Select Git revision
ComputingTasks.hh
Forked from
Dynare / dynare
Source project has a limited visibility.
cmaes.m 113.19 KiB
function [xmin, ... % minimum search point of last iteration
fmin, ... % function value of xmin
counteval, ... % number of function evaluations done
stopflag, ... % stop criterion reached
out, ... % struct with various histories and solutions
bestever ... % struct containing overall best solution (for convenience)
] = cmaes( ...
fitfun, ... % name of objective/fitness function
xstart, ... % objective variables initial point, determines N
insigma, ... % initial coordinate wise standard deviation(s)
inopts, ... % options struct, see defopts below
varargin ) % arguments passed to objective function
% cmaes.m, Version 3.56.beta, last change: February, 2012
% CMAES implements an Evolution Strategy with Covariance Matrix
% Adaptation (CMA-ES) for nonlinear function minimization. For
% introductory comments and copyright (GPL) see end of file (type
% 'type cmaes'). cmaes.m runs with MATLAB (Windows, Linux) and,
% without data logging and plotting, it should run under Octave
% (Linux, package octave-forge is needed).
%
% OPTS = CMAES returns default options.
% OPTS = CMAES('defaults') returns default options quietly.
% OPTS = CMAES('displayoptions') displays options.
% OPTS = CMAES('defaults', OPTS) supplements options OPTS with default
% options.
%
% XMIN = CMAES(FUN, X0, SIGMA[, OPTS]) locates the minimum XMIN of
% function FUN starting from column vector X0 with the initial
% coordinate wise search standard deviation SIGMA.
%
% Input arguments:
%
% FUN is a string function name like 'myfun'. FUN takes as argument a
% column vector of size of X0 and returns a scalar. An easy way to
% implement a hard non-linear constraint is to return NaN. Then,
% this function evaluation is not counted and a newly sampled
% point is tried immediately.
%
% X0 is a column vector, or a matrix, or a string. If X0 is a matrix,
% mean(X0, 2) is taken as initial point. If X0 is a string like
% '2*rand(10,1)-1', the string is evaluated first.
%
% SIGMA is a scalar, or a column vector of size(X0,1), or a string
% that can be evaluated into one of these. SIGMA determines the
% initial coordinate wise standard deviations for the search.
% Setting SIGMA one third of the initial search region is
% appropriate, e.g., the initial point in [0, 6]^10 and SIGMA=2
% means cmaes('myfun', 3*rand(10,1), 2). If SIGMA is missing and
% size(X0,2) > 1, SIGMA is set to sqrt(var(X0')'). That is, X0 is
% used as a sample for estimating initial mean and variance of the
% search distribution.
%
% OPTS (an optional argument) is a struct holding additional input
% options. Valid field names and a short documentation can be
% discovered by looking at the default options (type 'cmaes'
% without arguments, see above). Empty or missing fields in OPTS
% invoke the default value, i.e. OPTS needs not to have all valid
% field names. Capitalization does not matter and unambiguous
% abbreviations can be used for the field names. If a string is
% given where a numerical value is needed, the string is evaluated
% by eval, where 'N' expands to the problem dimension
% (==size(X0,1)) and 'popsize' to the population size.
%
% [XMIN, FMIN, COUNTEVAL, STOPFLAG, OUT, BESTEVER] = ...
% CMAES(FITFUN, X0, SIGMA)
% returns the best (minimal) point XMIN (found in the last
% generation); function value FMIN of XMIN; the number of needed
% function evaluations COUNTEVAL; a STOPFLAG value as cell array,
% where possible entries are 'fitness', 'tolx', 'tolupx', 'tolfun',
% 'maxfunevals', 'maxiter', 'stoptoresume', 'manual',% 'warnconditioncov', 'warnnoeffectcoord', 'warnnoeffectaxis',
% 'warnequalfunvals', 'warnequalfunvalhist', 'bug' (use
% e.g. any(strcmp(STOPFLAG, 'tolx')) or findstr(strcat(STOPFLAG,
% 'tolx')) for further processing); a record struct OUT with some
% more output, where the struct SOLUTIONS.BESTEVER contains the overall
% best evaluated point X with function value F evaluated at evaluation
% count EVALS. The last output argument BESTEVER equals
% OUT.SOLUTIONS.BESTEVER. Moreover a history of solutions and
% parameters is written to files according to the Log-options.
%
% A regular manual stop can be achieved via the file signals.par. The
% program is terminated if the first two non-white sequences in any
% line of this file are 'stop' and the value of the LogFilenamePrefix
% option (by default 'outcmaes'). Also a run can be skipped.
% Given, for example, 'skip outcmaes run 2', skips the second run
% if option Restarts is at least 2, and another run will be started.
%
% To run the code completely silently set Disp, Save, and Log options
% to 0. With OPTS.LogModulo > 0 (1 by default) the most important
% data are written to ASCII files permitting to investigate the
% results (e.g. plot with function plotcmaesdat) even while CMAES is
% still running (which can be quite useful on expensive objective
% functions). When OPTS.SaveVariables==1 (default) everything is saved
% in file OPTS.SaveFilename (default 'variablescmaes.mat') allowing to
% resume the search afterwards by using the resume option.
%
% To find the best ever evaluated point load the variables typing
% "es=load('variablescmaes')" and investigate the variable
% es.out.solutions.bestever.
%
% In case of a noisy objective function (uncertainties) set
% OPTS.Noise.on = 1. This option interferes presumably with some
% termination criteria, because the step-size sigma will presumably
% not converge to zero anymore. If CMAES was provided with a
% fifth argument (P1 in the below example, which is passed to the
% objective function FUN), this argument is multiplied with the
% factor given in option Noise.alphaevals, each time the detected
% noise exceeds a threshold. This argument can be used within
% FUN, for example, as averaging number to reduce the noise level.
%
% OPTS.DiagonalOnly > 1 defines the number of initial iterations,
% where the covariance matrix remains diagonal and the algorithm has
% internally linear time complexity. OPTS.DiagonalOnly = 1 means
% keeping the covariance matrix always diagonal and this setting
% also exhibits linear space complexity. This can be particularly
% useful for dimension > 100. The default is OPTS.DiagonalOnly = 0.
%
% OPTS.CMA.active = 1 turns on "active CMA" with a negative update
% of the covariance matrix and checks for positive definiteness.
% OPTS.CMA.active = 2 does not check for pos. def. and is numerically
% faster. Active CMA usually speeds up the adaptation and might
% become a default in near future.
%
% The primary strategy parameter to play with is OPTS.PopSize, which
% can be increased from its default value. Increasing the population
% size (by default linked to increasing parent number OPTS.ParentNumber)
% improves global search properties in exchange to speed. Speed
% decreases, as a rule, at most linearely with increasing population
% size. It is advisable to begin with the default small population
% size. The options Restarts and IncPopSize can be used for an
% automated multistart where the population size is increased by the
% factor IncPopSize (two by default) before each restart. X0 (given as
% string) is reevaluated for each restart. Stopping options
% StopFunEvals, StopIter, MaxFunEvals, and Fitness terminate the
% program, all others including MaxIter invoke another restart, where
% the iteration counter is reset to zero.
%
% Examples:
%
% XMIN = cmaes('myfun', 5*ones(10,1), 1.5); starts the search at% 10D-point 5 and initially searches mainly between 5-3 and 5+3
% (+- two standard deviations), but this is not a strict bound.
% 'myfun' is a name of a function that returns a scalar from a 10D
% column vector.
%
% opts.LBounds = 0; opts.UBounds = 10;
% X=cmaes('myfun', 10*rand(10,1), 5, opts);
% search within lower bound of 0 and upper bound of 10. Bounds can
% also be given as column vectors. If the optimum is not located
% on the boundary, use rather a penalty approach to handle bounds.
%
% opts=cmaes; opts.StopFitness=1e-10;
% X=cmaes('myfun', rand(5,1), 0.5, opts); stops the search, if
% the function value is smaller than 1e-10.
%
% [X, F, E, STOP, OUT] = cmaes('myfun2', 'rand(5,1)', 1, [], P1, P2);
% passes two additional parameters to the function MYFUN2.
%
% Copyright (C) 2001-2012 Nikolaus Hansen,
% Copyright (C) 2012 Dynare Team
%
% This file is part of Dynare.
%
% Dynare is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% Dynare is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with Dynare. If not, see <http://www.gnu.org/licenses/>.
cmaVersion = '3.60.beta';
% ----------- Set Defaults for Input Parameters and Options -------------
% These defaults may be edited for convenience
% Input Defaults (obsolete, these are obligatory now)
definput.fitfun = 'felli'; % frosen; fcigar; see end of file for more
definput.xstart = rand(10,1); % 0.50*ones(10,1);
definput.sigma = 0.3;
% Options defaults: Stopping criteria % (value of stop flag)
defopts.StopFitness = '-Inf % stop if f(xmin) < stopfitness, minimization';
defopts.MaxFunEvals = 'Inf % maximal number of fevals';
defopts.MaxIter = '1e3*(N+5)^2/sqrt(popsize) % maximal number of iterations';
defopts.StopFunEvals = 'Inf % stop after resp. evaluation, possibly resume later';
defopts.StopIter = 'Inf % stop after resp. iteration, possibly resume later';
defopts.TolX = '1e-11*max(insigma) % stop if x-change smaller TolX';
defopts.TolUpX = '1e3*max(insigma) % stop if x-changes larger TolUpX';
defopts.TolFun = '1e-12 % stop if fun-changes smaller TolFun';
defopts.TolHistFun = '1e-13 % stop if back fun-changes smaller TolHistFun';
defopts.StopOnStagnation = 'on % stop when fitness stagnates for a long time';
defopts.StopOnWarnings = 'yes % ''no''==''off''==0, ''on''==''yes''==1 ';
defopts.StopOnEqualFunctionValues = '2 + N/3 % number of iterations';
% Options defaults: Other
defopts.DiffMaxChange = 'Inf % maximal variable change(s), can be Nx1-vector';
defopts.DiffMinChange = '0 % minimal variable change(s), can be Nx1-vector';
defopts.WarnOnEqualFunctionValues = ...
'yes % ''no''==''off''==0, ''on''==''yes''==1 ';
defopts.LBounds = '-Inf % lower bounds, scalar or Nx1-vector';
defopts.UBounds = 'Inf % upper bounds, scalar or Nx1-vector';
defopts.EvalParallel = 'no % objective function FUN accepts NxM matrix, with M>1?';
defopts.EvalInitialX = 'yes % evaluation of initial solution';defopts.Restarts = '0 % number of restarts ';
defopts.IncPopSize = '2 % multiplier for population size before each restart';
defopts.PopSize = '(4 + floor(3*log(N))) % population size, lambda';
defopts.ParentNumber = 'floor(popsize/2) % AKA mu, popsize equals lambda';
defopts.RecombinationWeights = 'superlinear decrease % or linear, or equal';
defopts.DiagonalOnly = '0*(1+100*N/sqrt(popsize))+(N>=1000) % C is diagonal for given iterations, 1==always';
defopts.Noise.on = '0 % uncertainty handling is off by default';
defopts.Noise.reevals = '1*ceil(0.05*lambda) % nb. of re-evaluated for uncertainty measurement';
defopts.Noise.theta = '0.5 % threshold to invoke uncertainty treatment'; % smaller: more likely to diverge
defopts.Noise.cum = '0.3 % cumulation constant for uncertainty';
defopts.Noise.cutoff = '2*lambda/3 % rank change cutoff for summation';
defopts.Noise.alphasigma = '1+2/(N+10) % factor for increasing sigma'; % smaller: slower adaptation
defopts.Noise.epsilon = '1e-7 % additional relative perturbation before reevaluation';
defopts.Noise.minmaxevals = '[1 inf] % min and max value of 2nd arg to fitfun, start value is 5th arg to cmaes';
defopts.Noise.alphaevals = '1+2/(N+10) % factor for increasing 2nd arg to fitfun';
defopts.Noise.callback = '[] % callback function when uncertainty threshold is exceeded';
% defopts.TPA = 0;
defopts.CMA.cs = '(mueff+2)/(N+mueff+3) % cumulation constant for step-size';
%qqq cs = (mueff^0.5)/(N^0.5+mueff^0.5) % the short time horizon version
defopts.CMA.damps = '1 + 2*max(0,sqrt((mueff-1)/(N+1))-1) + cs % damping for step-size';
% defopts.CMA.ccum = '4/(N+4) % cumulation constant for covariance matrix';
defopts.CMA.ccum = '(4 + mueff/N) / (N+4 + 2*mueff/N) % cumulation constant for pc';
defopts.CMA.ccov1 = '2 / ((N+1.3)^2+mueff) % learning rate for rank-one update';
defopts.CMA.ccovmu = '2 * (mueff-2+1/mueff) / ((N+2)^2+mueff) % learning rate for rank-mu update';
defopts.CMA.on = 'yes';
defopts.CMA.active = '0 % active CMA 1: neg. updates with pos. def. check, 2: neg. updates';
flg_future_setting = 0; % testing for possible future variant(s)
if flg_future_setting
disp('in the future')
% damps setting from Brockhoff et al 2010
% this damps diverges with popsize 400:
% cmaeshtml('benchmarkszero', ones(20,1)*2, 5, o, 15);
defopts.CMA.damps = '2*mueff/lambda + 0.3 + cs % damping for step-size'; % cs: for large mueff
% how about:
% defopts.CMA.damps = '2*mueff/lambda + 0.3 + 2*max(0,sqrt((mueff-1)/(N+1))-1) + cs % damping for step-size';
% ccum adjusted for large mueff, better on schefelmult?
% TODO: this should also depend on diagonal option!?
defopts.CMA.ccum = '(4 + mueff/N) / (N+4 + 2*mueff/N) % cumulation constant for pc';
defopts.CMA.active = '1 % active CMA 1: neg. updates with pos. def. check, 2: neg. updates';
end
defopts.Resume = 'no % resume former run from SaveFile';
defopts.Science = 'on % off==do some additional (minor) problem capturing, NOT IN USE';
defopts.ReadSignals = 'on % from file signals.par for termination, yet a stumb';
defopts.Seed = 'sum(100*clock) % evaluated if it is a string';
defopts.DispFinal = 'on % display messages like initial and final message';
defopts.DispModulo = '100 % [0:Inf], disp messages after every i-th iteration';
defopts.SaveVariables = 'on % [on|final|off][-v6] save variables to .mat file';
defopts.SaveFilename = 'variablescmaes.mat % save all variables, see SaveVariables';
defopts.LogModulo = '1 % [0:Inf] if >1 record data less frequently after gen=100';
defopts.LogTime = '25 % [0:100] max. percentage of time for recording data';
defopts.LogFilenamePrefix = 'outcmaes % files for output data';
defopts.LogPlot = 'off % plot while running using output data files';
%qqqkkk
%defopts.varopt1 = ''; % 'for temporary and hacking purposes';
%defopts.varopt2 = ''; % 'for temporary and hacking purposes';
defopts.UserData = 'for saving data/comments associated with the run';
defopts.UserDat2 = ''; 'for saving data/comments associated with the run';
% ---------------------- Handling Input Parameters ----------------------
if nargin < 1 || isequal(fitfun, 'defaults') % pass default options
if nargin < 1
disp('Default options returned (type "help cmaes" for help).'); end
xmin = defopts;
if nargin > 1 % supplement second argument with default options
xmin = getoptions(xstart, defopts);
end
return;
end
if isequal(fitfun, 'displayoptions')
names = fieldnames(defopts);
for name = names'
disp([name{:} repmat(' ', 1, 20-length(name{:})) ': ''' defopts.(name{:}) '''']);
end
return;
end
input.fitfun = fitfun; % record used input
if isempty(fitfun)
% fitfun = definput.fitfun;
% warning(['Objective function not determined, ''' fitfun ''' used']);
error(['Objective function not determined']);
end
if ~ischar(fitfun)
error('first argument FUN must be a string');
end
if nargin < 2
xstart = [];
end
input.xstart = xstart;
if isempty(xstart)
% xstart = definput.xstart; % objective variables initial point
% warning('Initial search point, and problem dimension, not determined');
error('Initial search point, and problem dimension, not determined');
end
if nargin < 3
insigma = [];
end
if isa(insigma, 'struct')
error(['Third argument SIGMA must be (or eval to) a scalar '...
'or a column vector of size(X0,1)']);
end
input.sigma = insigma;
if isempty(insigma)
if all(size(myeval(xstart)) > 1)
insigma = std(xstart, 0, 2);
if any(insigma == 0)
error(['Initial search volume is zero, choose SIGMA or X0 appropriate']);
end
else
% will be captured later
% error(['Initial step sizes (SIGMA) not determined']);
end
end
% Compose options opts
if nargin < 4 || isempty(inopts) % no input options available
inopts = [];
opts = defopts;
else
opts = getoptions(inopts, defopts);
end
i = strfind(opts.SaveFilename, ' '); % remove everything after white space
if ~isempty(i)
opts.SaveFilename = opts.SaveFilename(1:i(1)-1);
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
counteval = 0; countevalNaN = 0;
irun = 0;
while irun <= myeval(opts.Restarts) % for-loop does not work with resume
irun = irun + 1;
% ------------------------ Initialization -------------------------------
% Handle resuming of old run
flgresume = myevalbool(opts.Resume);
xmean = myeval(xstart);
if all(size(xmean) > 1)
xmean = mean(xmean, 2); % in case if xstart is a population
elseif size(xmean, 2) > 1
xmean = xmean';
end
if ~flgresume % not resuming a former run
% Assign settings from input parameters and options for myeval...
N = size(xmean, 1); numberofvariables = N;
lambda0 = floor(myeval(opts.PopSize) * myeval(opts.IncPopSize)^(irun-1));
% lambda0 = floor(myeval(opts.PopSize) * 3^floor((irun-1)/2));
popsize = lambda0;
lambda = lambda0;
insigma = myeval(insigma);
if all(size(insigma) == [N 2])
insigma = 0.5 * (insigma(:,2) - insigma(:,1));
end
else % flgresume is true, do resume former run
tmp = whos('-file', opts.SaveFilename);
for i = 1:length(tmp)
if strcmp(tmp(i).name, 'localopts');
error('Saved variables include variable "localopts", please remove');
end
end
local.opts = opts; % keep stopping and display options
local.varargin = varargin;
load(opts.SaveFilename);
varargin = local.varargin;
flgresume = 1;
% Overwrite old stopping and display options
opts.StopFitness = local.opts.StopFitness;
%%opts.MaxFunEvals = local.opts.MaxFunEvals;
%%opts.MaxIter = local.opts.MaxIter;
opts.StopFunEvals = local.opts.StopFunEvals;
opts.StopIter = local.opts.StopIter;
opts.TolX = local.opts.TolX;
opts.TolUpX = local.opts.TolUpX;
opts.TolFun = local.opts.TolFun;
opts.TolHistFun = local.opts.TolHistFun;
opts.StopOnStagnation = local.opts.StopOnStagnation;
opts.StopOnWarnings = local.opts.StopOnWarnings;
opts.ReadSignals = local.opts.ReadSignals;
opts.DispFinal = local.opts.DispFinal;
opts.LogPlot = local.opts.LogPlot;
opts.DispModulo = local.opts.DispModulo;
opts.SaveVariables = local.opts.SaveVariables;
opts.LogModulo = local.opts.LogModulo;
opts.LogTime = local.opts.LogTime;
clear local; % otherwise local would be overwritten during load
end
%--------------------------------------------------------------
% Evaluate options
stopFitness = myeval(opts.StopFitness);
stopMaxFunEvals = myeval(opts.MaxFunEvals);
stopMaxIter = myeval(opts.MaxIter);
stopFunEvals = myeval(opts.StopFunEvals);
stopIter = myeval(opts.StopIter);
stopTolX = myeval(opts.TolX);stopTolUpX = myeval(opts.TolUpX);
stopTolFun = myeval(opts.TolFun);
stopTolHistFun = myeval(opts.TolHistFun);
stopOnStagnation = myevalbool(opts.StopOnStagnation);
stopOnWarnings = myevalbool(opts.StopOnWarnings);
flgreadsignals = myevalbool(opts.ReadSignals);
flgWarnOnEqualFunctionValues = myevalbool(opts.WarnOnEqualFunctionValues);
flgEvalParallel = myevalbool(opts.EvalParallel);
stopOnEqualFunctionValues = myeval(opts.StopOnEqualFunctionValues);
arrEqualFunvals = zeros(1,10+N);
flgDiagonalOnly = myeval(opts.DiagonalOnly);
flgActiveCMA = myeval(opts.CMA.active);
noiseHandling = myevalbool(opts.Noise.on);
noiseMinMaxEvals = myeval(opts.Noise.minmaxevals);
noiseAlphaEvals = myeval(opts.Noise.alphaevals);
noiseCallback = myeval(opts.Noise.callback);
flgdisplay = myevalbool(opts.DispFinal);
flgplotting = myevalbool(opts.LogPlot);
verbosemodulo = myeval(opts.DispModulo);
flgscience = myevalbool(opts.Science);
flgsaving = [];
strsaving = [];
if strfind(opts.SaveVariables, '-v6')
i = strfind(opts.SaveVariables, '%');
if isempty(i) || i == 0 || strfind(opts.SaveVariables, '-v6') < i
strsaving = '-v6';
flgsaving = 1;
flgsavingfinal = 1;
end
end
if strncmp('final', opts.SaveVariables, 5)
flgsaving = 0;
flgsavingfinal = 1;
end
if isempty(flgsaving)
flgsaving = myevalbool(opts.SaveVariables);
flgsavingfinal = flgsaving;
end
savemodulo = myeval(opts.LogModulo);
savetime = myeval(opts.LogTime);
i = strfind(opts.LogFilenamePrefix, ' '); % remove everything after white space
if ~isempty(i)
opts.LogFilenamePrefix = opts.LogFilenamePrefix(1:i(1)-1);
end
% TODO here silent option? set disp, save and log options to 0
%--------------------------------------------------------------
if (isfinite(stopFunEvals) || isfinite(stopIter)) && ~flgsaving
warning('To resume later the saving option needs to be set');
end
% Do more checking and initialization
if flgresume % resume is on
time.t0 = clock;
if flgdisplay
disp([' resumed from ' opts.SaveFilename ]);
end
if counteval >= stopMaxFunEvals
error(['MaxFunEvals exceeded, use StopFunEvals as stopping ' ...
'criterion before resume']);
end
if countiter >= stopMaxIter
error(['MaxIter exceeded, use StopIter as stopping criterion ' ...
'before resume']);
end
else % flgresume
% xmean = mean(myeval(xstart), 2); % evaluate xstart again, because of irun
maxdx = myeval(opts.DiffMaxChange); % maximal sensible variable change
mindx = myeval(opts.DiffMinChange); % minimal sensible variable change
% can both also be defined as Nx1 vectors
lbounds = myeval(opts.LBounds);
ubounds = myeval(opts.UBounds);
if length(lbounds) == 1
lbounds = repmat(lbounds, N, 1);
end
if length(ubounds) == 1
ubounds = repmat(ubounds, N, 1);
end
if isempty(insigma) % last chance to set insigma
if all(lbounds > -Inf) && all(ubounds < Inf)
if any(lbounds>=ubounds)
error('upper bound must be greater than lower bound');
end
insigma = 0.3*(ubounds-lbounds);
stopTolX = myeval(opts.TolX); % reevaluate these
stopTolUpX = myeval(opts.TolUpX);
else
error(['Initial step sizes (SIGMA) not determined']);
end
end
% Check all vector sizes
if size(xmean, 2) > 1 || size(xmean,1) ~= N
error(['intial search point should be a column vector of size ' ...
num2str(N)]);
elseif ~(all(size(insigma) == [1 1]) || all(size(insigma) == [N 1]))
error(['input parameter SIGMA should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(stopTolX, 2) > 1 || ~ismember(size(stopTolX, 1), [1 N])
error(['option TolX should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(stopTolUpX, 2) > 1 || ~ismember(size(stopTolUpX, 1), [1 N])
error(['option TolUpX should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(maxdx, 2) > 1 || ~ismember(size(maxdx, 1), [1 N])
error(['option DiffMaxChange should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(mindx, 2) > 1 || ~ismember(size(mindx, 1), [1 N])
error(['option DiffMinChange should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(lbounds, 2) > 1 || ~ismember(size(lbounds, 1), [1 N])
error(['option lbounds should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
elseif size(ubounds, 2) > 1 || ~ismember(size(ubounds, 1), [1 N])
error(['option ubounds should be (or eval to) a scalar '...
'or a column vector of size ' num2str(N)] );
end
% Initialize dynamic internal state parameters
if any(insigma <= 0)
error(['Initial search volume (SIGMA) must be greater than zero']);
end
if max(insigma)/min(insigma) > 1e6
error(['Initial search volume (SIGMA) badly conditioned']);
end
sigma = max(insigma); % overall standard deviation
pc = zeros(N,1); ps = zeros(N,1); % evolution paths for C and sigma
if length(insigma) == 1
insigma = insigma * ones(N,1) ;
end
diagD = insigma/max(insigma); % diagonal matrix D defines the scaling
diagC = diagD.^2;
if flgDiagonalOnly ~= 1 % use at some point full covariance matrix
B = eye(N,N); % B defines the coordinate system BD = B.*repmat(diagD',N,1); % B*D for speed up only
C = diag(diagC); % covariance matrix == BD*(BD)'
end
if flgDiagonalOnly
B = 1;
end
fitness.hist=NaN*ones(1,10+ceil(3*10*N/lambda)); % history of fitness values
fitness.histsel=NaN*ones(1,10+ceil(3*10*N/lambda)); % history of fitness values
fitness.histbest=[]; % history of fitness values
fitness.histmedian=[]; % history of fitness values
% Initialize boundary handling
bnd.isactive = any(lbounds > -Inf) || any(ubounds < Inf);
if bnd.isactive
if any(lbounds>ubounds)
error('lower bound found to be greater than upper bound');
end
[xmean ti] = xintobounds(xmean, lbounds, ubounds); % just in case
if any(ti)
warning('Initial point was out of bounds, corrected');
end
bnd.weights = zeros(N,1); % weights for bound penalty
% scaling is better in axis-parallel case, worse in rotated
bnd.flgscale = 0; % scaling will be omitted if zero
if bnd.flgscale ~= 0
bnd.scale = diagC/mean(diagC);
else
bnd.scale = ones(N,1);
end
idx = (lbounds > -Inf) | (ubounds < Inf);
if length(idx) == 1
idx = idx * ones(N,1);
end
bnd.isbounded = zeros(N,1);
bnd.isbounded(find(idx)) = 1;
maxdx = min(maxdx, (ubounds - lbounds)/2);
if any(sigma*sqrt(diagC) > maxdx)
fac = min(maxdx ./ sqrt(diagC))/sigma;
sigma = min(maxdx ./ sqrt(diagC));
warning(['Initial SIGMA multiplied by the factor ' num2str(fac) ...
', because it was larger than half' ...
' of one of the boundary intervals']);
end
idx = (lbounds > -Inf) & (ubounds < Inf);
dd = diagC;
if any(5*sigma*sqrt(dd(idx)) < ubounds(idx) - lbounds(idx))
warning(['Initial SIGMA is, in at least one coordinate, ' ...
'much smaller than the '...
'given boundary intervals. For reasonable ' ...
'global search performance SIGMA should be ' ...
'between 0.2 and 0.5 of the bounded interval in ' ...
'each coordinate. If all coordinates have ' ...
'lower and upper bounds SIGMA can be empty']);
end
bnd.dfithist = 1; % delta fit for setting weights
bnd.aridxpoints = []; % remember complete outside points
bnd.arfitness = []; % and their fitness
bnd.validfitval = 0;
bnd.iniphase = 1;
end
% ooo initial feval, for output only
if irun == 1
out.solutions.bestever.x = xmean;
out.solutions.bestever.f = Inf; % for simpler comparison below
out.solutions.bestever.evals = counteval;
bestever = out.solutions.bestever;
end if myevalbool(opts.EvalInitialX)
fitness.hist(1)=feval(fitfun, xmean, varargin{:});
fitness.histsel(1)=fitness.hist(1);
counteval = counteval + 1;
if fitness.hist(1) < out.solutions.bestever.f
out.solutions.bestever.x = xmean;
out.solutions.bestever.f = fitness.hist(1);
out.solutions.bestever.evals = counteval;
bestever = out.solutions.bestever;
end
else
fitness.hist(1)=NaN;
fitness.histsel(1)=NaN;
end
% initialize random number generator
if ischar(opts.Seed)
randn('state', eval(opts.Seed)); % random number generator state
else
randn('state', opts.Seed);
end
%qqq
% load(opts.SaveFilename, 'startseed');
% randn('state', startseed);
% disp(['SEED RELOADED FROM ' opts.SaveFilename]);
startseed = randn('state'); % for retrieving in saved variables
% Initialize further constants
chiN=N^0.5*(1-1/(4*N)+1/(21*N^2)); % expectation of
% ||N(0,I)|| == norm(randn(N,1))
countiter = 0;
% Initialize records and output
if irun == 1
time.t0 = clock;
% TODO: keep also median solution?
out.evals = counteval; % should be first entry
out.stopflag = {};
outiter = 0;
% Write headers to output data files
filenameprefix = opts.LogFilenamePrefix;
if savemodulo && savetime
filenames = {};
filenames(end+1) = {'axlen'};
filenames(end+1) = {'fit'};
filenames(end+1) = {'stddev'};
filenames(end+1) = {'xmean'};
filenames(end+1) = {'xrecentbest'};
str = [' (startseed=' num2str(startseed(2)) ...
', ' num2str(clock, '%d/%02d/%d %d:%d:%2.2f') ')'];
for namecell = filenames(:)'
name = namecell{:};
[fid, err] = fopen(['./' filenameprefix name '.dat'], 'w');
if fid < 1 % err ~= 0
warning(['could not open ' filenameprefix name '.dat']);
filenames(find(strcmp(filenames,name))) = [];
else
% fprintf(fid, '%s\n', ...
% ['<CMAES-OUTPUT version="' cmaVersion '">']);
% fprintf(fid, [' <NAME>' name '</NAME>\n']);
% fprintf(fid, [' <DATE>' date() '</DATE>\n']);
% fprintf(fid, ' <PARAMETERS>\n');
% fprintf(fid, [' dimension=' num2str(N) '\n']);
% fprintf(fid, ' </PARAMETERS>\n');
% different cases for DATA columns annotations here
% fprintf(fid, ' <DATA'); if strcmp(name, 'axlen')
fprintf(fid, ['%% columns="iteration, evaluation, sigma, ' ...
'max axis length, min axis length, ' ...
'all principal axes lengths (sorted square roots ' ...
'of eigenvalues of C)"' str]);
elseif strcmp(name, 'fit')
fprintf(fid, ['%% columns="iteration, evaluation, sigma, axis ratio, bestever,' ...
' best, median, worst fitness function value,' ...
' further objective values of best"' str]);
elseif strcmp(name, 'stddev')
fprintf(fid, ['%% columns=["iteration, evaluation, sigma, void, void, ' ...
'stds==sigma*sqrt(diag(C))"' str]);
elseif strcmp(name, 'xmean')
fprintf(fid, ['%% columns="iteration, evaluation, void, ' ...
'void, void, xmean"' str]);
elseif strcmp(name, 'xrecentbest')
fprintf(fid, ['%% columns="iteration, evaluation, fitness, ' ...
'void, void, xrecentbest"' str]);
end
fprintf(fid, '\n'); % DATA
if strcmp(name, 'xmean')
fprintf(fid, '%ld %ld 0 0 0 ', 0, counteval);
% fprintf(fid, '%ld %ld 0 0 %e ', countiter, counteval, fmean);
%qqq fprintf(fid, msprintf('%e ', genophenotransform(out.genopheno, xmean)) + '\n');
fprintf(fid, '%e ', xmean);
fprintf(fid, '\n');
end
fclose(fid);
clear fid; % preventing
end
end % for files
end % savemodulo
end % irun == 1
end % else flgresume
% -------------------- Generation Loop --------------------------------
stopflag = {};
while isempty(stopflag)
% set internal parameters
if countiter == 0 || lambda ~= lambda_last
if countiter > 0 && floor(log10(lambda)) ~= floor(log10(lambda_last)) ...
&& flgdisplay
disp([' lambda = ' num2str(lambda)]);
lambda_hist(:,end+1) = [countiter+1; lambda];
else
lambda_hist = [countiter+1; lambda];
end
lambda_last = lambda;
% Strategy internal parameter setting: Selection
mu = myeval(opts.ParentNumber); % number of parents/points for recombination
if strncmp(lower(opts.RecombinationWeights), 'equal', 3)
weights = ones(mu,1); % (mu_I,lambda)-CMA-ES
elseif strncmp(lower(opts.RecombinationWeights), 'linear', 3)
weights = mu+0.5-(1:mu)';
elseif strncmp(lower(opts.RecombinationWeights), 'superlinear', 3)
weights = log(mu+0.5)-log(1:mu)'; % muXone array for weighted recombination
% qqq mu can be non-integer and
% should become ceil(mu-0.5) (minor correction)
else
error(['Recombination weights to be "' opts.RecombinationWeights ...
'" is not implemented']);
end
mueff=sum(weights)^2/sum(weights.^2); % variance-effective size of mu
weights = weights/sum(weights); % normalize recombination weights array
if mueff == lambda
error(['Combination of values for PopSize, ParentNumber and ' ...
' and RecombinationWeights is not reasonable']);
end
% Strategy internal parameter setting: Adaptation
cc = myeval(opts.CMA.ccum); % time constant for cumulation for covariance matrix
cs = myeval(opts.CMA.cs);
% old way TODO: remove this at some point
% mucov = mueff; % size of mu used for calculating learning rate ccov
% ccov = (1/mucov) * 2/(N+1.41)^2 ... % learning rate for covariance matrix
% + (1-1/mucov) * min(1,(2*mucov-1)/((N+2)^2+mucov));
% new way
if myevalbool(opts.CMA.on)
ccov1 = myeval(opts.CMA.ccov1);
ccovmu = min(1-ccov1, myeval(opts.CMA.ccovmu));
else
ccov1 = 0;
ccovmu = 0;
end
% flgDiagonalOnly = -lambda*4*1/ccov; % for ccov==1 it is not needed
% 0 : C will never be diagonal anymore
% 1 : C will always be diagonal
% >1: C is diagonal for first iterations, set to 0 afterwards
if flgDiagonalOnly < 1
flgDiagonalOnly = 0;
end
if flgDiagonalOnly
ccov1_sep = min(1, ccov1 * (N+1.5) / 3);
ccovmu_sep = min(1-ccov1_sep, ccovmu * (N+1.5) / 3);
elseif N > 98 && flgdisplay && countiter == 0
disp('consider option DiagonalOnly for high-dimensional problems');
end
% ||ps|| is close to sqrt(mueff/N) for mueff large on linear fitness
%damps = ... % damping for step size control, usually close to one
% (1 + 2*max(0,sqrt((mueff-1)/(N+1))-1)) ... % limit sigma increase
% * max(0.3, ... % reduce damps, if max. iteration number is small
% 1 - N/min(stopMaxIter,stopMaxFunEvals/lambda)) + cs;
damps = myeval(opts.CMA.damps);
if noiseHandling
noiseReevals = min(myeval(opts.Noise.reevals), lambda);
noiseAlpha = myeval(opts.Noise.alphasigma);
noiseEpsilon = myeval(opts.Noise.epsilon);
noiseTheta = myeval(opts.Noise.theta);
noisecum = myeval(opts.Noise.cum);
noiseCutOff = myeval(opts.Noise.cutoff); % arguably of minor relevance
else
noiseReevals = 0; % more convenient in later coding
end
%qqq hacking of a different parameter setting, e.g. for ccov or damps,
% can be done here, but is not necessary anymore, see opts.CMA.
% ccov1 = 0.0*ccov1; disp(['CAVE: ccov1=' num2str(ccov1)]);
% ccovmu = 0.0*ccovmu; disp(['CAVE: ccovmu=' num2str(ccovmu)]);
% damps = inf*damps; disp(['CAVE: damps=' num2str(damps)]);
% cc = 1; disp(['CAVE: cc=' num2str(cc)]);
end
% Display initial message
if countiter == 0 && flgdisplay
if mu == 1
strw = '100';
elseif mu < 8
strw = [sprintf('%.0f', 100*weights(1)) ...
sprintf(' %.0f', 100*weights(2:end)')];
else
strw = [sprintf('%.2g ', 100*weights(1:2)') ...
sprintf('%.2g', 100*weights(3)') '...' ...
sprintf(' %.2g', 100*weights(end-1:end)') ']%, '];
end if irun > 1
strrun = [', run ' num2str(irun)];
else
strrun = '';
end
disp([' n=' num2str(N) ': (' num2str(mu) ',' ...
num2str(lambda) ')-CMA-ES(w=[' ...
strw ']%, ' ...
'mu_eff=' num2str(mueff,'%.1f') ...
') on function ' ...
(fitfun) strrun]);
if flgDiagonalOnly == 1
disp(' C is diagonal');
elseif flgDiagonalOnly
disp([' C is diagonal for ' num2str(floor(flgDiagonalOnly)) ' iterations']);
end
end
flush;
countiter = countiter + 1;
% Generate and evaluate lambda offspring
fitness.raw = repmat(NaN, 1, lambda + noiseReevals);
% parallel evaluation
if flgEvalParallel
arz = randn(N,lambda);
if ~flgDiagonalOnly
arx = repmat(xmean, 1, lambda) + sigma * (BD * arz); % Eq. (1)
else
arx = repmat(xmean, 1, lambda) + repmat(sigma * diagD, 1, lambda) .* arz;
end
if noiseHandling
if noiseEpsilon == 0
arx = [arx arx(:,1:noiseReevals)];
elseif flgDiagonalOnly
arx = [arx arx(:,1:noiseReevals) + ...
repmat(noiseEpsilon * sigma * diagD, 1, noiseReevals) ...
.* randn(N,noiseReevals)];
else
arx = [arx arx(:,1:noiseReevals) + ...
noiseEpsilon * sigma * ...
(BD * randn(N,noiseReevals))];
end
end
% You may handle constraints here. You may either resample
% arz(:,k) and/or multiply it with a factor between -1 and 1
% (the latter will decrease the overall step size) and
% recalculate arx accordingly. Do not change arx or arz in any
% other way.
if ~bnd.isactive
arxvalid = arx;
else
arxvalid = xintobounds(arx, lbounds, ubounds);
end
% You may handle constraints here. You may copy and alter
% (columns of) arxvalid(:,k) only for the evaluation of the
% fitness function. arx and arxvalid should not be changed.
fitness.raw = feval(fitfun, arxvalid, varargin{:});
countevalNaN = countevalNaN + sum(isnan(fitness.raw));
counteval = counteval + sum(~isnan(fitness.raw));
end
% non-parallel evaluation and remaining NaN-values % set also the reevaluated solution to NaN
fitness.raw(lambda + find(isnan(fitness.raw(1:noiseReevals)))) = NaN;
for k=find(isnan(fitness.raw)),
% fitness.raw(k) = NaN;
tries = 0;
% Resample, until fitness is not NaN
while isnan(fitness.raw(k))
if k <= lambda % regular samples (not the re-evaluation-samples)
arz(:,k) = randn(N,1); % (re)sample
if flgDiagonalOnly
arx(:,k) = xmean + sigma * diagD .* arz(:,k); % Eq. (1)
else
arx(:,k) = xmean + sigma * (BD * arz(:,k)); % Eq. (1)
end
else % re-evaluation solution with index > lambda
if flgDiagonalOnly
arx(:,k) = arx(:,k-lambda) + (noiseEpsilon * sigma) * diagD .* randn(N,1);
else
arx(:,k) = arx(:,k-lambda) + (noiseEpsilon * sigma) * (BD * randn(N,1));
end
end
% You may handle constraints here. You may either resample
% arz(:,k) and/or multiply it with a factor between -1 and 1
% (the latter will decrease the overall step size) and
% recalculate arx accordingly. Do not change arx or arz in any
% other way.
if ~bnd.isactive
arxvalid(:,k) = arx(:,k);
else
arxvalid(:,k) = xintobounds(arx(:,k), lbounds, ubounds);
end
% You may handle constraints here. You may copy and alter
% (columns of) arxvalid(:,k) only for the evaluation of the
% fitness function. arx should not be changed.
fitness.raw(k) = feval(fitfun, arxvalid(:,k), varargin{:});
tries = tries + 1;
if isnan(fitness.raw(k))
countevalNaN = countevalNaN + 1;
end
if mod(tries, 100) == 0
warning([num2str(tries) ...
' NaN objective function values at evaluation ' ...
num2str(counteval)]);
end
end
counteval = counteval + 1; % retries due to NaN are not counted
end
fitness.sel = fitness.raw;
% ----- handle boundaries -----
if 1 < 3 && bnd.isactive
% Get delta fitness values
val = myprctile(fitness.raw, [25 75]);
% more precise would be exp(mean(log(diagC)))
val = (val(2) - val(1)) / N / mean(diagC) / sigma^2;
%val = (myprctile(fitness.raw, 75) - myprctile(fitness.raw, 25)) ...
% / N / mean(diagC) / sigma^2;
% Catch non-sensible values
if ~isfinite(val)
warning('Non-finite fitness range');
val = max(bnd.dfithist);
elseif val == 0 % happens if all points are out of bounds
val = min(bnd.dfithist(bnd.dfithist>0)); % seems not to make sense, given all solutions are out of bounds
elseif bnd.validfitval == 0 % flag that first sensible val was found
bnd.dfithist = [];
bnd.validfitval = 1; end
% Store delta fitness values
if length(bnd.dfithist) < 20+(3*N)/lambda
bnd.dfithist = [bnd.dfithist val];
else
bnd.dfithist = [bnd.dfithist(2:end) val];
end
[tx ti] = xintobounds(xmean, lbounds, ubounds);
% Set initial weights
if bnd.iniphase
if any(ti)
bnd.weights(find(bnd.isbounded)) = 2.0002 * median(bnd.dfithist);
if bnd.flgscale == 0 % scale only initial weights then
dd = diagC;
idx = find(bnd.isbounded);
dd = dd(idx) / mean(dd); % remove mean scaling
bnd.weights(idx) = bnd.weights(idx) ./ dd;
end
if bnd.validfitval && countiter > 2
bnd.iniphase = 0;
end
end
end
% Increase weights
if 1 < 3 && any(ti) % any coordinate of xmean out of bounds
% judge distance of xmean to boundary
tx = xmean - tx;
idx = (ti ~= 0 & abs(tx) > 3*max(1,sqrt(N)/mueff) ...
* sigma*sqrt(diagC)) ;
% only increase if xmean is moving away
idx = idx & (sign(tx) == sign(xmean - xold));
if ~isempty(idx) % increase
% the factor became 1.2 instead of 1.1, because
% changed from max to min in version 3.52
bnd.weights(idx) = 1.2^(min(1, mueff/10/N)) * bnd.weights(idx);
end
end
% Calculate scaling biased to unity, product is one
if bnd.flgscale ~= 0
bnd.scale = exp(0.9*(log(diagC)-mean(log(diagC))));
end
% Assigned penalized fitness
bnd.arpenalty = (bnd.weights ./ bnd.scale)' * (arxvalid - arx).^2;
fitness.sel = fitness.raw + bnd.arpenalty;
end % handle boundaries
% ----- end handle boundaries -----
% compute noise measurement and reduce fitness arrays to size lambda
if noiseHandling
[noiseS] = local_noisemeasurement(fitness.sel(1:lambda), ...
fitness.sel(lambda+(1:noiseReevals)), ...
noiseReevals, noiseTheta, noiseCutOff);
if countiter == 1 % TODO: improve this very rude way of initialization
noiseSS = 0;
noiseN = 0; % counter for mean
end
noiseSS = noiseSS + noisecum * (noiseS - noiseSS);
% noise-handling could be done here, but the original sigma is still needed
% disp([noiseS noiseSS noisecum])
fitness.rawar12 = fitness.raw; % just documentary fitness.selar12 = fitness.sel; % just documentary
% qqq refine fitness based on both values
if 11 < 3 % TODO: in case of outliers this mean is counterproductive
% median out of three would be ok
fitness.raw(1:noiseReevals) = ... % not so raw anymore
(fitness.raw(1:noiseReevals) + fitness.raw(lambda+(1:noiseReevals))) / 2;
fitness.sel(1:noiseReevals) = ...
(fitness.sel(1:noiseReevals) + fitness.sel(lambda+(1:noiseReevals))) / 2;
end
fitness.raw = fitness.raw(1:lambda);
fitness.sel = fitness.sel(1:lambda);
end
% Sort by fitness
[fitness.raw, fitness.idx] = sort(fitness.raw);
[fitness.sel, fitness.idxsel] = sort(fitness.sel); % minimization
fitness.hist(2:end) = fitness.hist(1:end-1); % record short history of
fitness.hist(1) = fitness.raw(1); % best fitness values
if length(fitness.histbest) < 120+ceil(30*N/lambda) || ...
(mod(countiter, 5) == 0 && length(fitness.histbest) < 2e4) % 20 percent of 1e5 gen.
fitness.histbest = [fitness.raw(1) fitness.histbest]; % best fitness values
fitness.histmedian = [median(fitness.raw) fitness.histmedian]; % median fitness values
else
fitness.histbest(2:end) = fitness.histbest(1:end-1);
fitness.histmedian(2:end) = fitness.histmedian(1:end-1);
fitness.histbest(1) = fitness.raw(1); % best fitness values
fitness.histmedian(1) = median(fitness.raw); % median fitness values
end
fitness.histsel(2:end) = fitness.histsel(1:end-1); % record short history of
fitness.histsel(1) = fitness.sel(1); % best sel fitness values
% Calculate new xmean, this is selection and recombination
xold = xmean; % for speed up of Eq. (2) and (3)
xmean = arx(:,fitness.idxsel(1:mu))*weights;
zmean = arz(:,fitness.idxsel(1:mu))*weights;%==D^-1*B'*(xmean-xold)/sigma
if mu == 1
fmean = fitness.sel(1);
else
fmean = NaN; % [] does not work in the latter assignment
% fmean = feval(fitfun, xintobounds(xmean, lbounds, ubounds), varargin{:});
% counteval = counteval + 1;
end
% Cumulation: update evolution paths
ps = (1-cs)*ps + sqrt(cs*(2-cs)*mueff) * (B*zmean); % Eq. (4)
hsig = norm(ps)/sqrt(1-(1-cs)^(2*countiter))/chiN < 1.4 + 2/(N+1);
if flg_future_setting
hsig = sum(ps.^2) / (1-(1-cs)^(2*countiter)) / N < 2 + 4/(N+1); % just simplified
end
% hsig = norm(ps)/sqrt(1-(1-cs)^(2*countiter))/chiN < 1.4 + 2/(N+1);
% hsig = norm(ps)/sqrt(1-(1-cs)^(2*countiter))/chiN < 1.5 + 1/(N-0.5);
% hsig = norm(ps) < 1.5 * sqrt(N);
% hsig = 1;
pc = (1-cc)*pc ...
+ hsig*(sqrt(cc*(2-cc)*mueff)/sigma) * (xmean-xold); % Eq. (2)
if hsig == 0
% disp([num2str(countiter) ' ' num2str(counteval) ' pc update stalled']);
end
% Adapt covariance matrix
neg.ccov = 0; % TODO: move parameter setting upwards at some point
if ccov1 + ccovmu > 0 % Eq. (3)
if flgDiagonalOnly % internal linear(?) complexity
diagC = (1-ccov1_sep-ccovmu_sep+(1-hsig)*ccov1_sep*cc*(2-cc)) * diagC ... % regard old matrix
+ ccov1_sep * pc.^2 ... % plus rank one update
+ ccovmu_sep ... % plus rank mu update
* (diagC .* (arz(:,fitness.idxsel(1:mu)).^2 * weights));
% * (repmat(diagC,1,mu) .* arz(:,fitness.idxsel(1:mu)).^2 * weights);
diagD = sqrt(diagC); % replaces eig(C) else
arpos = (arx(:,fitness.idxsel(1:mu))-repmat(xold,1,mu)) / sigma;
% "active" CMA update: negative update, in case controlling pos. definiteness
if flgActiveCMA > 0
% set parameters
neg.mu = mu;
neg.mueff = mueff;
if flgActiveCMA > 10 % flat weights with mu=lambda/2
neg.mu = floor(lambda/2);
neg.mueff = neg.mu;
end
% neg.mu = ceil(min([N, lambda/4, mueff])); neg.mueff = mu; % i.e. neg.mu <= N
% Parameter study: in 3-D lambda=50,100, 10-D lambda=200,400, 30-D lambda=1000,2000 a
% three times larger neg.ccov does not work.
% increasing all ccov rates three times does work (probably because of the factor (1-ccovmu))
% in 30-D to looks fine
neg.ccov = (1 - ccovmu) * 0.25 * neg.mueff / ((N+2)^1.5 + 2*neg.mueff);
neg.minresidualvariance = 0.66; % keep at least 0.66 in all directions, small popsize are most critical
neg.alphaold = 0.5; % where to make up for the variance loss, 0.5 means no idea what to do
% 1 is slightly more robust and gives a better "guaranty" for pos. def.,
% but does it make sense from the learning perspective for large ccovmu?
neg.ccovfinal = neg.ccov;
% prepare vectors, compute negative updating matrix Cneg and checking matrix Ccheck
arzneg = arz(:,fitness.idxsel(lambda:-1:lambda - neg.mu + 1));
% i-th longest becomes i-th shortest
% TODO: this is not in compliance with the paper Hansen&Ros2010,
% where simply arnorms = arnorms(end:-1:1) ?
[arnorms idxnorms] = sort(sqrt(sum(arzneg.^2, 1)));
[ignore idxnorms] = sort(idxnorms); % inverse index
arnormfacs = arnorms(end:-1:1) ./ arnorms;
% arnormfacs = arnorms(randperm(neg.mu)) ./ arnorms;
arnorms = arnorms(end:-1:1); % for the record
if flgActiveCMA < 20
arzneg = arzneg .* repmat(arnormfacs(idxnorms), N, 1); % E x*x' is N
% arzneg = sqrt(N) * arzneg ./ repmat(sqrt(sum(arzneg.^2, 1)), N, 1); % E x*x' is N
end
if flgActiveCMA < 10 && neg.mu == mu % weighted sum
if mod(flgActiveCMA, 10) == 1 % TODO: prevent this with a less tight but more efficient check (see below)
Ccheck = arzneg * diag(weights) * arzneg'; % in order to check the largest EV
end
artmp = BD * arzneg;
Cneg = artmp * diag(weights) * artmp';
else % simple sum
if mod(flgActiveCMA, 10) == 1
Ccheck = (1/neg.mu) * arzneg*arzneg'; % in order to check largest EV
end
artmp = BD * arzneg;
Cneg = (1/neg.mu) * artmp*artmp';
end
% check pos.def. and set learning rate neg.ccov accordingly,
% this check makes the original choice of neg.ccov extremly failsafe
% still assuming C == BD*BD', which is only approxim. correct
if mod(flgActiveCMA, 10) == 1 && 1 - neg.ccov * arnorms(idxnorms).^2 * weights < neg.minresidualvariance
% TODO: the simple and cheap way would be to set
% fac = 1 - ccovmu - ccov1 OR 1 - mueff/lambda and
% neg.ccov = fac*(1 - neg.minresidualvariance) / (arnorms(idxnorms).^2 * weights)
% this is the more sophisticated way:
% maxeigenval = eigs(arzneg * arzneg', 1, 'lm', eigsopts); % not faster
maxeigenval = max(eig(Ccheck)); % norm is much slower, because (norm()==max(svd())
%disp([countiter log10([neg.ccov, maxeigenval, arnorms(idxnorms).^2 * weights, max(arnorms)^2]), ...
% neg.ccov * arnorms(idxnorms).^2 * weights])
% pause
% remove less than ??34*(1-cmu)%?? of variance in any direction
% 1-ccovmu is the variance left from the old C
neg.ccovfinal = min(neg.ccov, (1-ccovmu)*(1-neg.minresidualvariance)/maxeigenval); % -ccov1 removed to avoid error message??
if neg.ccovfinal < neg.ccov
disp(['active CMA at iteration ' num2str(countiter) ...
': max EV ==', num2str([maxeigenval, neg.ccov, neg.ccovfinal])]);
end
end
% xmean = xold; % the distribution does not degenerate!?
% update C
C = (1-ccov1-ccovmu+neg.alphaold*neg.ccovfinal+(1-hsig)*ccov1*cc*(2-cc)) * C ... % regard old matrix
+ ccov1 * pc*pc' ... % plus rank one update
+ (ccovmu + (1-neg.alphaold)*neg.ccovfinal) ... % plus rank mu update
* arpos * (repmat(weights,1,N) .* arpos') ...
- neg.ccovfinal * Cneg; % minus rank mu update
else % no active (negative) update
C = (1-ccov1-ccovmu+(1-hsig)*ccov1*cc*(2-cc)) * C ... % regard old matrix
+ ccov1 * pc*pc' ... % plus rank one update
+ ccovmu ... % plus rank mu update
* arpos * (repmat(weights,1,N) .* arpos');
% is now O(mu*N^2 + mu*N), was O(mu*N^2 + mu^2*N) when using diag(weights)
% for mu=30*N it is now 10 times faster, overall 3 times faster
end
diagC = diag(C);
end
end
% the following is de-preciated and will be removed in future
% better setting for cc makes this hack obsolete
if 11 < 2 && ~flgscience
% remove momentum in ps, if ps is large and fitness is getting worse.
% this should rarely happen.
% this might very well be counterproductive in dynamic environments
if sum(ps.^2)/N > 1.5 + 10*(2/N)^.5 && ...
fitness.histsel(1) > max(fitness.histsel(2:3))
ps = ps * sqrt(N*(1+max(0,log(sum(ps.^2)/N))) / sum(ps.^2));
if flgdisplay
disp(['Momentum in ps removed at [niter neval]=' ...
num2str([countiter counteval]) ']']);
end
end
end
% Adapt sigma
if flg_future_setting % according to a suggestion from Dirk Arnold (2000)
% exp(1) is still not reasonably small enough
sigma = sigma * exp(min(1, (sum(ps.^2)/N - 1)/2 * cs/damps)); % Eq. (5)
else
% exp(1) is still not reasonably small enough
sigma = sigma * exp(min(1, (sqrt(sum(ps.^2))/chiN - 1) * cs/damps)); % Eq. (5)
end
% disp([countiter norm(ps)/chiN]);
if 11 < 3 % testing with optimal step-size
if countiter == 1
disp('*********** sigma set to const * ||x|| ******************');
end
sigma = 0.04 * mueff * sqrt(sum(xmean.^2)) / N; % 20D,lam=1000:25e3
sigma = 0.3 * mueff * sqrt(sum(xmean.^2)) / N; % 20D,lam=(40,1000):17e3
% 75e3 with def (1.5)
% 35e3 with damps=0.25
end
if 11 < 3
if countiter == 1
disp('*********** xmean set to const ******************');
end
xmean = ones(N,1);
end
% Update B and D from C
if ~flgDiagonalOnly && (ccov1+ccovmu+neg.ccov) > 0 && mod(countiter, 1/(ccov1+ccovmu+neg.ccov)/N/10) < 1 C=triu(C)+triu(C,1)'; % enforce symmetry to prevent complex numbers
[B,tmp] = eig(C); % eigen decomposition, B==normalized eigenvectors
% effort: approx. 15*N matrix-vector multiplications
diagD = diag(tmp);
if any(~isfinite(diagD))
clear idx; % prevents error under octave
save(['tmp' opts.SaveFilename]);
error(['function eig returned non-finited eigenvalues, cond(C)=' ...
num2str(cond(C)) ]);
end
if any(any(~isfinite(B)))
clear idx; % prevents error under octave
save(['tmp' opts.SaveFilename]);
error(['function eig returned non-finited eigenvectors, cond(C)=' ...
num2str(cond(C)) ]);
end
% limit condition of C to 1e14 + 1
if min(diagD) <= 0
if stopOnWarnings
stopflag(end+1) = {'warnconditioncov'};
else
warning(['Iteration ' num2str(countiter) ...
': Eigenvalue (smaller) zero']);
diagD(diagD<0) = 0;
tmp = max(diagD)/1e14;
C = C + tmp*eye(N,N); diagD = diagD + tmp*ones(N,1);
end
end
if max(diagD) > 1e14*min(diagD)
if stopOnWarnings
stopflag(end+1) = {'warnconditioncov'};
else
warning(['Iteration ' num2str(countiter) ': condition of C ' ...
'at upper limit' ]);
tmp = max(diagD)/1e14 - min(diagD);
C = C + tmp*eye(N,N); diagD = diagD + tmp*ones(N,1);
end
end
diagC = diag(C);
diagD = sqrt(diagD); % D contains standard deviations now
% diagD = diagD / prod(diagD)^(1/N); C = C / prod(diagD)^(2/N);
BD = B.*repmat(diagD',N,1); % O(n^2)
end % if mod
% Align/rescale order of magnitude of scales of sigma and C for nicer output
% not a very usual case
if 1 < 2 && sigma > 1e10*max(diagD)
fac = sigma / max(diagD);
sigma = sigma/fac;
pc = fac * pc;
diagD = fac * diagD;
if ~flgDiagonalOnly
C = fac^2 * C; % disp(fac);
BD = B.*repmat(diagD',N,1); % O(n^2), but repmat might be inefficient todo?
end
diagC = fac^2 * diagC;
end
if flgDiagonalOnly > 1 && countiter > flgDiagonalOnly
% full covariance matrix from now on
flgDiagonalOnly = 0;
B = eye(N,N);
BD = diag(diagD);
C = diag(diagC); % is better, because correlations are spurious anyway
end
if noiseHandling if countiter == 1 % assign firstvarargin for noise treatment e.g. as #reevaluations
if ~isempty(varargin) && length(varargin{1}) == 1 && isnumeric(varargin{1})
if irun == 1
firstvarargin = varargin{1};
else
varargin{1} = firstvarargin; % reset varargin{1}
end
else
firstvarargin = 0;
end
end
if noiseSS < 0 && noiseMinMaxEvals(2) > noiseMinMaxEvals(1) && firstvarargin
varargin{1} = max(noiseMinMaxEvals(1), varargin{1} / noiseAlphaEvals^(1/4)); % still experimental
elseif noiseSS > 0
if ~isempty(noiseCallback) % to be removed?
res = feval(noiseCallback); % should also work without output argument!?
if ~isempty(res) && res > 1 % TODO: decide for interface of callback
% also a dynamic popsize could be done here
sigma = sigma * noiseAlpha;
end
else
if noiseMinMaxEvals(2) > noiseMinMaxEvals(1) && firstvarargin
varargin{1} = min(noiseMinMaxEvals(2), varargin{1} * noiseAlphaEvals);
end
sigma = sigma * noiseAlpha;
% lambda = ceil(0.1 * sqrt(lambda) + lambda);
% TODO: find smallest increase of lambda with log-linear
% convergence in iterations
end
% qqq experimental: take a mean to estimate the true optimum
noiseN = noiseN + 1;
if noiseN == 1
noiseX = xmean;
else
noiseX = noiseX + (3/noiseN) * (xmean - noiseX);
end
end
end
% ----- numerical error management -----
% Adjust maximal coordinate axis deviations
if any(sigma*sqrt(diagC) > maxdx)
sigma = min(maxdx ./ sqrt(diagC));
%warning(['Iteration ' num2str(countiter) ': coordinate axis std ' ...
% 'deviation at upper limit of ' num2str(maxdx)]);
% stopflag(end+1) = {'maxcoorddev'};
end
% Adjust minimal coordinate axis deviations
if any(sigma*sqrt(diagC) < mindx)
sigma = max(mindx ./ sqrt(diagC)) * exp(0.05+cs/damps);
%warning(['Iteration ' num2str(countiter) ': coordinate axis std ' ...
% 'deviation at lower limit of ' num2str(mindx)]);
% stopflag(end+1) = {'mincoorddev'};;
end
% Adjust too low coordinate axis deviations
if any(xmean == xmean + 0.2*sigma*sqrt(diagC))
if stopOnWarnings
stopflag(end+1) = {'warnnoeffectcoord'};
else
warning(['Iteration ' num2str(countiter) ': coordinate axis std ' ...
'deviation too low' ]);
if flgDiagonalOnly
diagC = diagC + (ccov1_sep+ccovmu_sep) * (diagC .* ...
(xmean == xmean + 0.2*sigma*sqrt(diagC)));
else
C = C + (ccov1+ccovmu) * diag(diagC .* ...
(xmean == xmean + 0.2*sigma*sqrt(diagC)));
end
sigma = sigma * exp(0.05+cs/damps); end
end
% Adjust step size in case of (numerical) precision problem
if flgDiagonalOnly
tmp = 0.1*sigma*diagD;
else
tmp = 0.1*sigma*BD(:,1+floor(mod(countiter,N)));
end
if all(xmean == xmean + tmp)
i = 1+floor(mod(countiter,N));
if stopOnWarnings
stopflag(end+1) = {'warnnoeffectaxis'};
else
warning(['Iteration ' num2str(countiter) ...
': main axis standard deviation ' ...
num2str(sigma*diagD(i)) ' has no effect' ]);
sigma = sigma * exp(0.2+cs/damps);
end
end
% Adjust step size in case of equal function values (flat fitness)
% isequalfuncvalues = 0;
if fitness.sel(1) == fitness.sel(1+ceil(0.1+lambda/4))
% isequalfuncvalues = 1;
if stopOnEqualFunctionValues
arrEqualFunvals = [countiter arrEqualFunvals(1:end-1)];
% stop if this happens in more than 33%
if arrEqualFunvals(end) > countiter - 3 * length(arrEqualFunvals)
stopflag(end+1) = {'equalfunvals'};
end
else
if flgWarnOnEqualFunctionValues
warning(['Iteration ' num2str(countiter) ...
': equal function values f=' num2str(fitness.sel(1)) ...
' at maximal main axis sigma ' ...
num2str(sigma*max(diagD))]);
end
sigma = sigma * exp(0.2+cs/damps);
end
end
% Adjust step size in case of equal function values
if countiter > 2 && myrange([fitness.hist fitness.sel(1)]) == 0
if stopOnWarnings
stopflag(end+1) = {'warnequalfunvalhist'};
else
warning(['Iteration ' num2str(countiter) ...
': equal function values in history at maximal main ' ...
'axis sigma ' num2str(sigma*max(diagD))]);
sigma = sigma * exp(0.2+cs/damps);
end
end
% ----- end numerical error management -----
% Keep overall best solution
out.evals = counteval;
out.solutions.evals = counteval;
out.solutions.mean.x = xmean;
out.solutions.mean.f = fmean;
out.solutions.mean.evals = counteval;
out.solutions.recentbest.x = arxvalid(:, fitness.idx(1));
out.solutions.recentbest.f = fitness.raw(1);
out.solutions.recentbest.evals = counteval + fitness.idx(1) - lambda;
out.solutions.recentworst.x = arxvalid(:, fitness.idx(end));
out.solutions.recentworst.f = fitness.raw(end);
out.solutions.recentworst.evals = counteval + fitness.idx(end) - lambda;
if fitness.hist(1) < out.solutions.bestever.f
out.solutions.bestever.x = arxvalid(:, fitness.idx(1));
out.solutions.bestever.f = fitness.hist(1);
out.solutions.bestever.evals = counteval + fitness.idx(1) - lambda;
bestever = out.solutions.bestever; end
% Set stop flag
if fitness.raw(1) <= stopFitness, stopflag(end+1) = {'fitness'}; end
if counteval >= stopMaxFunEvals, stopflag(end+1) = {'maxfunevals'}; end
if countiter >= stopMaxIter, stopflag(end+1) = {'maxiter'}; end
if all(sigma*(max(abs(pc), sqrt(diagC))) < stopTolX)
stopflag(end+1) = {'tolx'};
end
if any(sigma*sqrt(diagC) > stopTolUpX)
stopflag(end+1) = {'tolupx'};
end
if sigma*max(diagD) == 0 % should never happen
stopflag(end+1) = {'bug'};
end
if countiter > 2 && myrange([fitness.sel fitness.hist]) <= stopTolFun
stopflag(end+1) = {'tolfun'};
end
if countiter >= length(fitness.hist) && myrange(fitness.hist) <= stopTolHistFun
stopflag(end+1) = {'tolhistfun'};
end
l = floor(length(fitness.histbest)/3);
if 1 < 2 && stopOnStagnation && ... % leads sometimes early stop on ftablet, fcigtab
countiter > N * (5+100/lambda) && ...
length(fitness.histbest) > 100 && ...
median(fitness.histmedian(1:l)) >= median(fitness.histmedian(end-l:end)) && ...
median(fitness.histbest(1:l)) >= median(fitness.histbest(end-l:end))
stopflag(end+1) = {'stagnation'};
end
if counteval >= stopFunEvals || countiter >= stopIter
stopflag(end+1) = {'stoptoresume'};
if length(stopflag) == 1 && flgsaving == 0
error('To resume later the saving option needs to be set');
end
end
% read stopping message from file signals.par
if flgreadsignals
fid = fopen('./signals.par', 'rt'); % can be performance critical
while fid > 0
strline = fgetl(fid); %fgets(fid, 300);
if strline < 0 % fgets and fgetl returns -1 at end of file
break;
end
% 'stop filename' sets stopflag to manual
str = sscanf(strline, ' %s %s', 2);
if strcmp(str, ['stop' opts.LogFilenamePrefix])
stopflag(end+1) = {'manual'};
break;
end
% 'skip filename run 3' skips a run, but not the last
str = sscanf(strline, ' %s %s %s', 3);
if strcmp(str, ['skip' opts.LogFilenamePrefix 'run'])
i = strfind(strline, 'run');
if irun == sscanf(strline(i+3:end), ' %d ', 1) && irun <= myeval(opts.Restarts)
stopflag(end+1) = {'skipped'};
end
end
end % while, break
if fid > 0
fclose(fid);
clear fid; % prevents strange error under octave
end
end
out.stopflag = stopflag;
% ----- output generation -----
if verbosemodulo > 0 && isfinite(verbosemodulo)
if countiter == 1 || mod(countiter, 10*verbosemodulo) < 1 disp(['Iterat, #Fevals: Function Value (median,worst) ' ...
'|Axis Ratio|' ...
'idx:Min SD idx:Max SD']);
end
if mod(countiter, verbosemodulo) < 1 ...
|| (verbosemodulo > 0 && isfinite(verbosemodulo) && ...
(countiter < 3 || ~isempty(stopflag)))
[minstd minstdidx] = min(sigma*sqrt(diagC));
[maxstd maxstdidx] = max(sigma*sqrt(diagC));
% format display nicely
disp([repmat(' ',1,4-floor(log10(countiter))) ...
num2str(countiter) ' , ' ...
repmat(' ',1,5-floor(log10(counteval))) ...
num2str(counteval) ' : ' ...
num2str(fitness.hist(1), '%.13e') ...
' +(' num2str(median(fitness.raw)-fitness.hist(1), '%.0e ') ...
',' num2str(max(fitness.raw)-fitness.hist(1), '%.0e ') ...
') | ' ...
num2str(max(diagD)/min(diagD), '%4.2e') ' | ' ...
repmat(' ',1,1-floor(log10(minstdidx))) num2str(minstdidx) ':' ...
num2str(minstd, ' %.1e') ' ' ...
repmat(' ',1,1-floor(log10(maxstdidx))) num2str(maxstdidx) ':' ...
num2str(maxstd, ' %.1e')]);
end
end
% measure time for recording data
if countiter < 3
time.c = 0.05;
time.nonoutput = 0;
time.recording = 0;
time.saving = 0.15; % first saving after 3 seconds of 100 iterations
time.plotting = 0;
elseif countiter > 300
% set backward horizon, must be long enough to cover infrequent plotting etc
% time.c = min(1, time.nonoutput/3 + 1e-9);
time.c = max(1e-5, 0.1/sqrt(countiter)); % mean over all or 1e-5
end
% get average time per iteration
time.t1 = clock;
time.act = max(0,etime(time.t1, time.t0));
time.nonoutput = (1-time.c) * time.nonoutput ...
+ time.c * time.act;
time.recording = (1-time.c) * time.recording; % per iteration
time.saving = (1-time.c) * time.saving;
time.plotting = (1-time.c) * time.plotting;
% record output data, concerning time issues
if savemodulo && savetime && (countiter < 1e2 || ~isempty(stopflag) || ...
countiter >= outiter + savemodulo)
outiter = countiter;
% Save output data to files
for namecell = filenames(:)'
name = namecell{:};
[fid, err] = fopen(['./' filenameprefix name '.dat'], 'a');
if fid < 1 % err ~= 0
warning(['could not open ' filenameprefix name '.dat']);
else
if strcmp(name, 'axlen')
fprintf(fid, '%d %d %e %e %e ', countiter, counteval, sigma, ...
max(diagD), min(diagD));
fprintf(fid, '%e ', sort(diagD));
fprintf(fid, '\n');
elseif strcmp(name, 'disp') % TODO
elseif strcmp(name, 'fit')
fprintf(fid, '%ld %ld %e %e %25.18e %25.18e %25.18e %25.18e', ...
countiter, counteval, sigma, max(diagD)/min(diagD), ...
out.solutions.bestever.f, ... fitness.raw(1), median(fitness.raw), fitness.raw(end));
if ~isempty(varargin) && length(varargin{1}) == 1 && isnumeric(varargin{1}) && varargin{1} ~= 0
fprintf(fid, ' %f', varargin{1});
end
fprintf(fid, '\n');
elseif strcmp(name, 'stddev')
fprintf(fid, '%ld %ld %e 0 0 ', countiter, counteval, sigma);
fprintf(fid, '%e ', sigma*sqrt(diagC));
fprintf(fid, '\n');
elseif strcmp(name, 'xmean')
if isnan(fmean)
fprintf(fid, '%ld %ld 0 0 0 ', countiter, counteval);
else
fprintf(fid, '%ld %ld 0 0 %e ', countiter, counteval, fmean);
end
fprintf(fid, '%e ', xmean);
fprintf(fid, '\n');
elseif strcmp(name, 'xrecentbest')
% TODO: fitness is inconsistent with x-value
fprintf(fid, '%ld %ld %25.18e 0 0 ', countiter, counteval, fitness.raw(1));
fprintf(fid, '%e ', arx(:,fitness.idx(1)));
fprintf(fid, '\n');
end
fclose(fid);
end
end
% get average time for recording data
time.t2 = clock;
time.recording = time.recording + time.c * max(0,etime(time.t2, time.t1));
% plot
if flgplotting && countiter > 1
if countiter == 2
iterplotted = 0;
end
if ~isempty(stopflag) || ...
((time.nonoutput+time.recording) * (countiter - iterplotted) > 1 && ...
time.plotting < 0.05 * (time.nonoutput+time.recording))
local_plotcmaesdat(324, filenameprefix);
iterplotted = countiter;
% outplot(out); % outplot defined below
if time.plotting == 0 % disregard opening of the window
time.plotting = time.nonoutput+time.recording;
else
time.plotting = time.plotting + time.c * max(0,etime(clock, time.t2));
end
end
end
if countiter > 100 + 20 && savemodulo && ...
time.recording * countiter > 0.1 && ... % absolute time larger 0.1 second
time.recording > savetime * (time.nonoutput+time.recording) / 100
savemodulo = floor(1.1 * savemodulo) + 1;
% disp('++savemodulo'); %qqq
end
end % if output
% save everything
time.t3 = clock;
if ~isempty(stopflag) || time.saving < 0.05 * time.nonoutput || countiter == 100
xmin = arxvalid(:, fitness.idx(1));
fmin = fitness.raw(1);
if flgsaving && countiter > 2
clear idx; % prevents error under octave
% -v6 : non-compressed non-unicode for version 6 and earlier
if ~isempty(strsaving) && ~isoctave
save('-mat', strsaving, opts.SaveFilename); % for inspection and possible restart
else
save('-mat', opts.SaveFilename); % for inspection and possible restart
end time.saving = time.saving + time.c * max(0,etime(clock, time.t3));
end
end
time.t0 = clock;
% ----- end output generation -----
end % while, end generation loop
% -------------------- Final Procedures -------------------------------
% Evaluate xmean and return best recent point in xmin
fmin = fitness.raw(1);
xmin = arxvalid(:, fitness.idx(1)); % Return best point of last generation.
if length(stopflag) > sum(strcmp(stopflag, 'stoptoresume')) % final stopping
out.solutions.mean.f = ...
feval(fitfun, xintobounds(xmean, lbounds, ubounds), varargin{:});
counteval = counteval + 1;
out.solutions.mean.evals = counteval;
if out.solutions.mean.f < fitness.raw(1)
fmin = out.solutions.mean.f;
xmin = xintobounds(xmean, lbounds, ubounds); % Return xmean as best point
end
if out.solutions.mean.f < out.solutions.bestever.f
out.solutions.bestever = out.solutions.mean; % Return xmean as bestever point
out.solutions.bestever.x = xintobounds(xmean, lbounds, ubounds);
bestever = out.solutions.bestever;
end
end
% Save everything and display final message
if flgsavingfinal
clear idx; % prevents error under octave
if ~isempty(strsaving) && ~isoctave
save('-mat', strsaving, opts.SaveFilename); % for inspection and possible restart
else
save('-mat', opts.SaveFilename); % for inspection and possible restart
end
message = [' (saved to ' opts.SaveFilename ')'];
else
message = [];
end
if flgdisplay
disp(['#Fevals: f(returned x) | bestever.f | stopflag' ...
message]);
if isoctave
strstop = stopflag(:);
else
strcat(stopflag(:), '.');
end
strstop = stopflag(:); %strcat(stopflag(:), '.');
disp([repmat(' ',1,6-floor(log10(counteval))) ...
num2str(counteval, '%6.0f') ': ' num2str(fmin, '%.11e') ' | ' ...
num2str(out.solutions.bestever.f, '%.11e') ' | ' ...
strstop{1:end}]);
if N < 102
disp(['mean solution:' sprintf(' %+.1e', xmean)]);
disp(['std deviation:' sprintf(' %.1e', sigma*sqrt(diagC))]);
disp(sprintf('use plotcmaesdat.m for plotting the output at any time (option LogModulo must not be zero)'));
end
if exist('sfile', 'var')
disp(['Results saved in ' sfile]);
end
end
out.arstopflags{irun} = stopflag;
if any(strcmp(stopflag, 'fitness')) ...
|| any(strcmp(stopflag, 'maxfunevals')) ...
|| any(strcmp(stopflag, 'stoptoresume')) ... || any(strcmp(stopflag, 'manual'))
break;
end
end % while irun <= Restarts
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function [x, idx] = xintobounds(x, lbounds, ubounds)
%
% x can be a column vector or a matrix consisting of column vectors
%
if ~isempty(lbounds)
if length(lbounds) == 1
idx = x < lbounds;
x(idx) = lbounds;
else
arbounds = repmat(lbounds, 1, size(x,2));
idx = x < arbounds;
x(idx) = arbounds(idx);
end
else
idx = 0;
end
if ~isempty(ubounds)
if length(ubounds) == 1
idx2 = x > ubounds;
x(idx2) = ubounds;
else
arbounds = repmat(ubounds, 1, size(x,2));
idx2 = x > arbounds;
x(idx2) = arbounds(idx2);
end
else
idx2 = 0;
end
idx = idx2-idx;
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function opts=getoptions(inopts, defopts)
% OPTS = GETOPTIONS(INOPTS, DEFOPTS) handles an arbitrary number of
% optional arguments to a function. The given arguments are collected
% in the struct INOPTS. GETOPTIONS matches INOPTS with a default
% options struct DEFOPTS and returns the merge OPTS. Empty or missing
% fields in INOPTS invoke the default value. Fieldnames in INOPTS can
% be abbreviated.
%
% The returned struct OPTS is first assigned to DEFOPTS. Then any
% field value in OPTS is replaced by the respective field value of
% INOPTS if (1) the field unambiguously (case-insensitive) matches
% with the fieldname in INOPTS (cut down to the length of the INOPTS
% fieldname) and (2) the field is not empty.
%
% Example:
% In the source-code of the function that needs optional
% arguments, the last argument is the struct of optional
% arguments:
%
% function results = myfunction(mandatory_arg, inopts)
% % Define four default options
% defopts.PopulationSize = 200;
% defopts.ParentNumber = 50;
% defopts.MaxIterations = 1e6;
% defopts.MaxSigma = 1;
%
% % merge default options with input options
% opts = getoptions(inopts, defopts);
%
% % Thats it! From now on the values in opts can be used
% for i = 1:opts.PopulationSize% % do whatever
% if sigma > opts.MaxSigma
% % do whatever
% end
% end
%
% For calling the function myfunction with default options:
% myfunction(argument1, []);
% For calling the function myfunction with modified options:
% opt.pop = 100; % redefine PopulationSize option
% opt.PAR = 10; % redefine ParentNumber option
% opt.maxiter = 2; % opt.max=2 is ambiguous and would result in an error
% myfunction(argument1, opt);
%
% 04/07/19: Entries can be structs itself leading to a recursive
% call to getoptions.
%
if nargin < 2 || isempty(defopts) % no default options available
opts=inopts;
return;
elseif isempty(inopts) % empty inopts invoke default options
opts = defopts;
return;
elseif ~isstruct(defopts) % handle a single option value
if isempty(inopts)
opts = defopts;
elseif ~isstruct(inopts)
opts = inopts;
else
error('Input options are a struct, while default options are not');
end
return;
elseif ~isstruct(inopts) % no valid input options
error('The options need to be a struct or empty');
end
opts = defopts; % start from defopts
% if necessary overwrite opts fields by inopts values
defnames = fieldnames(defopts);
idxmatched = []; % indices of defopts that already matched
for name = fieldnames(inopts)'
name = name{1}; % name of i-th inopts-field
if isoctave
for i = 1:size(defnames, 1)
idx(i) = strncmp(lower(defnames(i)), lower(name), length(name));
end
else
idx = strncmpi(defnames, name, length(name));
end
if sum(idx) > 1
error(['option "' name '" is not an unambigous abbreviation. ' ...
'Use opts=RMFIELD(opts, ''' name, ...
''') to remove the field from the struct.']);
end
if sum(idx) == 1
defname = defnames{find(idx)};
if ismember(find(idx), idxmatched)
error(['input options match more than ones with "' ...
defname '". ' ...
'Use opts=RMFIELD(opts, ''' name, ...
''') to remove the field from the struct.']);
end
idxmatched = [idxmatched find(idx)];
val = getfield(inopts, name);
% next line can replace previous line from MATLAB version 6.5.0 on and in octave
% val = inopts.(name);
if isstruct(val) % valid syntax only from version 6.5.0
opts = setfield(opts, defname, ... getoptions(val, getfield(defopts, defname)));
elseif isstruct(getfield(defopts, defname))
% next three lines can replace previous three lines from MATLAB
% version 6.5.0 on
% opts.(defname) = ...
% getoptions(val, defopts.(defname));
% elseif isstruct(defopts.(defname))
warning(['option "' name '" disregarded (must be struct)']);
elseif ~isempty(val) % empty value: do nothing, i.e. stick to default
opts = setfield(opts, defnames{find(idx)}, val);
% next line can replace previous line from MATLAB version 6.5.0 on
% opts.(defname) = inopts.(name);
end
else
warning(['option "' name '" disregarded (unknown field name)']);
end
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function res=myeval(s)
if ischar(s)
res = evalin('caller', s);
else
res = s;
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function res=myevalbool(s)
if ~ischar(s) % s may not and cannot be empty
res = s;
else % evaluation string s
if strncmpi(lower(s), 'yes', 3) || strncmpi(s, 'on', 2) ...
|| strncmpi(s, 'true', 4) || strncmp(s, '1 ', 2)
res = 1;
elseif strncmpi(s, 'no', 2) || strncmpi(s, 'off', 3) ...
|| strncmpi(s, 'false', 5) || strncmp(s, '0 ', 2)
res = 0;
else
try res = evalin('caller', s); catch
error(['String value "' s '" cannot be evaluated']);
end
try res ~= 0; catch
error(['String value "' s '" cannot be evaluated reasonably']);
end
end
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function res = isoctave
% any hack to find out whether we are running octave
s = version;
res = 0;
if exist('fflush', 'builtin') && eval(s(1)) < 7
res = 1;
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function flush
if isoctave
feval('fflush', stdout);
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% ----- replacements for statistic toolbox functions ------------% ---------------------------------------------------------------
% ---------------------------------------------------------------
function res=myrange(x)
res = max(x) - min(x);
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function res = myprctile(inar, perc, idx)
%
% Computes the percentiles in vector perc from vector inar
% returns vector with length(res)==length(perc)
% idx: optional index-array indicating sorted order
%
N = length(inar);
flgtranspose = 0;
% sizes
if size(perc,1) > 1
perc = perc';
flgtranspose = 1;
if size(perc,1) > 1
error('perc must not be a matrix');
end
end
if size(inar, 1) > 1 && size(inar,2) > 1
error('data inar must not be a matrix');
end
% sort inar
if nargin < 3 || isempty(idx)
[sar idx] = sort(inar);
else
sar = inar(idx);
end
res = [];
for p = perc
if p <= 100*(0.5/N)
res(end+1) = sar(1);
elseif p >= 100*((N-0.5)/N)
res(end+1) = sar(N);
else
% find largest index smaller than required percentile
availablepercentiles = 100*((1:N)-0.5)/N;
i = max(find(p > availablepercentiles));
% interpolate linearly
res(end+1) = sar(i) ...
+ (sar(i+1)-sar(i))*(p - availablepercentiles(i)) ...
/ (availablepercentiles(i+1) - availablepercentiles(i));
end
end
if flgtranspose
res = res';
end
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% ---------------------------------------------------------------
function [s ranks rankDelta] = local_noisemeasurement(arf1, arf2, lamreev, theta, cutlimit)
% function [s ranks rankDelta] = noisemeasurement(arf1, arf2, lamreev, theta)
%
% Input: % arf1, arf2 : two arrays of function values. arf1 is of size 1xlambda,
% arf2 may be of size 1xlamreev or 1xlambda. The first lamreev values
% in arf2 are (re-)evaluations of the respective solutions, i.e.
% arf1(1) and arf2(1) are two evaluations of "the first" solution.
% lamreev: number of reevaluated individuals in arf2
% theta : parameter theta for the rank change limit, between 0 and 1,
% typically between 0.2 and 0.7.
% cutlimit (optional): output s is computed as a mean of rankchange minus
% threshold, where rankchange is <=2*(lambda-1). cutlimit limits
% abs(rankchange minus threshold) in this calculation to cutlimit.
% cutlimit=1 evaluates basically the sign only. cutlimit=2 could be
% the rank change with one solution (both evaluations of it).
%
% Output:
% s : noise measurement, s>0 means the noise measure is above the
% acceptance threshold
% ranks : 2xlambda array, corresponding to [arf1; arf2], of ranks
% of arf1 and arf2 in the set [arf1 arf2], values are in [1:2*lambda]
% rankDelta: 1xlambda array of rank movements of arf2 compared to
% arf1. rankDelta(i) agrees with the number of values from
% the set [arf1 arf2] that lie between arf1(i) and arf2(i).
%
% Note: equal function values might lead to somewhat spurious results.
% For this case a revision is advisable.
%%% verify input argument sizes
if size(arf1,1) ~= 1
error('arf1 must be an 1xlambda array');
elseif size(arf2,1) ~= 1
error('arf2 must be an 1xsomething array');
elseif size(arf1,2) < size(arf2,2) % not really necessary, but saver
error('arf2 must not be smaller than arf1 in length');
end
lam = size(arf1, 2);
if size(arf1,2) ~= size(arf2,2)
arf2(end+1:lam) = arf1((size(arf2,2)+1):lam);
end
if nargin < 5
cutlimit = inf;
end
%%% capture unusual values
if any(diff(arf1) == 0)
% this will presumably interpreted as rank change, because
% sort(ones(...)) returns 1,2,3,...
warning([num2str(sum(diff(arf1)==0)) ' equal function values']);
end
%%% compute rank changes into rankDelta
% compute ranks
[ignore, idx] = sort([arf1 arf2]);
[ignore, ranks] = sort(idx);
ranks = reshape(ranks, lam, 2)';
rankDelta = ranks(1,:) - ranks(2,:) - sign(ranks(1,:) - ranks(2,:));
%%% compute rank change limits using both ranks(1,...) and ranks(2,...)
for i = 1:lamreev
sumlim(i) = ...
0.5 * (...
myprctile(abs((1:2*lam-1) - (ranks(1,i) - (ranks(1,i)>ranks(2,i)))), ...
theta*50) ...
+ myprctile(abs((1:2*lam-1) - (ranks(2,i) - (ranks(2,i)>ranks(1,i)))), ...
theta*50));
end
%%% compute measurement
%s = abs(rankDelta(1:lamreev)) - sumlim; % lives roughly in 0..2*lambda
% max: 1 rankchange in 2*lambda is always fines = abs(rankDelta(1:lamreev)) - max(1, sumlim); % lives roughly in 0..2*lambda
% cut-off limit
idx = abs(s) > cutlimit;
s(idx) = sign(s(idx)) * cutlimit;
s = mean(s);
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% ---------------------------------------------------------------
% just a "local" copy of plotcmaesdat.m, with manual_mode set to zero
function local_plotcmaesdat(figNb, filenameprefix, filenameextension, objectvarname)
% PLOTCMAESDAT;
% PLOTCMAES(FIGURENUMBER_iBEGIN_iEND, FILENAMEPREFIX, FILENAMEEXTENSION, OBJECTVARNAME);
% plots output from CMA-ES, e.g. cmaes.m, Java class CMAEvolutionStrategy...
% mod(figNb,100)==1 plots versus iterations.
%
% PLOTCMAES([101 300]) plots versus iteration, from iteration 300.
% PLOTCMAES([100 150 800]) plots versus function evaluations, between iteration 150 and 800.
%
% Upper left subplot: blue/red: function value of the best solution in the
% recent population, cyan: same function value minus best
% ever seen function value, green: sigma, red: ratio between
% longest and shortest principal axis length which is equivalent
% to sqrt(cond(C)).
% Upper right plot: time evolution of the distribution mean (default) or
% the recent best solution vector.
% Lower left: principal axes lengths of the distribution ellipsoid,
% equivalent with the sqrt(eig(C)) square root eigenvalues of C.
% Lower right: magenta: minimal and maximal "true" standard deviation
% (with sigma included) in the coordinates, other colors: sqrt(diag(C))
% of all diagonal elements of C, if C is diagonal they equal to the
% lower left.
%
% Files [FILENAMEPREFIX name FILENAMEEXTENSION] are used, where
% name = axlen, OBJECTVARNAME (xmean|xrecentbest), fit, or stddev.
%
manual_mode = 0;
if nargin < 1 || isempty(figNb)
figNb = 325;
end
if nargin < 2 || isempty(filenameprefix)
filenameprefix = 'outcmaes';
end
if nargin < 3 || isempty(filenameextension)
filenameextension = '.dat';
end
if nargin < 4 || isempty(objectvarname)
objectvarname = 'xmean';
objectvarname = 'xrecentbest';
end
% load data
d.x = load([filenameprefix objectvarname filenameextension]);
% d.x = load([filenameprefix 'xmean' filenameextension]);
% d.x = load([filenameprefix 'xrecentbest' filenameextension]);
d.f = load([filenameprefix 'fit' filenameextension]);
d.std = load([filenameprefix 'stddev' filenameextension]);
d.D = load([filenameprefix 'axlen' filenameextension]);
% interpret entries in figNb for cutting out some data
if length(figNb) > 1
iend = inf;
istart = figNb(2);
if length(figNb) > 2
iend = figNb(3);
end
figNb = figNb(1); d.x = d.x(d.x(:,1) >= istart & d.x(:,1) <= iend, :);
d.f = d.f(d.f(:,1) >= istart & d.f(:,1) <= iend, :);
d.std = d.std(d.std(:,1) >= istart & d.std(:,1) <= iend, :);
d.D = d.D(d.D(:,1) >= istart & d.D(:,1) <= iend, :);
end
% decide for x-axis
iabscissa = 2; % 1== versus iterations, 2==versus fevals
if mod(figNb,100) == 1
iabscissa = 1; % a short hack
end
if iabscissa == 1
xlab ='iterations';
elseif iabscissa == 2
xlab = 'function evaluations';
end
if size(d.x, 2) < 1000
minxend = 1.03*d.x(end, iabscissa);
else
minxend = 0;
end
% set up figure window
if manual_mode
figure(figNb); % just create and raise the figure window
else
if 1 < 3 && evalin('caller', 'iterplotted') == 0 && evalin('caller', 'irun') == 1
figure(figNb); % incomment this, if figure raise in the beginning is not desired
elseif ismember(figNb, findobj('Type', 'figure'))
set(0, 'CurrentFigure', figNb); % prevents raise of existing figure window
else
figure(figNb);
end
end
% plot fitness etc
foffset = 1e-99;
dfit = d.f(:,6)-min(d.f(:,6));
[ignore idxbest] = min(dfit);
dfit(dfit<1e-98) = NaN;
subplot(2,2,1); hold off;
dd = abs(d.f(:,7:8)) + foffset;
dd(d.f(:,7:8)==0) = NaN;
semilogy(d.f(:,iabscissa), dd, '-k'); hold on;
% additional fitness data, for example constraints values
if size(d.f,2) > 8
dd = abs(d.f(:,9:end)) + 10*foffset; % a hack
% dd(d.f(:,9:end)==0) = NaN;
semilogy(d.f(:,iabscissa), dd, '-m'); hold on;
if size(d.f,2) > 12
semilogy(d.f(:,iabscissa),abs(d.f(:,[7 8 11 13]))+foffset,'-k'); hold on;
end
end
idx = find(d.f(:,6)>1e-98); % positive values
if ~isempty(idx) % otherwise non-log plot gets hold
semilogy(d.f(idx,iabscissa), d.f(idx,6)+foffset, '.b'); hold on;
end
idx = find(d.f(:,6) < -1e-98); % negative values
if ~isempty(idx)
semilogy(d.f(idx, iabscissa), abs(d.f(idx,6))+foffset,'.r'); hold on;
end
semilogy(d.f(:,iabscissa),abs(d.f(:,6))+foffset,'-b'); hold on;
semilogy(d.f(:,iabscissa),dfit,'-c'); hold on;
semilogy(d.f(:,iabscissa),(d.f(:,4)),'-r'); hold on; % AR
semilogy(d.std(:,iabscissa), [max(d.std(:,6:end)')' ...
min(d.std(:,6:end)')'], '-m'); % max,min std
maxval = max(d.std(end,6:end));
minval = min(d.std(end,6:end)); text(d.std(end,iabscissa), maxval, sprintf('%.0e', maxval));
text(d.std(end,iabscissa), minval, sprintf('%.0e', minval));
semilogy(d.std(:,iabscissa),(d.std(:,3)),'-g'); % sigma
% plot best f
semilogy(d.f(idxbest,iabscissa),min(dfit),'*c'); hold on;
semilogy(d.f(idxbest,iabscissa),abs(d.f(idxbest,6))+foffset,'*r'); hold on;
ax = axis;
ax(2) = max(minxend, ax(2));
axis(ax);
yannote = 10^(log10(ax(3)) + 0.05*(log10(ax(4))-log10(ax(3))));
text(ax(1), yannote, ...
[ 'f=' num2str(d.f(end,6), '%.15g') ]);
title('blue:abs(f), cyan:f-min(f), green:sigma, red:axis ratio');
grid on;
subplot(2,2,2); hold off;
plot(d.x(:,iabscissa), d.x(:,6:end),'-'); hold on;
ax = axis;
ax(2) = max(minxend, ax(2));
axis(ax);
% add some annotation lines
[ignore idx] = sort(d.x(end,6:end));
% choose no more than 25 indices
idxs = round(linspace(1, size(d.x,2)-5, min(size(d.x,2)-5, 25)));
yy = repmat(NaN, 2, size(d.x,2)-5);
yy(1,:) = d.x(end, 6:end);
yy(2,idx(idxs)) = linspace(ax(3), ax(4), length(idxs));
plot([d.x(end,iabscissa) ax(2)], yy, '-');
plot(repmat(d.x(end,iabscissa),2), [ax(3) ax(4)], 'k-');
for i = idx(idxs)
text(ax(2), yy(2,i), ...
['x(' num2str(i) ')=' num2str(yy(1,i))]);
end
lam = 'NA';
if size(d.x, 1) > 1 && d.x(end, 1) > d.x(end-1, 1)
lam = num2str((d.x(end, 2) - d.x(end-1, 2)) / (d.x(end, 1) - d.x(end-1, 1)));
end
title(['Object Variables (' num2str(size(d.x, 2)-5) ...
'-D, popsize~' lam ')']);grid on;
subplot(2,2,3); hold off; semilogy(d.D(:,iabscissa), d.D(:,6:end), '-');
ax = axis;
ax(2) = max(minxend, ax(2));
axis(ax);
title('Principal Axes Lengths');grid on;
xlabel(xlab);
subplot(2,2,4); hold off;
% semilogy(d.std(:,iabscissa), d.std(:,6:end), 'k-'); hold on;
% remove sigma from stds
d.std(:,6:end) = d.std(:,6:end) ./ (d.std(:,3) * ones(1,size(d.std,2)-5));
semilogy(d.std(:,iabscissa), d.std(:,6:end), '-'); hold on;
if 11 < 3 % max and min std
semilogy(d.std(:,iabscissa), [d.std(:,3).*max(d.std(:,6:end)')' ...
d.std(:,3).*min(d.std(:,6:end)')'], '-m', 'linewidth', 2);
maxval = max(d.std(end,6:end));
minval = min(d.std(end,6:end));
text(d.std(end,iabscissa), d.std(end,3)*maxval, sprintf('max=%.0e', maxval));
text(d.std(end,iabscissa), d.std(end,3)*minval, sprintf('min=%.0e', minval));
end
ax = axis;
ax(2) = max(minxend, ax(2));
axis(ax);
% add some annotation lines [ignore idx] = sort(d.std(end,6:end));
% choose no more than 25 indices
idxs = round(linspace(1, size(d.x,2)-5, min(size(d.x,2)-5, 25)));
yy = repmat(NaN, 2, size(d.std,2)-5);
yy(1,:) = d.std(end, 6:end);
yy(2,idx(idxs)) = logspace(log10(ax(3)), log10(ax(4)), length(idxs));
semilogy([d.std(end,iabscissa) ax(2)], yy, '-');
semilogy(repmat(d.std(end,iabscissa),2), [ax(3) ax(4)], 'k-');
for i = idx(idxs)
text(ax(2), yy(2,i), [' ' num2str(i)]);
end
title('Standard Deviations in Coordinates divided by sigma');grid on;
xlabel(xlab);
if figNb ~= 324
% zoom on; % does not work in Octave
end
drawnow;
% ---------------------------------------------------------------
% --------------- TEST OBJECTIVE FUNCTIONS ----------------------
% ---------------------------------------------------------------
%%% Unimodal functions
function f=fjens1(x)
%
% use population size about 2*N
%
f = sum((x>0) .* x.^1, 1);
if any(any(x<0))
idx = sum(x < 0, 1) > 0;
f(idx) = 1e3;
% f = f + 1e3 * sum(x<0, 1);
% f = f + 10 * sum((x<0) .* x.^2, 1);
f(idx) = f(idx) + 1e-3*abs(randn(1,sum(idx)));
% f(idx) = NaN;
end
function f=fsphere(x)
f = sum(x.^2,1);
function f=fmax(x)
f = max(abs(x), [], 1);
function f=fssphere(x)
f=sqrt(sum(x.^2, 1));
% lb = -0.512; ub = 512;
% xfeas = x;
% xfeas(x<lb) = lb;
% xfeas(x>ub) = ub;
% f=sum(xfeas.^2, 1);
% f = f + 1e-9 * sum((xfeas-x).^2);
function f=fspherenoise(x, Nevals)
if nargin < 2 || isempty(Nevals)
Nevals = 1;
end
[N,popsi] = size(x);
% x = x .* (1 + 0.3e-0 * randn(N, popsi)/(2*N)); % actuator noise
fsum = 10.^(0*(0:N-1)/(N-1)) * x.^2;
% f = 0*rand(1,1) ...
% + fsum ...
% + fsum .* (2*randn(1,popsi) ./ randn(1,popsi).^0 / (2*N)) ...
% + 1*fsum.^0.9 .* 2*randn(1,popsi) / (2*N); %
% f = fsum .* exp(0.1*randn(1,popsi));
f = fsum .* (1 + (10/(N+10)/sqrt(Nevals))*randn(1,popsi));
% f = fsum .* (1 + (0.1/N)*randn(1,popsi)./randn(1,popsi).^1);
idx = rand(1,popsi) < 0.0;
if sum(idx) > 0
f(idx) = f(idx) + 1e3*exp(randn(1,sum(idx)));
end
function f=fmixranks(x)
N = size(x,1);
f=(10.^(0*(0:(N-1))/(N-1))*x.^2).^0.5;
if size(x, 2) > 1 % compute ranks, if it is a population
[ignore, idx] = sort(f);
[ignore, ranks] = sort(idx);
k = 9; % number of solutions randomly permuted, lambda/2-1
% works still quite well (two time slower)
for i = k+1:k-0:size(x,2)
idx(i-k+(1:k)) = idx(i-k+randperm(k));
end
%disp([ranks' f'])
[ignore, ranks] = sort(idx);
%disp([ranks' f'])
%pause
f = ranks+1e-9*randn(1,1);
end
function f = fsphereoneax(x)
f = x(1)^2;
f = mean(x)^2;
function f=frandsphere(x)
N = size(x,1);
idx = ceil(N*rand(7,1));
f=sum(x(idx).^2);
function f=fspherelb0(x, M) % lbound at zero for 1:M needed
if nargin < 2 M = 0; end
N = size(x,1);
% M active bounds, f_i = 1 for x = 0
f = -M + sum((x(1:M) + 1).^2);
f = f + sum(x(M+1:N).^2);
function f=fspherehull(x)
% Patton, Dexter, Goodman, Punch
% in -500..500
% spherical ridge through zeros(N,1)
% worst case start point seems x = 2*100*sqrt(N)
% and small step size
N = size(x,1);
f = norm(x) + (norm(x-100*sqrt(N)) - 100*N)^2;
function f=fellilb0(x, idxM, scal) % lbound at zero for 1:M needed
N = size(x,1);
if nargin < 3 || isempty(scal)
scal = 100;
end
scale=scal.^((0:N-1)/(N-1));
if nargin < 2 || isempty(idxM)
idxM = 1:N;
end
%scale(N) = 1e0;
% M active bounds
xopt = 0.1;
x(idxM) = x(idxM) + xopt;
f = scale.^2*x.^2;
f = f - sum((xopt*scale(idxM)).^2);
% f = exp(f) - 1;
% f = log10(f+1e-19) + 19;
f = f + 1e-19;
function f=fcornersphere(x) w = ones(size(x,1));
w(1) = 2.5; w(2)=2.5;
idx = x < 0;
f = sum(x(idx).^2);
idx = x > 0;
f = f + 2^2*sum(w(idx).*x(idx).^2);
function f=fsectorsphere(x, scal)
%
% This is deceptive for cumulative sigma control CSA in large dimension:
% The strategy (initially) diverges for N=50 and popsize = 150. (Even
% for cs==1 this can be observed for larger settings of N and
% popsize.) The reason is obvious from the function topology.
% Divergence can be avoided by setting boundaries or adding a
% penalty for large ||x||. Then, convergence can be observed again.
% Conclusion: for popsize>N cumulative sigma control is not completely
% reasonable, but I do not know better alternatives. In particular:
% TPA takes longer to converge than CSA when the latter still works.
%
if nargin < 2 || isempty (scal)
scal = 1e3;
end
f=sum(x.^2,1);
idx = x<0;
f = f + (scal^2 - 1) * sum((idx.*x).^2,1);
if 11 < 3
idxpen = find(f>1e9);
if ~isempty(idxpen)
f(idxpen) = f(idxpen) + 1e8*sum(x(:,idxpen).^2,1);
end
end
function f=fstepsphere(x, scal)
if nargin < 2 || isempty (scal)
scal = 1e0;
end
N = size(x,1);
f=1e-11+sum(scal.^((0:N-1)/(N-1))*floor(x+0.5).^2);
f=1e-11+sum(floor(scal.^((0:N-1)/(N-1))'.*x+0.5).^2);
% f=1e-11+sum(floor(x+0.5).^2);
function f=fstep(x)
% in -5.12..5.12 (bounded)
N = size(x,1);
f=1e-11+6*N+sum(floor(x));
function f=flnorm(x, scal, e)
if nargin < 2 || isempty(scal)
scal = 1;
end
if nargin < 3 || isempty(e)
e = 1;
end
if e==inf
f = max(abs(x));
else
N = size(x,1);
scale = scal.^((0:N-1)/(N-1))';
f=sum(abs(scale.*x).^e);
end
function f=fneumaier3(x)
% in -n^2..n^2
% x^*-i = i(n+1-i)
N = size(x,1);
% f = N*(N+4)*(N-1)/6 + sum((x-1).^2) - sum(x(1:N-1).*x(2:N));
f = sum((x-1).^2) - sum(x(1:N-1).*x(2:N));
function f = fmaxmindist(y)
% y in [-1,1], y(1:2) is first point on a plane, y(3:4) second etc % points best
% 5 1.4142
% 8 1.03527618
% 10 0.842535997
% 20 0.5997
pop = size(y,2);
N = size(y,1)/2;
f = [];
for ipop = 1:pop
if any(abs(y(:,ipop)) > 1)
f(ipop) = NaN;
else
x = reshape(y(:,ipop), [2, N]);
f(ipop) = inf;
for i = 1:N
f(ipop) = min(f(ipop), min(sqrt(sum((x(:,[1:i-1 i+1:N]) - repmat(x(:,i), 1, N-1)).^2, 1))));
end
end
end
f = -f;
function f=fchangingsphere(x)
N = size(x,1);
global scale_G; global count_G; if isempty(count_G) count_G=-1; end
count_G = count_G+1;
if mod(count_G,10) == 0
scale_G = 10.^(2*rand(1,N));
end
%disp(scale(1));
f = scale_G*x.^2;
function f= flogsphere(x)
f = 1-exp(-sum(x.^2));
function f= fexpsphere(x)
f = exp(sum(x.^2)) - 1;
function f=fbaluja(x)
% in [-0.16 0.16]
y = x(1);
for i = 2:length(x)
y(i) = x(i) + y(i-1);
end
f = 1e5 - 1/(1e-5 + sum(abs(y)));
function f=fschwefel(x)
f = 0;
for i = 1:size(x,1),
f = f+sum(x(1:i))^2;
end
function f=fcigar(x, ar)
if nargin < 2 || isempty(ar)
ar = 1e3;
end
f = x(1,:).^2 + ar^2*sum(x(2:end,:).^2,1);
function f=fcigtab(x)
f = x(1,:).^2 + 1e8*x(end,:).^2 + 1e4*sum(x(2:(end-1),:).^2, 1);
function f=ftablet(x)
f = 1e6*x(1,:).^2 + sum(x(2:end,:).^2, 1);
function f=felli(x, lgscal, expon, expon2)
% lgscal: log10(axis ratio)
% expon: x_i^expon, sphere==2
N = size(x,1); if N < 2 error('dimension must be greater one'); end
% x = x - repmat(-0.5+(1:N)',1,size(x,2)); % optimum in 1:N
if nargin < 2 || isempty(lgscal), lgscal = 3; end if nargin < 3 || isempty(expon), expon = 2; end
if nargin < 4 || isempty(expon2), expon2 = 1; end
f=((10^(lgscal*expon)).^((0:N-1)/(N-1)) * abs(x).^expon).^(1/expon2);
% if rand(1,1) > 0.015
% f = NaN;
% end
% f = f + randn(size(f));
function f=fellitest(x)
beta = 0.9;
N = size(x,1);
f = (1e6.^((0:(N-1))/(N-1))).^beta * (x.^2).^beta;
function f=fellii(x, scal)
N = size(x,1); if N < 2 error('dimension must be greater one'); end
if nargin < 2
scal = 1;
end
f= (scal*(1:N)).^2 * (x).^2;
function f=fellirot(x)
N = size(x,1);
global ORTHOGONALCOORSYSTEM_G
if isempty(ORTHOGONALCOORSYSTEM_G) ...
|| length(ORTHOGONALCOORSYSTEM_G) < N ...
|| isempty(ORTHOGONALCOORSYSTEM_G{N})
coordinatesystem(N);
end
f = felli(ORTHOGONALCOORSYSTEM_G{N}*x);
function f=frot(x, fun, varargin)
N = size(x,1);
global ORTHOGONALCOORSYSTEM_G
if isempty(ORTHOGONALCOORSYSTEM_G) ...
|| length(ORTHOGONALCOORSYSTEM_G) < N ...
|| isempty(ORTHOGONALCOORSYSTEM_G{N})
coordinatesystem(N);
end
f = feval(fun, ORTHOGONALCOORSYSTEM_G{N}*x, varargin{:});
function coordinatesystem(N)
if nargin < 1 || isempty(N)
arN = 2:30;
else
arN = N;
end
global ORTHOGONALCOORSYSTEM_G
ORTHOGONALCOORSYSTEM_G{1} = 1;
for N = arN
ar = randn(N,N);
for i = 1:N
for j = 1:i-1
ar(:,i) = ar(:,i) - ar(:,i)'*ar(:,j) * ar(:,j);
end
ar(:,i) = ar(:,i) / norm(ar(:,i));
end
ORTHOGONALCOORSYSTEM_G{N} = ar;
end
function f=fplane(x)
f=x(1);
function f=ftwoaxes(x)
f = sum(x(1:floor(end/2),:).^2, 1) + 1e6*sum(x(floor(1+end/2):end,:).^2, 1);
function f=fparabR(x)
f = -x(1,:) + 100*sum(x(2:end,:).^2,1);
function f=fsharpR(x)
f = abs(-x(1, :)).^2 + 100 * sqrt(sum(x(2:end,:).^2, 1));
function f=frosen(x)
if size(x,1) < 2 error('dimension must be greater one'); end
N = size(x,1);
popsi = size(x,2);
f = 1e2*sum((x(1:end-1,:).^2 - x(2:end,:)).^2,1) + sum((x(1:end-1,:)-1).^2,1);
% f = f + f^0.9 .* (2*randn(1,popsi) ./ randn(1,popsi).^0 / (2*N));
function f=frosenlin(x)
if size(x,1) < 2 error('dimension must be greater one'); end
x_org = x;
x(x>30) = 30;
x(x<-30) = -30;
f = 1e2*sum(-(x(1:end-1,:).^2 - x(2:end,:)),1) + ...
sum((x(1:end-1,:)-1).^2,1);
f = f + sum((x-x_org).^2,1);
% f(any(abs(x)>30,1)) = NaN;
function f=frosenrot(x)
N = size(x,1);
global ORTHOGONALCOORSYSTEM_G
if isempty(ORTHOGONALCOORSYSTEM_G) ...
|| length(ORTHOGONALCOORSYSTEM_G) < N ...
|| isempty(ORTHOGONALCOORSYSTEM_G{N})
coordinatesystem(N);
end
f = frosen(ORTHOGONALCOORSYSTEM_G{N}*x);
function f=frosenmodif(x)
f = 74 + 100*(x(2)-x(1)^2)^2 + (1-x(1))^2 ...
- 400*exp(-sum((x+1).^2)/2/0.05);
function f=fschwefelrosen1(x)
% in [-10 10]
f=sum((x.^2-x(1)).^2 + (x-1).^2);
function f=fschwefelrosen2(x)
% in [-10 10]
f=sum((x(2:end).^2-x(1)).^2 + (x(2:end)-1).^2);
function f=fdiffpow(x)
[N popsi] = size(x); if N < 2 error('dimension must be greater one'); end
f = sum(abs(x).^repmat(2+10*(0:N-1)'/(N-1), 1, popsi), 1);
f = sqrt(f);
function f=fabsprod(x)
f = sum(abs(x),1) + prod(abs(x),1);
function f=ffloor(x)
f = sum(floor(x+0.5).^2,1);
function f=fmaxx(x)
f = max(abs(x), [], 1);
%%% Multimodal functions
function f=fbirastrigin(x)
% todo: the volume needs to be a constant
N = size(x,1);
idx = (sum(x, 1) < 0.5*N); % global optimum
f = zeros(1,size(x,2));
f(idx) = 10*(N-sum(cos(2*pi*x(:,idx)),1)) + sum(x(:,idx).^2,1);
idx = ~idx;
f(idx) = 0.1 + 10*(N-sum(cos(2*pi*(x(:,idx)-2)),1)) + sum((x(:,idx)-2).^2,1);
function f=fackley(x)
% -32.768..32.768
% Adding a penalty outside the interval is recommended,
% because for large step sizes, fackley imposes like frand
%
N = size(x,1);
f = 20-20*exp(-0.2*sqrt(sum(x.^2)/N));
f = f + (exp(1) - exp(sum(cos(2*pi*x))/N));
% add penalty outside the search interval
f = f + sum((x(x>32.768)-32.768).^2) + sum((x(x<-32.768)+32.768).^2);
function f = fbohachevsky(x)
% -15..15
f = sum(x(1:end-1).^2 + 2 * x(2:end).^2 - 0.3 * cos(3*pi*x(1:end-1)) ...
- 0.4 * cos(4*pi*x(2:end)) + 0.7);
function f=fconcentric(x)
% in +-600
s = sum(x.^2);
f = s^0.25 * (sin(50*s^0.1)^2 + 1);
function f=fgriewank(x)
% in [-600 600]
[N, P] = size(x);
f = 1 - prod(cos(x'./sqrt(1:N))) + sum(x.^2)/4e3;
scale = repmat(sqrt(1:N)', 1, P);
f = 1 - prod(cos(x./scale), 1) + sum(x.^2, 1)/4e3;
% f = f + 1e4*sum(x(abs(x)>5).^2);
% if sum(x(abs(x)>5).^2) > 0
% f = 1e4 * sum(x(abs(x)>5).^2) + 1e8 * sum(x(x>5)).^2;
% end
function f=fgriewrosen(x)
% F13 or F8F2
[N, P] = size(x);
scale = repmat(sqrt(1:N)', 1, P);
y = [x(2:end,:); x(1,:)];
x = 100 * (x.^2 - y) + (x - 1).^2; % Rosenbrock part
f = 1 - prod(cos(x./scale), 1) + sum(x.^2, 1)/4e3;
f = sum(1 - cos(x) + x.^2/4e3, 1);
function f=fspallpseudorastrigin(x, scal, skewfac, skewstart, amplitude)
% by default multi-modal about between -30 and 30
if nargin < 5 || isempty(amplitude)
amplitude = 40;
end
if nargin < 4 || isempty(skewstart)
skewstart = 0;
end
if nargin < 3 || isempty(skewfac)
skewfac = 1;
end
if nargin < 2 || isempty(scal)
scal = 1;
end
N = size(x,1);
scale = 1;
if N > 1
scale=scal.^((0:N-1)'/(N-1));
end
% simple version:
% f = amplitude*(N - sum(cos(2*pi*(scale.*x)))) + sum((scale.*x).^2);
% skew version:
y = repmat(scale, 1, size(x,2)) .* x;
idx = find(x > skewstart);
if ~isempty(idx)
y(idx) = skewfac*y(idx);
end f = amplitude * (0*N-prod(cos((2*pi)^0*y),1)) + 0.05 * sum(y.^2,1) ...
+ randn(1,1);
function f=frastrigin(x, scal, skewfac, skewstart, amplitude)
% by default multi-modal about between -30 and 30
if nargin < 5 || isempty(amplitude)
amplitude = 10;
end
if nargin < 4 || isempty(skewstart)
skewstart = 0;
end
if nargin < 3 || isempty(skewfac)
skewfac = 1;
end
if nargin < 2 || isempty(scal)
scal = 1;
end
N = size(x,1);
scale = 1;
if N > 1
scale=scal.^((0:N-1)'/(N-1));
end
% simple version:
% f = amplitude*(N - sum(cos(2*pi*(scale.*x)))) + sum((scale.*x).^2);
% skew version:
y = repmat(scale, 1, size(x,2)) .* x;
idx = find(x > skewstart);
% idx = intersect(idx, 2:2:10);
if ~isempty(idx)
y(idx) = skewfac*y(idx);
end
f = amplitude * (N-sum(cos(2*pi*y),1)) + sum(y.^2,1);
function f=frastriginmax(x)
N = size(x,1);
f = (N/20)*807.06580387678 - (10 * (N-sum(cos(2*pi*x),1)) + sum(x.^2,1));
f(any(abs(x) > 5.12)) = 1e2*N;
function f = fschaffer(x)
% -100..100
N = size(x,1);
s = x(1:N-1,:).^2 + x(2:N,:).^2;
f = sum(s.^0.25 .* (sin(50*s.^0.1).^2+1), 1);
function f=fschwefelmult(x)
% -500..500
%
N = size(x,1);
f = - sum(x.*sin(sqrt(abs(x))), 1);
f = 418.9829*N - 1.27275661e-5*N - sum(x.*sin(sqrt(abs(x))), 1);
% penalty term
f = f + 1e4*sum((abs(x)>500) .* (abs(x)-500).^2, 1);
function f=ftwomax(x)
% Boundaries at +/-5
N = size(x,1);
f = -abs(sum(x)) + 5*N;
function f=ftwomaxtwo(x)
% Boundaries at +/-10
N = size(x,1);
f = abs(sum(x));
if f > 30
f = f - 30;
end
f = -f;
function f=frand(x)
f=1./(1-rand(1, size(x,2))) - 1;
% CHANGES
% 12/02/19: "future" setting of ccum, correcting for large mueff, is default now
% 11/11/15: bug-fix: max value for ccovmu_sep setting corrected
% 10/11/11: (3.52.beta) boundary handling: replace max with min in change
% rate formula. Active CMA: check of pos.def. improved.
% Plotting: value of lambda appears in the title.
% 10/04/03: (3.51.beta) active CMA cleaned up. Equal fitness detection
% looks into history now.
% 10/03/08: (3.50.beta) "active CMA" revised and bug-fix of ambiguous
% option Noise.alpha -> Noise.alphasigma.
% 09/10/12: (3.40.beta) a slightly modified version of "active CMA",
% that is a negative covariance matrix update, use option
% CMA.active. In 10;30;90-D the gain on ftablet is a factor
% of 1.6;2.5;4.4 (the scaling improves by sqrt(N)). On
% Rosenbrock the gain is about 25%. On sharp ridge the
% behavior is improved. Cigar is unchanged.
% 09/08/10: local plotcmaesdat remains in backround
% 09/08/10: bug-fix in time management for data writing, logtime was not
% considered properly (usually not at all).
% 09/07/05: V3.24: stagnation termination added
% 08/09/27: V3.23: momentum alignment is out-commented and de-preciated
% 08/09/25: V3.22: re-alignment of sigma and C was buggy
% 08/07/15: V3.20, CMA-parameters are options now. ccov and mucov were replaced
% by ccov1 \approx ccov/mucov and ccovmu \approx (1-1/mucov)*ccov
% 08/06/30: file name xrecent was change to xrecentbest (compatible with other
% versions)
% 08/06/29: time stamp added to output files
% 08/06/28: bug fixed with resume option, commentary did not work
% 08/06/28: V3.10, uncertainty (noise) handling added (re-implemented), according
% to reference "A Method for Handling Uncertainty..." from below.
% 08/06/28: bug fix: file xrecent was empty
% 08/06/01: diagonalonly clean up. >1 means some iterations.
% 08/05/05: output is written to file preventing an increasing data
% array and ease long runs.
% 08/03/27: DiagonalOnly<0 learns for -DiagonalOnly iterations only the
% diagonal with a larger learning rate.
% 08/03 (2.60): option DiagonalOnly>=1 invokes a time- and space-linear
% variant with only diagonal elements of the covariance matrix
% updating. This can be useful for large dimensions, say > 100.
% 08/02: diag(weights) * ... replaced with repmat(weights,1,N) .* ...
% in C update, implies O(mu*N^2) instead of O(mu^2*N + mu*N^2).
% 07/09: tolhistfun as termination criterion added, "<" changed to
% "<=" also for TolFun to allow for stopping on zero difference.
% Name tolfunhist clashes with option tolfun.
% 07/07: hsig threshold made slighly smaller for large dimension,
% useful for lambda < lambda_default.
% 07/06: boundary handling: scaling in the boundary handling
% is omitted now, see bnd.flgscale. This seems not to
% have a big impact. Using the scaling is worse on rotated
% functions, but better on separable ones.
% 07/05: boundary handling: weight i is not incremented anymore
% if xmean(i) moves towards the feasible space. Increment
% factor changed to 1.2 instead of 1.1.
% 07/05: boundary handling code simplified not changing the algorithm
% 07/04: bug removed for saving in octave
% 06/11/10: more testing of outcome of eig, fixed max(D) to max(diag(D))
% 06/10/21: conclusive final bestever assignment in the end
% 06/10/21: restart and incpopsize option implemented for restarts
% with increasing population size, version 2.50.
% 06/09/16: output argument bestever inserted again for convenience and
% backward compatibility
% 06/08: output argument out and struct out reorganized.
% 06/01: Possible parallel evaluation included as option EvalParallel
% 05/11: Compatibility to octave implemented, package octave-forge
% is needed.
% 05/09: Raise of figure and waiting for first plots improved
% 05/01: Function coordinatesystem cleaned up.
% 05/01: Function prctile, which requires the statistics toolbox,
% replaced by myprctile. % 05/01: Option warnonequalfunctionvalues included.
% 04/12: Decrease of sigma removed. Problems on fsectorsphere can
% be addressed better by adding search space boundaries.
% 04/12: Boundary handling simpyfied.
% 04/12: Bug when stopping criteria tolx or tolupx are vectors.
% 04/11: Three input parameters are obligatory now.
% 04/11: Bug in boundary handling removed: Boundary weights can decrease now.
% 04/11: Normalization for boundary weights scale changed.
% 04/11: VerboseModulo option bug removed. Documentation improved.
% 04/11: Condition for increasing boundary weights changed.
% 04/10: Decrease of sigma when fitness is getting consistenly
% worse. Addresses the problems appearing on fsectorsphere for
% large population size.
% 04/10: VerboseModulo option included.
% 04/10: Bug for condition for increasing boundary weights removed.
% 04/07: tolx depends on initial sigma to achieve scale invariance
% for this stopping criterion.
% 04/06: Objective function value NaN is not counted as function
% evaluation and invokes resampling of the search point.
% 04/06: Error handling for eigenvalue beeing zero (never happens
% with default parameter setting)
% 04/05: damps further tuned for large mueff
% o Details for stall of pc-adaptation added (variable hsig
% introduced).
% 04/05: Bug in boundary handling removed: A large initial SIGMA was
% corrected not until *after* the first iteration, which could
% lead to a complete failure.
% 04/05: Call of function range (works with stats toolbox only)
% changed to myrange.
% 04/04: Parameter cs depends on mueff now and damps \propto sqrt(mueff)
% instead of \propto mueff.
% o Initial stall to adapt C (flginiphase) is removed and
% adaptation of pc is stalled for large norm(ps) instead.
% o Returned default options include documentation.
% o Resume part reorganized.
% 04/03: Stopflag becomes cell-array.
% ---------------------------------------------------------------
% CMA-ES: Evolution Strategy with Covariance Matrix Adaptation for
% nonlinear function minimization. To be used under the terms of the
% GNU General Public License (http://www.gnu.org/copyleft/gpl.html).
% Author (copyright): Nikolaus Hansen, 2001-2008.
% e-mail: nikolaus.hansen AT inria.fr
% URL:http://www.bionik.tu-berlin.de/user/niko
% References: See below.
% ---------------------------------------------------------------
%
% GENERAL PURPOSE: The CMA-ES (Evolution Strategy with Covariance
% Matrix Adaptation) is a robust search method which should be
% applied, if derivative based methods, e.g. quasi-Newton BFGS or
% conjucate gradient, (supposably) fail due to a rugged search
% landscape (e.g. noise, local optima, outlier, etc.). On smooth
% landscapes CMA-ES is roughly ten times slower than BFGS. For up to
% N=10 variables even the simplex direct search method (Nelder & Mead)
% is often faster, but far less robust than CMA-ES. To see the
% advantage of the CMA, it will usually take at least 30*N and up to
% 300*N function evaluations, where N is the search problem dimension.
% On considerably hard problems the complete search (a single run) is
% expected to take at least 30*N^2 and up to 300*N^2 function
% evaluations.
%
% SOME MORE COMMENTS:
% The adaptation of the covariance matrix (e.g. by the CMA) is
% equivalent to a general linear transformation of the problem
% coding. Nevertheless every problem specific knowlegde about the best
% linear transformation should be exploited before starting the
% search. That is, an appropriate a priori transformation should be
% applied to the problem. This also makes the identity matrix as
% initial covariance matrix the best choice.
%% The strategy parameter lambda (population size, opts.PopSize) is the
% preferred strategy parameter to play with. If results with the
% default strategy are not satisfactory, increase the population
% size. (Remark that the crucial parameter mu (opts.ParentNumber) is
% increased proportionally to lambda). This will improve the
% strategies capability of handling noise and local minima. We
% recomment successively increasing lambda by a factor of about three,
% starting with initial values between 5 and 20. Casually, population
% sizes even beyond 1000+100*N can be sensible.
%
%
% ---------------------------------------------------------------
%%% REFERENCES
%
% The equation numbers refer to
% Hansen, N. and S. Kern (2004). Evaluating the CMA Evolution
% Strategy on Multimodal Test Functions. Eighth International
% Conference on Parallel Problem Solving from Nature PPSN VIII,
% Proceedings, pp. 282-291, Berlin: Springer.
% (http://www.bionik.tu-berlin.de/user/niko/ppsn2004hansenkern.pdf)
%
% Further references:
% Hansen, N. and A. Ostermeier (2001). Completely Derandomized
% Self-Adaptation in Evolution Strategies. Evolutionary Computation,
% 9(2), pp. 159-195.
% (http://www.bionik.tu-berlin.de/user/niko/cmaartic.pdf).
%
% Hansen, N., S.D. Mueller and P. Koumoutsakos (2003). Reducing the
% Time Complexity of the Derandomized Evolution Strategy with
% Covariance Matrix Adaptation (CMA-ES). Evolutionary Computation,
% 11(1). (http://mitpress.mit.edu/journals/pdf/evco_11_1_1_0.pdf).
%
% Ros, R. and N. Hansen (2008). A Simple Modification in CMA-ES
% Achieving Linear Time and Space Complexity. To appear in Tenth
% International Conference on Parallel Problem Solving from Nature
% PPSN X, Proceedings, Berlin: Springer.
%
% Hansen, N., A.S.P. Niederberger, L. Guzzella and P. Koumoutsakos
% (2009?). A Method for Handling Uncertainty in Evolutionary
% Optimization with an Application to Feedback Control of
% Combustion. To appear in IEEE Transactions on Evolutionary
% Computation.