|
Forum Archive :
Opening Rolls
Rollouts of opening 21 and replies
|
|
Im rolling out all replies to the standard opening moves. The method is
as follows: I take an opening move, eg 21 played 13/11 24/23. Then I
take all replies from 11 to 66 and make rollouts for each move which is
evaluated by JF 3.0 level 7 within 0.05 equity of the best move (take or
give a move by my consideration). The rollouts are full cubeless 864
games level 6.
This is one of the hardest subjects for rollout studies because of some
reasons:
1) Opening rollouts lasts very long compared to other types of
positions.
2) A very high number of games is needed to get statistically
significant results.
3) Jellyfish evaluates these opening positions quite good, so why do
these rollouts?
As you can imagine, this takes a long time even on a fast computer. I
have a 233 MHz Pentium and a 133 MHz Pentium. One rollout lasts about
1.5 hours and 2.5 hours respectively. For about 1500 rollouts this
procedure needs only 10 weeks 24 hours a day :-). If someone out there
is willing to help me with these rollouts, tell me, and I am willing to
coordinate the work.
Now to the results for the replies to 21 if played 13/11 24/23:
The first two columns are the roll and the move (self explanatory). The
next three columns are the equity estimates of JF 3.0 level 6 and level
7 and the rollout result. An asterisk marks the best play in each column.
In the next six columns are the figures for wins, gammons and
backgammons for both players where player 1 is the player who won the
opening roll and player 2 is the replying player.
The column SD contains the standard deviation of the rollout equity.
Based on these SD and the rollout equities I have calculated the
probability that the rollout came out with the wrong best move (wrong
in a statistical sense of true infinite rollout result, not wrong in a
sense of perfect play on both sides).
Equities Player 2 Player 1
roll move L6 EV L7 EV L6 RO wins G&BG BG wins G&BG BG SD error
11 24/22 6/5(2) *.122 *.105 *.111 54.2 14.6 .7 45.8 12.2 .3 .010
11 24/23 6/5(3) .109 .099 .106 53.8 14.7 .6 46.2 11.7 .6 .010 36.2%
11 24/23(2) 6/5(2) .103 .090 .084 53.1 14.6 .6 46.9 12.4 .1 .010 2.8%
11 8/7(2) 6/5(2) .079 .064 .039 51.2 13.9 .7 48.8 12.6 .6 .011 .0%
11 8/5 6/5 .068 .058 .070 52.5 14.9 .7 47.5 13.0 .6 .010 .2%
8/7(2) 6/5(2) is clearly inferior, surprisingly close is the ugly 24/23
6/5(3). Maybe it is generally better to split to 23 than to 22.
21 24/21 -.099 *-.098 *-.099 46.4 10.1 .5 53.6 12.8 .5 .009
21 13/11 24/23 *-.099 -.101 -.117 46.0 11.3 .5 54.0 14.8 .8 .010 9.0%
21 24/22 24/23 -.115 -.113 -.129 45.5 10.5 .3 54.5 14.3 .4 .010 1.3%
21 13/10 -.129 -.132 -.124 45.4 11.3 .5 54.6 14.4 .6 .010 3.2%
24/21 came out clearly best, a surprise for me and maybe for level 6 too
;-)
22 13/11(2) 6/4(2) *.141 *.138 *.154 55.5 15.9 .8 44.5 11.8 .5 .011
22 24/22(2) 6/4(2) .113 .117 .125 54.1 13.9 .7 45.9 9.9 .4 .010 2.6%
22 24/20(2) .105 .104 .137 54.5 12.0 .5 45.5 7.6 .3 .009 11.6%
22 24/22 13/11 6/4(2) .089 .087 .110 53.9 15.2 .5 46.1 12.1 .4 .010 .2%
22 13/9 6/4(2) .090 .086 .062 52.1 14.7 .6 47.9 12.7 .7 .011 .0%
24/20(2) is more close to 13/11(2) 6/4(2) than I would have thought.
31 8/5 6/5 *.068 *.058 *.071 52.6 14.9 .7 47.4 13.0 .6 .010
32 24/21 13/11 *-.084 *-.080 -.100 46.9 11.3 .4 53.1 14.9 .7 .010 44.1%
32 13/10 24/22 -.091 -.095 -.100 46.6 11.8 .6 53.4 15.2 .4 .010 44.1%
32 24/21 24/22 -.099 -.098 *-.098 47.0 10.1 .5 53.0 13.8 .6 .009
32 13/8 -.116 -.112 -.146 44.8 11.3 .3 55.2 15.2 .6 .010 .0%
32 13/10 13/11 -.110 -.116 *-.098 46.5 12.3 .4 53.5 14.4 1.1 .010
32 24/21 6/4 -.147 -.144 -.148 45.5 10.0 .4 54.5 15.6 .7 .010 .0%
Here you have 4 equally good moves. Pick your choice!
33 24/21(2) 13/10(2) .139 *.148 *.180 56.1 13.8 .5 43.9 8.2 .2
.009
33 24/21(2) 6/3(2) *.140 .141 .155 55.0 13.6 .5 45.0 8.4 .3 .009 2.5%
33 13/10(2) 6/3(2) .119 .124 .114 54.0 15.7 .7 46.0 12.2 .8 .010 .0%
33 8/5(2) 6/3(2) .126 .117 .092 52.1 16.9 .7 47.9 12.1 .5 .010 .0%
33 24/18(2) .098 .089 .108 54.6 12.1 .4 45.4 10.5 .3 .009 .0%
Level 6 makes the wrong move here. This affects level 6 rollout for the
opening move as I will discuss below.
41 6/2* 2/1* *-.064 *-.069 *-.093 46.7 11.9 .5 53.3 14.3 .8 .009
41 13/9 24/23 -.089 -.092 -.109 46.1 12.0 .6 53.9 15.1 .6 .010 11.7%
41 13/8 -.116 -.112 -.144 44.9 11.3 .4 55.1 15.3 .6 .010 .0%
41 24/20 24/23 -.105 -.114 -.109 46.4 10.7 .4 53.6 14.4 .4 .010 11.7%
The double hit is best, but it is only close; it is worse than JF
estimates.
42 8/4 6/4 *.016 *.012 *.008 50.1 14.0 .7 49.9 13.5 .8 .010
43 13/9 24/21 *-.077 *-.077 *-.076 47.2 12.4 .5 52.8 14.3 .6 .010
43 24/20 13/10 -.079 -.091 -.083 47.3 12.1 .6 52.7 14.9 .7 .011 31.9%
43 24/20 24/21 -.104 -.094 -.088 47.1 10.4 .6 52.9 13.6 .5 .009 18.6%
43 13/9 13/10 -.097 -.095 -.112 45.6 13.1 .6 54.4 15.2 .9 .011 .8%
43 6/2* 24/21 -.101 -.105 -.081 47.0 11.7 .5 53.0 13.7 .7 .010 36.2%
6/2* 24/21 follows the rule Hit and Split. So ugly a hit on the 2 with
so many return hits is, this move yields the second place very close
behind the conventional 13/9 24/21.
44 24/20(2) 13/9(2) .277 *.276 *.308 60.8 14.7 .5 39.2 5.9 .2 .009
44 24/20(2) 6/2*(2) *.282 .275 .263 58.6 16.4 .6 41.4 7.6 .3 .010 .0%
44 8/4(2) 6/2*(2) .254 .253 .265 57.1 22.7 1.4 42.9 11.3 .5 .011 .1%
44 13/9(2) 6/2*(2) .240 .250 .219 56.7 18.7 1.2 43.3 11.0 .5 .010 .0%
Again a big error of JF level 6. Note: Trailing 2 away 1 away Crawford
you should very clearly play 8/4(2) 6/2*(2).
51 13/8 24/23 *-.067 *-.069 *-.071 47.6 12.0 .4 52.4 14.1 .7 .010
51 24/18 -.098 -.111 -.103 47.2 10.7 .5 52.8 15.4 .7 .011 1.6%
51 6/1* 24/23 -.101 -.115 -.127 45.5 11.2 .3 54.5 14.5 .7 .010 .0%
52 13/8 24/22 *-.065 *-.075 *-.089 47.5 10.7 .4 52.5 14.4 .6 .010
52 6/1* 24/22 -.088 -.093 -.125 45.5 10.9 .4 54.5 14.1 .6 .010 .5%
52 13/8 13/11 -.097 -.097 -.100 46.4 12.1 .6 53.6 14.9 .8 .010 21.8%
52 6/1* 13/11 -.112 -.118 -.121 46.0 12.1 .5 54.0 16.2 .7 .011 1.6%
52 13/6 -.125 -.127 -.147 44.7 11.7 .5 55.3 15.7 .7 .010 .0%
It isnt that good to hit on the ace point as I thought.
53 8/3 6/3 *-.031 *-.033 *-.034 48.6 13.7 .8 51.4 14.3 .6 .010
53 13/8 24/21 -.050 -.044 -.053 47.9 11.7 .6 52.1 13.0 .5 .009 7.9%
54 13/8 24/20 *-.050 *-.067 *-.075 47.5 11.8 .5 52.5 14.2 .6 .010
54 13/8 13/9 -.089 -.086 -.127 45.3 11.8 .6 54.7 15.0 .7 .010 .0%
54 24/15 -.068 -.089 -.085 47.9 10.7 .4 52.1 15.0 .4 .010 24.0%
54 13/8 6/2* -.098 -.101 -.110 46.2 11.4 .6 53.8 14.7 .7 .009 .5%
54 6/1* 24/20 -.088 -.107 -.102 46.4 11.5 .3 53.6 14.4 .4 .010 2.8%
54 6/1* 13/9 -.101 -.109 -.136 45.2 12.2 .7 54.8 15.7 1.1 .010 .0%
In this situation you must get your back checkers going. 13/8 13/9 is
clearly inferior.
55 8/3(2) 6/1*(2) *.164 *.133 *.098 51.5 19.6 .6 48.5 12.8 .6 .011
55 13/8(2) 6/1*(2) .121 .112 .087 52.2 17.2 .6 47.8 12.8 .5 .010 23.0%
55 13/3(2) .066 .078 .067 51.6 16.2 .7 48.4 12.9 .5 .010 1.9%
This is a really big surprise: The most ugly 13/8(2) 6/1*(2) is darn
close to the more natural 8/3(2) 6/1*(2). In double match point this is
even the best move!
61 13/7 8/7 *-.005 *-.010 *-.010 50.2 12.8 .5 49.8 14.1 .7 .010
62 24/18 13/11 *-.076 *-.083 *-.075 48.0 11.3 .5 52.0 14.8 .6 .010
62 24/16 -.087 -.098 -.113 46.9 10.5 .6 53.1 15.7 .6 .010 .4%
62 8/2* 24/22 -.120 -.122 -.158 44.2 11.0 .5 55.8 15.0 .8 .010 .0%
62 24/18 24/22 -.115 -.131 -.121 46.7 11.2 .6 53.3 16.7 .5 .010 .1%
I expected a close result, but 24/18 13/11 is the clear winner.
63 24/18 13/10 -.071 *-.086 -.090 47.5 11.2 .4 52.5 14.8 .8 .010 36.2%
63 24/15 *-.068 -.089 *-.085 47.9 10.7 .4 52.1 15.0 .4 .010
63 24/18 24/21 -.112 -.107 -.088 47.5 11.1 .3 52.5 14.8 .5 .010 41.6%
Here the results are really close, you can choose either move.
64 24/14* *.140 *.136 *.129 55.1 13.7 .7 44.9 11.2 .5 .011
65 24/13 *-.002 *-.018 *-.058 49.3 9.6 .3 50.7 14.0 .3 .009
66 24/18(2) 13/7(2) *.363 *.352 *.374 63.9 16.3 .6 36.1 7.1 .2
.009
As a side effect of these opening reply rollouts I am able to calculate
equities for the opening moves with 18144 games played on level 6, and
so with an incredible low standard deviation. The calculated equity is
even more precise than a normal level 6 rollout, because I take the best
replies to the openings based on above rollouts rather than the replies
of level 6 in a normal level 6 rollout.
Rollout result for 21 (13/11 24/23) with best replies: .002
Rollout result for 21 (13/11 24/23) with replies of level 6: .005
Using the best replies gains an equity of .003 compared to level 6 play.
The equity of .002 now may be underestimating for the opener because
only the replies are corrected but not the replies to the replies and so
on.
-- Alexander (acey_deucey@FIBS)
|
|
|
|
Opening Rolls
- At different match scores (Louis Nardy Pillards, July 2002)
- Average advantage of winning opening roll (Chuck Bower, Oct 1998)
- Choosing a strategy (Daniel Murphy, June 2001)
- Early game rule of thumb (Rich Munitz, Feb 2009)
- Factors to consider (Kit Woolsey, July 1994)
- How computers play (Kit Woolsey, Mar 1995)
![[Recommended reading]](pics/Star2.gif)
- Magriel's Chapter 5 (Hayden Alfano+, May 2006)
![[Long message]](pics/Long.gif)
- Mloner vs Jellyfish (Kit Woolsey, Dec 1995)
- Nactating a whole game (Nack Ballard+, Jan 2011)
- Nactation (Jim Stutz+, June 2010)
- Nactation overview (Nack Ballard, Oct 2009)
- Nactation--Why use it? (leobueno+, Jan 2011)
- Opening 1's: Split or slot? (Douglas Zare, Dec 2003)
- Opening 21: Rollout (Stick, Mar 2006)
![[GammOnLine forum]](pics/GammOnLine.gif)
- Opening 21: Split or slot? (Dick Adams+, Dec 2003)
- Opening 32: Rollout (Stick, Feb 2006)
- Opening 43: In GOL online match (Raccoon+, Feb 2004)
- Opening 43: Pros and cons (Stick+, Jan 2006)
- Opening 43: Which split is better? (Peter Backgren+, Aug 2000)
- Opening 43: Which split is better? (Michael J. Zehr+, Mar 1996)
- Opening 51: Rollout (Stick, Feb 2006)
- Opening 52: Merits of splitting (Peter Bell, Apr 1995)
- Opening 53: Magriel's recommendation (George Parker+, July 1997)
- Opening 53: Split to 21? (Alex Zamanian, Aug 2000)
- Opening 53: Why make the three point? (Kit Woolsey+, Feb 1996)
- Opening 6's: Slot the bar point? (Chuck Bower+, Feb 2000)
- Opening 6's: Slot the bar point? (David Montgomery, June 1995)
- Opening 62: Could running be best? (Gary Wong, Sept 1997)
- Opening 62: Split, run, or slot? (Chuck Bower, May 1997)
- Opening 63: Middle Eastern split? (Mark+, Apr 2002)
- Opening 63: Slot the four point? (Dennis Cartwright+, Mar 2002)
- Opening 64: Make the two point? (William Hill+, Jan 1998)
- Opening 64: Make the two point? (Darse Billings, Feb 1995)
- Opening 64: Rollout (Peter Grotrian, Jan 2006)
- Opening 64: Split to 20? (Peter Bell, June 1995)
- Opening 64: Three choices (Brian Sheppard, July 1997)
- Opening 65: Becker on lover's leap (Jeffrey Spiegler+, Aug 1991)
- Opening 65: Computer rankings (Chuck Bower, Jan 1997)
- Opening rolls ranked (Arthur+, Apr 2005)
- Rollouts of opening 21 and replies (Alexander Nitschke, Oct 1997)
- Rollouts of openings (Tom Keith+, Jan 2006)
- Rollouts: Expert Backgammon (Tom Fahland, Aug 1994)
- Rollouts: Jellyfish 3.0 (Midas+, Sept 1997)
- Rollouts: Jellyfish 3.0 level 6 (Chuck Bower, Feb 1999)
- Rollouts: Snowie 4.1 (Rene Cerutti, Apr 2004)
- Slotting the four point (Joe Loria+, Oct 1999)
- Snowie's openers and replies (rcerutti, Feb 1999)
- Splitting versus building (Dave Slayton+, Aug 2000)
- Splitting versus slotting (Daniel Murphy, Apr 2001)
- Splitting versus slotting (Daniel Murphy, Sept 1997)
- Trice's rankings (Marty Storer, Feb 1992)
From GammOnLine
Long message
Recommended reading
![[Recent addition]](pics/New.gif) Recent addition
|
| |
|
