[Bug-gnubg] gnubg.bd

[Joern Thyssen]

Hi all,

I would very much like to do some changes to gnubg.bd. Of course, I'll
try to keep backwards compability.

Very shortly, gnubg will be able to read large one-sided databases
and large two-sided databases. To keep the interface general I would
like to split gnubg.bd into two separate files: one for the one-sided
database and one for the two-sided database. For example, gnubg will use
four databases:
    - gnubg_osa.bd  (small one-sided bearoff database read into memory)
    - gnubg_osb.bd  (huge one-sided bearoff database used from disk)
    - gnubg_tsa.bd  (small two-sided bearoff database read into memory)
    - gnubg_tsb.bd  (huge two-sided bearoff database used from disk)

Also, gnubg_osa.bd will be re-generated to include gammon probabilities
and gnubg_osb.bd will be re-generated to include cubeful equities for
money game.

An alternative to gnubg_osa.bd is to include it directly into the source
as Josephs does with fibs2html/sagnubg. This will increase the
executable size and memory footprint with up to 6 MB (can be decreased
to 2 MB if my compression scheme is used). 

One can play the same trick with gnubg_osb.bd but the two-sided database
is not as necesary as the one-sided which is used to calculate bg
probabilities for the race net (and it will be used for one-sided
rollouts).

Once my code is finished we will get:

- exact cubeful equities for money game (with the two-sided database)
- huge one-sided databases (I'm not sure this is an improvement as the
  race net is quite good)
- gnubg for windows will be distributed with gnubg_osa.bd and
  gnubg_tsa.bd, but programs will be included that can generate
  arbitary-size bearoff databases for gnubg, only limited by the user's
  patience and disk space.[*]

[*] Features of the new makebearoff programs:

- uses a cache instead of malloc'ing the entire bearoff database
   (this will of course slow down generation, but will make it possible
   to generate huge databases without owning 1 GB of memory)
  The cache can be improved a lot, for example, a better hash function
  on position IDs is needed.
- can use previosly generated databases. For example, the 12 pt
  databases has 37M positions, but the 17M is already in the 11 pt
  database. 
- cubeful/cubeless two-sided bearoff databases with n chequers on m
  points (1<=m<=6)
- one-sided bearoff databases with bearoff and gammon distributions
  (either exact distributions or distributions approximated by normal
  distributions) for 15 chequers on m points (1<=m<=23), although the
  12-14 points seems to be the practical limit. Also, the exact
  distributions can be compressed in the file storing only non-zero
  elements.

So far I've generated the 12 pt one-sided database (approximated with
normal dist). It's 300 MB and takes 15 hours to generate on my 1Ghz
laptop. 

With respect to one-sided bearoff databases, it might be possible to
generate some hybrids, e.g., less than 5 chequers on the outfield is 18M
positions. I'll investigate that.

Jørn

-- 
Joern Thyssen, PhD
Vendsysselgade 3, 3., DK-9000 Aalborg, Denmark
+45 9813 2791 (private) / +45 2818 0183 (mobile) / +45 9633 7036 (work)
Note: new mobile number!