776 trials instead of 1296 
doesn't seem unrealistic.

>From the other parameters :

s version 1.93 weights 1.00 moves2plyLimit 20 rolloutLimit 5 nRollOutGames 1296 
cubeAway 7 include0Ply 1 evalPlies 2 shortCuts 1 osrGames 1296 osrInRoll 1

experimenting with osrInRoll set to 0 may be interesting. I dont know if OSR 
is used for speed or for accuracy...

Another interesting thing to try, if it is practical, would be to use the 
software you mentionned a few weeks ago to calculate exact values.

For instance, sort the positions by leading player's pipcount. Start 
from the smallest ones with your software ; that should tackle the 
hypergammon-like positions with few checkers and ideally the very 
unbalanced ones where the trailer can only try to save the gammon. The 
latter may well be misplayed in the current rollouts.

reply via email to

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