Daniel Murphy wrote a very nice article on Jellyfish's style and
strengths and weaknesses, with a listing of several mistakes that
Jellyfish makes, mostly from the bearoff where we can calculate the
correct actions. I just want to comment on one paragraph Daniel
wrote on the relative importance of cube and checker play.
> Bad takes -- positional judgement and cube strategy are the most
> challenging aspects of modern backgammon.
I disagree. I think checker play is more challenging. Why?
1) You are faced with many, many more tough checker plays than
cube actions per game.
2) Even beyond the surplus of tough checker plays that occur in
a game, there are many more tough checker plays than cube
actions. Here's an exercise. Take a few interesting doubling
positions. Now, go through all the rolls for the leader and
see how you would play them. You will find that for each
interesting doubling decision, there are several interesting
checker plays. And of course most positions that are not
interesting with respect to the cube have several interesting
checker plays on particular rolls.
3) Prior to the availability of computer rollouts, checker play
theory necessarily lagged far behind cube theory because it
is much more difficult to roll out checker plays. You have to
roll out several candidates, and they are probably close.
4) Even now that we have computer rollouts, it is much more difficult
to investigate checker plays. You typically have to rollout
three times as many positions, and each has to be rolled out
for more games.
5) Beyond the extra effort of doing the more rollouts, at this point
you have to invent theory because there is so little out there.
6) There are many more kinds of decisions in checker plays. Most
interesting cube decisions are (re)double/no (re)double decisions
or take/drop decisions. In checker plays, you have questions
on whether to break an anchor, to make an offensive or defensive
point, to hit loose or hit and cover, etc. Generally, the same
kinds of factors go into both checker plays and cube decisions
-- how strong is my board? how strong is my opponents? how flexible
am I? will I be able to clear this point? etc. -- but with
checker plays you are applying them to several kinds of decisions.
(One might protest, ah but you have take/drop decisions in holding
games, backgames, blitzes, etc. I would respond that for each of
these kinds of games, there are many types of checker play decisions.)
7) When you don't have a model for cube actions, you can often
get by using a reference position. But reference positions are
much less useful for checker plays. This is because the number of
types of decisions is greater, and since checker plays are often
closer than cube decisions, small inaccuracies in your mental
adjustments to your reference positions are more likely to
lead to the wrong decision.
> It's not easy to learn to play the checkers competitively, but
> it's much more difficult, especially in match play, to double
> neither too soon nor too late, and to drop or take correctly.
We haven't really defined "competitively", but lets look at it
this way -- consider the class A of players who on average maintain
a rating of X. Let's say you currently play at a lower level than
players in A. Is it easier to get your checker play up to the level
of A players or to get your cube action up to the A level?
I think it is easier to get your cube play up to this new level,
although I'm not sure. I think there is less you need to study (i.e.,
you can often get by with reference positions) and there is less need
for an appreciation of subtle changes in a position. Also, for
players who don't study, I suspect their checker play is better
than their cube play (since they get more practice at it, and
comments on it in chouettes), so players who do study can have
relatively better cube play.
Regarding match adjustments, it is certainly true that cube play
becomes much tougher, but so does checker play. Good players have
a strong appreciation for how to change their cube action based
on a match situation, but I believe they are worse at adjusting
their checker play.
> Among advanced and even expert players, errors with the cube
> tend to be much more costly than mistakes in checker play.
This is true when you compare a single checker play error to
a single cube error. But since there are many more checker play
errors per game, I believe that in total, human experts (and
everyone else) give up more equity due to checker play mistakes
than due to cube errors.
> Perhaps this -- cube handling -- is JellyFish's biggest advantage
> over human players, including experts.
Perhaps, but especially against experts, I think it's the other
way around. I think Jellyfish's bigger advantage against strong
players is in checker play.
(In matches I think Jellyfish does adjust its cube action better than
human players, because it is uniformly good at estimating probabilities,
whereas human players are more keenly attuned to money doubling and
taking points. Where Jellyfish adjusts its checker play evaluations,
at DMP, Crawford, post-Crawford, and some other situations, I'm certain
it does a better job of adjusting its play that almost everyone. Where
it doesn't, humans can get a small edge.)
I think checker play theory is the frontier in backgammon research.
monty on FIBS