Backgammon Rollouts

Rollouts

Level-5 versus level-6 rollouts

From:   Chuck Bower
Address:   bower@bigbang.astro.indiana.edu
Date:   20 August 1997
Subject:   Re: An endgame problem
Forum:   rec.games.backgammon
Google:   5tfs86$6cj$1@dismay.ucs.indiana.edu

1) For rollouts, I go straight to level-6. Occasionally I'll use level-5 (with cube) after adjusting the settlement limit based on the level-6 result. THEN I MAKE SURE that the level-5 cubeless results are consistent with the (already completed benchmark) level-6 (cubeless). > What do you mean by "*make sure* (they) are consistent"? My level-5 > cubeless rollout result is not consistent with the level-6 rollout > result, since three out of four plays go up by about .02, while the last > one is up by .06. It only tells me that there's a difference in the > playing style between level 5 and level 6, and that I should trust the > level 6 rollouts more when it comes to checkers play. One problem is > though that there's usually a cube somewhere in this game, and level 6 > doesn't know anything about that. OK. If the level-5 rollout isn't consistent with level-6 rollout, don't believe level-5. (By "consistent" I mean "statistically consistent".) As far as "knowing" about the cube, one of us may be confused (or maybe neither) but here is the way I understand JF's checker play decisions: JF ASSUMES that the game is cubeless (always played to the end) and makes the play which maximizes equity in this case. (Note that gammons count twice simple games, and backgammons thrice, and JF uses these facts.) Human BG players sometimes think something like this: I have a legitimate cash (double, pass) next time if I make play A. Play B is probably better IF the game continues, but there are risks there that may allow a legitimate take (or even NO DOUBLE). Go with the "safe" sure thing. If my opp surprises me by taking, all the better! To my knowledge, JFv3.0 does not use this kind of thinking. Now the question arises "suppose I'm rolling out a CUBE position" (that is, the question is Double or not, and/or drop or take). Then it MAY help to have JF use the cube. BUT, Jellyfish only has a limited cube algorithm in its rollouts, and these only at level-5 (inferior player). "Limited" means that THERE CAN NEVER BE A TAKE. If a side has access to the cube AND is above the "settlement limit", then that sides scores a simple win. (NOTE: there is an exception if a large swing has occurred over the last 1296 dice chances AND there is a high enough gammon chance. This doesn't come up often and for this discussion can be ignored.) Here JF will roll the game out to the bitter end, but keep track of FOUR SEPARATE SITUATIONS: cubeless, centered cube, player on roll owns cube, and player NOT on roll owns cube. In the last three, the game is "ended" when a player on roll has access to the cube AND is above the settlement limit. "Ended" just means the game is scored as a win for that player, but the rollout ALWAYS continues until one side has 15 checkers off so that the "cubeless" statistics can be tabulated. SO, if you first do a level-6 rollout, (adjust the "settlement limit" if you so choose) and then do level-5 "limited cube" rollout, you can compare the level-6 result with the level-5 CUBELESS result. If they disagree, DON'T BELIEVE ANY OF THE FOUR LEVEL-5 RESULTS. The fact is, you can use the level-6 cubeless result to determine the proper CUBEFUL decisions (double/NO; drop/take), AND you can do it for money play and match scores. How do you do this? You need to absorb my previous posts. It takes too long to explain right now. > Another thing I do not adjust is the settlement limit. I think it's only > needed in short/long races and positions with high gammon rates (that > equals about all positions now, doesn't it?), otherwise, about. 0.550 > will do nicely. This is mainly because I do not know how to adjust it > (shoot me if you like, but please tell me how I should do it while you > kill me), and partly because I do not find it incredibly important. I don't like to shoot good analysts. (I don't even like shooting BAD ones, but that may sometimes prove necessary....) This "settlement limit" thing is not a closed book, although Kit seemed to say that it was in a post a couple weeks back. Maybe I'm the only one getting my fingers pinched, though... My method of determining settlement limit has probably not been posted. Unfortunately I'm already late for dinner. Let me get back to that one. But in short, I do a level-6 rollout, use those results and match equities (if applicable) to find the settlement limit, tweek it a bit (Bob Koca won't like this step!) and then plug it in to a level-5 limited cube rollout and go from there. Stay tuned... Chuck bower@bigbang.astro.indiana.edu c_ray on FIBS

Did you find the information in this article useful?

Do you have any comments you'd like to add?