The cross-entropy tuner

competition_type string: "ce_tuner".

The cross-entropy tuner uses the cross-entropy method described in [CE]:

[CE] G.M.J-B. Chaslot, M.H.M Winands, I. Szita, and H.J. van den Herik.
Cross-entropy for Monte-Carlo Tree Search. ICGA Journal, 31(3):145-156.


The cross-entropy tuner is experimental. It can take a very large number of games to converge.

The tuning algorithm

The algorithm is not described in detail in this documentation. See [CE] section 3 for the description. The tuner always uses a Gaussian distribution. The improvement suggested in section 5 is not implemented.

The parameter model

The parameter values taken from the Gaussian distribution are floating-point numbers known as optimiser parameters.

These parameters can be transformed before being used to configure the candidate (see 3.3 Normalising Parameters in [CE]). The transformed values are known as engine parameters. The transformation is implemented using a Python transform function defined in the control file.

Reports show engine parameters (see the format parameter setting), together with the mean and variance of the corresponding optimiser parameter distribution in the form mean~variance.

Sample control file

Here is a sample control file, illustrating most of the available settings for a cross-entropy tuning event:

competition_type = "ce_tuner"

description = """\
This is a sample control file.

It illustrates the available settings for the cross entropy tuner.

players = {
    'gnugo-l10' : Player("gnugo --mode=gtp --chinese-rules "
                         "--capture-all-dead --level=10"),

def fuego(max_games, additional_commands=[]):
    commands = [
        "go_param timelimit 999999",
        "uct_max_memory 350000000",
        "uct_param_search number_threads 1",
        "uct_param_player reuse_subtree 0",
        "uct_param_player ponder 0",
        "uct_param_player max_games %d" % max_games,
    return Player(
        "fuego --quiet",


def exp_10(f):
    return 10.0**f

parameters = [
              # Mean and variance are in terms of log_10 (rave_weight_initial)
              initial_mean = -1.0,
              initial_variance = 1.5,
              transform = exp_10,
              format = "I: %4.2f"),

              # Mean and variance are in terms of log_10 (rave_weight_final)
              initial_mean = 3.5,
              initial_variance = 1.5,
              transform = exp_10,
              format = "F: %4.2f"),

def make_candidate(rwi, rwf):
    return fuego(
        ["uct_param_search rave_weight_initial %f" % rwi,
         "uct_param_search rave_weight_final %f" % rwf])

board_size = 9
komi = 7.5
opponent = 'gnugo-l10'
candidate_colour = 'w'

number_of_generations = 5
samples_per_generation = 100
batch_size = 10
elite_proportion = 0.1
step_size = 0.8

Control file settings

The following settings can be set at the top level of the control file:

All common settings (the players dictionary is required, though it is used only to define the opponent).

The following game settings (only board_size and komi are required):

The following additional settings (they are all required):


String: "b", "w", or "random"

The colour for the candidates to take in every game.



The player code of the player to use as the candidates’ opponent.


List of Parameter definitions (see Parameter configuration).

Describes the parameters that the tuner will work with. See The parameter model for more details.

The order of the Parameter definitions is used for the arguments to make_candidate, and whenever parameters are described in reports or game records.


Python function

Function to create a Player from its engine parameters.

This function is passed one argument for each candidate parameter, and must return a Player definition. Each argument is the output of the corresponding Parameter’s transform.

The function will typically use its arguments to construct command line options or GTP commands for the player. For example:

def make_candidate(param1, param2):
    return Player(["goplayer", "--param1", str(param1),
                   "--param2", str(param2)])

def make_candidate(param1, param2):
    return Player("goplayer", startup_gtp_commands=[
                   ["param1", str(param1)],
                   ["param2", str(param2)],

Positive integer

The number of times to repeat the tuning algorithm (number of iterations or T in the terminology of [CE]).


Positive integer

The number of candidates to make in each generation (population_size or N in the terminology of [CE]).


Positive integer

The number of games played by each candidate.


Float between 0.0 and 1.0

The proportion of candidates to select from each generation as ‘elite’ (the selection ratio or ρ in the terminology of [CE]). A value between 0.01 and 0.1 is recommended.


Float between 0.0 and 1.0

The rate at which to update the distribution parameters between generations (α in the terminology of [CE]).


I can’t find anywhere in the paper the value they used for this, so I don’t know what to recommend.

Parameter configuration

A Parameter definition has the same syntax as a Python function call: Parameter(arguments). Apart from code, the arguments should be specified using keyword form (see Sample control file).

The code, initial_mean, and initial_variance arguments are required.

The arguments are:



A short string used to identify the parameter. This is used in error messages, and in the default for format.



The mean value for the parameter in the first generation’s distribution.


Float >= 0

The variance for the parameter in the first generation’s distribution.


Python function (default identity)

Function mapping an optimiser parameter to an engine parameter; see The parameter model.


def exp_10(f):
    return 10.0**f

Parameter('p1', initial_mean = …, initial_variance = …,
          transform = exp_10)

If the transform is not specified, the optimiser parameter is used directly as the engine parameter.


String (default "parameter_code: %s")

Format string used to display the parameter value. This should include a short abbreviation to indicate which parameter is being displayed, and also contain %s, which will be replaced with the engine parameter value.

You can use any Python conversion specifier instead of %s. For example, %.2f will format a floating point number to two decimal places. %s should be safe to use for all types of value. See string formatting operations for details.

Format strings should be kept short, as screen space is limited.


          initial_mean = 0.0, initial_variance = 1.0,
          format = "p1: %.2f")

          initial_mean = 5000, initial_variance = 250000,
          format = "p2: %d")


Currently, there aren’t any sophisticated reports.

The standard report shows the parameters of the current Gaussian distribution, and the number of wins for each candidate in the current generation.

After each generation, the details of the candidates are written to the history file. The candidates selected as elite are marked with a *.

Changing the control file between runs

Some settings can safely be changed between runs of the same cross-entropy tuning event:

safe to increase
not safe to change
safe to change
safe to change
safe to change
safe to change, but don’t alter play-affecting options
not safe to change
safe to change