[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums

Reference packets

Posted By: Nack Ballard
Date: Saturday, 16 January 2010, at 2:45 p.m.

In Response To: Rollouts (Stick)

These reference packets are marvelous. Thanks so much to Rich and Stick for creating them. It's a lot of work for the creator(s) but a treasure trove of information to the eternal backgammon student like me.

I agree with Stick that it is more useful to have all the information for money, unless what you really want is veracity at DMP, GS, GG, or something nearly as extreme.

If you are interested in knowing how much the match score affects the play (or cube) for a given position, I think it makes sense to do one eval/rollout of the position, an eval/rollout of the same position for money, and then consider the comparison positions for money only. If there is a difference in margin for score and money, that adjustment in most cases can be fairly assumed up and down the line. (You can spot-check the extreme positions when you think it might not be the case.)

To be clear, it's fine with me to skip the match score rollout altogether; my point is I only want to see the score applied to one of the packet's positions at most, and the result of that position would also be reported for money in order to obtain a benchmark adjustment.

That said, I do NOT think it is necessary to roll out all the positions. IMO, it's not even desirable unless at least 5k trials are done for each (and even that assumes a vacuum -- that the CPU comes without a cost in getting other rollouts done).

What I recommend for a reference packet is to roll out the main position 20k+ (or skate by with 5k if you must), compare to the strongest eval you have, then just do evals for all the positions. Whatever the difference in eval and rollout, just make that adjustment up and down the line. If you believe a blanket overlay assumption might be unjustifiable for a particular type of position, do another rollout at one or both extremes (2k might be enough for that purpose) to verify that the eval bias is reasonably consistent for the range (and, in the somewhat rare case it's not then extend those two rollouts for use as two additional benchmarks and interpolate margins for all positions within the range).

In my experience, that methodology gives you the biggest bang for the buck.

Nack

Messages In This Thread

 

Post Response
Your Name:
Your E-Mail Address:
Subject:
Message:

If necessary, enter your password below:

Password:
 

 

[ View Thread ] [ Post Response ] [ Return to Index ] [ Read Prev Msg ] [ Read Next Msg ]

BGonline.org Forums is maintained by Stick with WebBBS 5.12.