Game 3, move 10: Forum 3-1

[Zorba]

jDN4wEIzzmCCIQ:cImlACAAAAAA

[ah_clem]

Spoiler

We don't want to abandon our anchor or the midpoint.  And we want to keep a spare on the  6pt.  That narrows the choices somewhat to:
 

The aggressive 8/4*,
the passive 24/20,
and the moderate, thematic 15/11.

Any time you can put two men on the roof, you should take a good look at it, but I don't see the risk being worth the reward.  We're ahead in the pipcount and have fewer men back, so hyper-aggressive play is not called for.

But 24/20 is to passive, and the man on the 24 might be useful in holding rkb if things go poorly for us.

15/11 gets a checker out of harm's way (rkb will have to abandon the anchor to hit it) and brings in a builder.  With men on the bar we should bring in buliders to make home board points.  Seems like the right balance here.

15/11

[close]

[blitzxz]

Spoiler

This is getting lot harder then I thought. No hit is thematic but things have changed. If we make the loose hit and followed with dance we can cash and win immediatly. That's a big opportunity. If rkb doesn't hit back we will get nice double with quite many gammons and we really need those gammons in this score. So especially in this score and cube position I would do the loose hit. If the score were even I would be a lot more unsure about it and propably wouldn't have guts to do it. But here 8/4*. Let's go for it.

[close]

[stog]

Spoiler

tempo 8/4*

[close]

[diane]

Spoiler

I am liking the aggressive too, I know most will find that shocking 

[close]

[lewscannon]

Spoiler

Hit. If we don't hit, and he comes back in on the 4, that's going to make for a very effective backgame and he's going to be harder to evict than cockroaches in a South Bronx apartment. We have 3 points covered already, so more dancing or bad situations for rkb is a definite possibility, in which case we cube. If he comes back in on a 4, it's not a major problem as we are already pointed on the 20. This is why I was initially in favor of cubing before the roll, as this is the kind of play that can turn into 4 points quickly. (Of course, this can blow up in your face from time to time as well)

[close]

[socksey]

Spoiler

Hit!  For all the reasons above!

[close]

socksey



"I pray for: Wisdom to understand a man; Love, to forgive him; and Patience, for his moods. Because, Lord, if I pray for Strength, I'll just beat him to death." â€" Anonymous

[Zorba]

Spoiler

Well done Forum! 8/4* is clearly best, probably even at a normal score, and much more so at this matchscore as it is clearly more gammonish. I wimped out with 15/11, I guess I had not properly reckoned with the fact that hitting puts the second checker on the bar, which is pretty strong given our three point board. It reduces the risk normally involved with breaking your eightpoint (the only outfield point we have) and hitting loose. Another factor is that the alternatives moves just aren't very good.

[close]

-------------------------------------------------------------------------------------------------------------------> next!

[ah_clem]

Well done Forum! 8/4* is clearly best, probably even at a normal score, and much more so at this matchscore as it is clearly more gammonish. I wimped out with 15/11, I guess I had not properly reckoned with the fact that hitting puts the second checker on the bar, which is pretty strong given our three point board. It reduces the risk normally involved with breaking your eightpoint (the only outfield point we have) and hitting loose. Another factor is that the alternatives moves just aren't very good.

Agreed.  I haven't done a rollout yet, but hitting is clear here (despite my arguments to the contrary above)

The key thing is that since we hold the three and one points, there are not as many hitting numbers for rkb as a normal loose hit.  To hit our checker on the four requires a direct 4 or 2-2.  That's 12 out of 36 or a 33% chance.  Meanwhile, the chances of leaving one checker on the bar is high.  So, hitting loose is not as risky as it seems at first glance.  I didn't properly assess the risk when I looked at it yesterday, and overestimated the riskiness - hitting loose is really not very risky, especially with us holding an advanced anchor.

Rollout forthcoming...

[ah_clem]

gnubg results:

Spoiler

Big surprise here.

Evaluation says the hitting play is clearly superior:
    1. Cubeful 2-ply    8/4*                         MWC:  53.16%
       0.573 0.235 0.007 - 0.427 0.101 0.004
        2-ply cubeful prune [world class]
    2. Cubeful 2-ply    24/21 15/14                  MWC:  52.70% ( -0.46%)
       0.575 0.195 0.005 - 0.425 0.089 0.002
        2-ply cubeful prune [world class]
    3. Cubeful 2-ply    15/11                        MWC:  52.63% ( -0.53%)
       0.568 0.202 0.006 - 0.432 0.091 0.002
        2-ply cubeful prune [world class]
    4. Cubeful 2-ply    24/20                        MWC:  52.53% ( -0.63%)
       0.572 0.187 0.005 - 0.428 0.090 0.002
        2-ply cubeful prune [world class]

Rollout is not so clear.  The prosaic  24/21 15/14 is a statistical tie* with the hitting play, despite garnering no votes from the forum.  Why is it a good move?  It duplicates 2's.  2 to enter, and 2 to hit the remaining blots - obvious in retrospect, but I didn't see it over the board.  It also gets the runners moving, bringing them into contact with rkb's backmen. 

    1. Rollout          24/21 15/14                  MWC:  51.77%
       0.545 0.179 0.005 - 0.455 0.108 0.003 CL  51.27% CF  51.77%
      [0.001 0.001 0.000 - 0.001 0.000 0.000 CL   0.02% CF   0.03%]
        Truncated cubeful rollout (depth 11) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861634383 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    2. Rollout          8/4*                         MWC:  51.73% ( -0.04%)
       0.531 0.220 0.007 - 0.469 0.123 0.004 CL  51.39% CF  51.73%
      [0.001 0.001 0.000 - 0.001 0.001 0.000 CL   0.02% CF   0.04%]
        Truncated cubeful rollout (depth 11) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861634383 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    3. Rollout          24/20                        MWC:  51.60% ( -0.17%)
       0.541 0.175 0.005 - 0.459 0.109 0.002 CL  51.18% CF  51.60%
      [0.001 0.001 0.000 - 0.001 0.000 0.000 CL   0.02% CF   0.03%]
        Truncated cubeful rollout (depth 11) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861634383 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    4. Rollout          15/11                        MWC:  51.55% ( -0.22%)
       0.535 0.185 0.005 - 0.465 0.111 0.003 CL  51.16% CF  51.55%
      [0.001 0.001 0.000 - 0.001 0.000 0.000 CL   0.02% CF   0.03%]
        Truncated cubeful rollout (depth 11) with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861634383 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]


* I don't consider my rollout settings to be authoritative, so any result this close is suspect.  The point is that it's close, and the choice is what kind of game do we want to play.

[close]

[blitzxz]

* I don't consider my rollout settings to be authoritative, so any result this close is suspect.  The point is that it's close, and the choice is what kind of game do we want to play.

I don't recommend making truncated rollouts. It only takes about ten to 20 minutes to make full 0-ply so if I'm making rollout in the first place (to get more reliable result) I wouldn't spoil the results with truncation. I don't have the results any more but full 0-ply rollout makes hitting big favorite. I also recommend using EMG not MWC in checker moves. That gives a lot better picture of the size of the error.

In very close decision I would only accept full 2-ply rollouts with something like 10000 games to get statistical confidence.   That is propably too much to the most because it takes several days to make those but I want to be sure.  

[ah_clem]

I also recommend using EMG not MWC in checker moves. That gives a lot better picture of the size of the error.

I've looked into changing that setting, but I can't find it.  Do you know where the "switch" is in gnubg?  Thanks.

I'll look into doing full rollouts in the future.

[dorbel]

When you open a hint panel in gnu, there is a button bottom left marked MWC, click it.

[ah_clem]

When you open a hint panel in gnu, there is a button bottom left marked MWC, click it.

Thanks.  I looked in every menu item like options and preferences and settings and never could find it.

Here's a non-truncated non-MWC rollout

Spoiler

The prosaic move is still on top, but not by enough to say it's the "best" with any confidence. Maybe 2 ply     gives a different result?

    1. Rollout          24/21 15/14                  Eq.:  +0.296
       0.558 0.179 0.004 - 0.442 0.097 0.004 CL  +0.212 CF  +0.296
      [0.002 0.002 0.001 - 0.002 0.002 0.001 CL   0.005 CF   0.013]
        Full cubeful rollout with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861430848 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    2. Rollout          8/4*                         Eq.:  +0.286 ( -0.010)
       0.545 0.212 0.005 - 0.455 0.112 0.007 CL  +0.221 CF  +0.286
      [0.002 0.002 0.001 - 0.002 0.002 0.001 CL   0.006 CF   0.017]
        Full cubeful rollout with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861430848 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    3. Rollout          24/20                        Eq.:  +0.260 ( -0.036)
       0.554 0.171 0.004 - 0.446 0.101 0.004 CL  +0.191 CF  +0.260
      [0.002 0.002 0.000 - 0.002 0.002 0.001 CL   0.006 CF   0.014]
        Full cubeful rollout with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861430848 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]
    4. Rollout          15/11                        Eq.:  +0.239 ( -0.057)
       0.547 0.180 0.005 - 0.453 0.105 0.005 CL  +0.186 CF  +0.239
      [0.002 0.002 0.001 - 0.002 0.002 0.001 CL   0.006 CF   0.014]
        Full cubeful rollout with var.redn.
        1296 games, Mersenne Twister dice gen. with seed 861430848 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 0-ply cubeful prune [expert]

[close]

[Zorba]

ah_clem, I think you made a mistake somewhere in setting up the rollout. Perhaps you used a different matchscore, or the cube was placed wrong? Your first rollout shows us with over 50% MWC, that can't be right, it's more like 37%. I get similar results as blitzxz.


GnuBG rollout:

Spoiler

8/4* is clearly best, although plays 2 and 3 win more games overall. The extra ~4% gammons the loose hit creates make it worth it, even more so at this matchscore.

    1. Rollout          8/4*                         Eq.:  +0,4309
        54,70  21,29   0,62 -  45,30  11,75   1,09 CL  +0,2318 CF  +0,4309
      [  0,09   0,08   0,03 -   0,09   0,10   0,08 CL   0,0028 CF   0,0056]
    2. Rollout          24/21 15/14                  Eq.:  +0,3690 ( -0,0619)
        55,70  17,12   0,33 -  44,30  10,81   0,96 CL  +0,2422 CF  +0,3690
      [  0,09   0,08   0,02 -   0,09   0,10   0,07 CL   0,0027 CF   0,0057]
    3. Rollout          24/20                        Eq.:  +0,3513 ( -0,0796)
        55,56  16,52   0,29 -  44,44  10,82   0,86 CL  +0,2298 CF  +0,3513
      [  0,09   0,08   0,02 -   0,09   0,10   0,06 CL   0,0027 CF   0,0056]
    4. Rollout          15/11                        Eq.:  +0,3307 ( -0,1002)
        54,48  17,62   0,33 -  45,52  11,34   1,07 CL  +0,2169 CF  +0,3307
      [  0,09   0,08   0,03 -   0,09   0,10   0,07 CL   0,0027 CF   0,0058]
        Full cubeful rollout with var.redn.
        6480 games, Mersenne Twister dice gen. with seed 860861592 and quasi-random dice
        Play: 0-ply cubeful prune [expert]
        Cube: 2-ply cubeful prune [world class]

[close]

[ah_clem]

ah_clem, I think you made a mistake somewhere in setting up the rollout.

Yeah, the analysis and rollout are for a 0-0 match score.  I forgot to adjust it, although isn't the second string in the match supposed to include match score?  For some reason gnubg dosen't pick it up.