GNU rollout program

Usage

sagnubg [flags] [cmd-file] [output-file]
Process `cmd-file' and write results to `output-file'. `cmd-file' and `output-file' default to stdin and stdout respectively.

Command Line Flags

flagfull namedescription
-w--weights-dir=FILE Explicitly specify name of NN's files. Default is to use the environment variable GNUBGWEIGHTS if it exists, otherwise use `gnubg.weights' in the same directory as the executable.

--moves2p-limit=N Maximum number of moves to keep after 0-ply pruning. Default is 20.

--rollout-limit=N Maximum number of moves to rollout. Default is 5.

--rollout-games=N Number of games per rollout. Default is 1296.

--cube-away=N Use (N,N) away for cube decisions. Default is 7.

--include-ply0=1/0 Force inclusion of top 0-ply move. Default is 1.

--eval-plies=N Ply used in evaluation (`e' command).

--n-osr=N Number of games for OSR evaluation (`O' command).

--no-shortcuts Disable usage of small net pruning.

--resume Resume an interrupted run. Both `cmd-file' and `output-file' must be given.

-v--verbose Print status report to stderr during run.
-h--help

Command File Format

Command file is a simple ASCII file. Each command occupies one line, where the first character specifies command type.

# ...
Comment. Echoed to output.

s SETTINGS
Change run settings in the middle.
s nRollOutGames 12960
r N
Set randomizer seed to N for next `m' or `o' command. Echoed to output.

m position dice0 dice1
Rollout top 2-ply moves. Output is

m position dice0 dice1 move1 eqt move2 dif2 move3 dif3 ...
eqt is the (money) equity of the best move, and difi if the equity difference of movei from best.

c position
Perform Cube-full Rollout of position with centered cube and 2 cube follwong a Double/Take. Output is

c position Pnd Pdt action
Pnd and Pdt are match win%, in the range 0 to 100. action is one of `TG',`D/T',`D/D' or `ND'.

o position
Rollout position. Output is
o position p0 p1 p2 p3 p4
Where the pi's are the GNUbg net evaluation probabilities.

e position
Evaluate position using the net.

Output is,

e position p0 p1 p2 p3 p4
O position
Evaluate race position using the OSR method.

Output is,

O position p0 p1 p2 p3 p4
b position dice0 dice1
Find evaluation best move.

Output is,

b position dice0 dice1 move p0 p1 p2 p3 p4
move is the position after the move, pi are the position evaluation.

Note that output has the same format as the input, i.e. can serve as input for another run.

The positions are encoded as 20 character in the range of `A' to `F'. Each letter stands for 4 bits (A == 0x0, F == 0xf), and concatenated they form the 80 bit GNUbg position ID.