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

BGonline.org Forums

GnuBG 0-ply vs 2-ply

Posted By: Nack Ballard
Date: Wednesday, 20 April 2011, at 8:54 p.m.

In Response To: 65R-54S-33A-61U-C-21H-31U-54H-42S-64S-51 - Rollout (Stick)

Nack: If one or the other has to be sacrificed due to CPU considerations, I guess I'd rather see 5k of 0-ply than 1k of 2-ply. However, I haven't seen enough comparisons of 2-ply vs 0-ply rollouts to be clear on how much the lower ply hurts, whereas I am quite clear on how wild variance is on a mere 1k trials.

Stick: Ewww. 0 ply, ewww. The nice part about getting a 5k of 0 ply (even with GNU) is that you will spend virtually no time on the rollout. The bad part about it is have you ever seen 0 ply play? When someone posts a 0 ply rollout I don't even look at the results. Give me 1k at 2 ply every day of the week.

Nack: No, I haven't seen plays made with 0-ply. However, there may be a greater difference in quality or impact on later positions than on early game positions (the latter of which characterizes the arena in which I'm interested and the vast(?) majority of the rollouts Leobueno has posted).

As badly as a bot might play on a given setting, I think it's very easy to overestimate or imagine how much it affects the bottom line of a simulation.

I've compiled results for all the 0-ply GnuBG rollouts (I could find) that Leobueno has posted (including only plays for which a minimum of 5k trials were rolled) for which someone else did a comparison rollout on GnuBG 2-ply or XG 3-ply in the same thread. Each stronger-setting result is listed beneath Leobueno's 0-ply result.

Nacbracs are clickable if you want to view the details of the rollouts.

    64-d [R S4] /7 (0-ply)
    64-d [R S2] ^46

    64-g [P S16] /7 (0-ply)
    64-g [P S21] <15

    51S-63R-41s-66 –5–5 [H P11] /7 (0-ply)
    51S-63R-41s-66 –5–5 [H P13] <10

    53P-52D-64H-61U-52H-42 [W S31 Z42] /12 (0-ply)
    53P-52D-64H-61U-52H-42 [W S27 Z61] <10*1 (only 1k for Z)

    63R-53D-22M-33A-43T-43S-52H-51 [E R41] /7 (0-ply)
    63R-53D-22M-33A-43T-43S-52H-51 [E R58] <1 (2-ply, but only 1k)

The differences in margins of the top two plays of the first four positions are: .002, .005, .002 and .004, respectively. As it "happens," the differences of .019 and .017 exist only in the cases where only 1k trials were rolled at one of the settings.

Also (not listed above), in the fourth position the U play is -.087 in the 0-ply rollout and -.046 in the XG 3-ply rollout, which is a difference of .041(!). In this case, the play is shorthanded at both settings: the first representation of U is 1k and the second is 2k.

The reason I'm bringing all this up is that for people doing rollouts for themselves, with limited CPU resources, it seems to me that 0-ply is a very sensible option for early game positions (not sure about later-game positions, that would have to be tested).

OTOH, to meet the standard of this forum, where rollouts are being viewed and digested by a large number of people, I think that Snowie 3-ply, GnuBG 2-ply and XG 3-ply (or higher) should be considered the automatic settings. Even more importantly, though, for early game positions I'd like to see a minimum of 5k trials per play (with the possible exception of blunders — plays worse than –.06), and I apologize for sounding like a broken record.


Below is one more 0-ply rollout result, in case someone is interested in rolling it out on a higher setting.

43Z-61P-33C-51 [U S3 $8] /7 (0-ply)

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.